blob: c69fbaca7b43953f395d30212a83241751bd37e6 [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
76SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -080077 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080078 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -080079 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070080 "src/f32-avgpool/mp9p8q-scalar.c",
81 "src/f32-avgpool/up9-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -080082 "src/f32-bilinear/gen/scalar-c1.c",
83 "src/f32-bilinear/gen/scalar-c2.c",
84 "src/f32-bilinear/gen/scalar-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070085 "src/f32-clamp/scalar.c",
Marat Dukhan441e2212019-12-04 18:30:49 -080086 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -080087 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen563df5f2019-10-23 08:02:21 -070088 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen0cc2c532019-10-15 04:44:18 -070089 "src/f32-dwconv-spchw/3x3p1-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -070090 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080091 "src/f32-dwconv-spchw/5x5p2-scalar.c",
Erich Elsen38709a62019-11-08 11:58:45 -080092 "src/f32-dwconv-spchw/5x5s2p2-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -080093 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
94 "src/f32-dwconv/gen/up1x25-scalar.c",
95 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
96 "src/f32-dwconv/gen/up1x4-scalar.c",
97 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
98 "src/f32-dwconv/gen/up1x9-scalar.c",
99 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
100 "src/f32-dwconv/gen/up2x25-scalar.c",
101 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
102 "src/f32-dwconv/gen/up2x4-scalar.c",
103 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
104 "src/f32-dwconv/gen/up2x9-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -0700105 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700106 "src/f32-gavgpool/mp7p7q-scalar.c",
107 "src/f32-gavgpool/up7-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800108 "src/f32-gemm/gen-inc/1x4-scalar.c",
109 "src/f32-gemm/gen-inc/2x4-scalar.c",
110 "src/f32-gemm/gen-inc/4x4-scalar.c",
111 "src/f32-gemm/gen/1x4-scalar.c",
112 "src/f32-gemm/gen/2x4-scalar.c",
113 "src/f32-gemm/gen/4x2-scalar.c",
114 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800115 "src/f32-hswish/gen/scalar-x1.c",
116 "src/f32-hswish/gen/scalar-x2.c",
117 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800118 "src/f32-igemm/gen/1x4-scalar.c",
119 "src/f32-igemm/gen/2x4-scalar.c",
120 "src/f32-igemm/gen/4x2-scalar.c",
121 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800122 "src/f32-maxpool/9p8x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700123 "src/f32-pavgpool/mp9p8q-scalar.c",
124 "src/f32-pavgpool/up9-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800125 "src/f32-ppmm/gen/2x4-scalar.c",
126 "src/f32-ppmm/gen/3x3-scalar.c",
127 "src/f32-ppmm/gen/4x2-scalar.c",
128 "src/f32-ppmm/gen/4x4-scalar.c",
129 "src/f32-prelu/gen/scalar-2x1.c",
130 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700131 "src/f32-rmax/scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800132 "src/f32-spmm/gen/1x1-scalar-pipelined.c",
133 "src/f32-spmm/gen/1x1-scalar.c",
134 "src/f32-spmm/gen/2x1-scalar-pipelined.c",
135 "src/f32-spmm/gen/2x1-scalar.c",
136 "src/f32-spmm/gen/4x1-scalar-pipelined.c",
137 "src/f32-spmm/gen/4x1-scalar.c",
138 "src/f32-spmm/gen/8x1-scalar-pipelined.c",
139 "src/f32-spmm/gen/8x1-scalar.c",
140 "src/f32-spmm/gen/8x2-scalar.c",
141 "src/f32-spmm/gen/8x4-scalar.c",
142 "src/f32-vbinary/gen/vadd-scalar-x1.c",
143 "src/f32-vbinary/gen/vadd-scalar-x2.c",
144 "src/f32-vbinary/gen/vadd-scalar-x4.c",
145 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
146 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
147 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800148 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
149 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
150 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
151 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
152 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
153 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800154 "src/f32-vbinary/gen/vmax-scalar-x1.c",
155 "src/f32-vbinary/gen/vmax-scalar-x2.c",
156 "src/f32-vbinary/gen/vmax-scalar-x4.c",
157 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
158 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
159 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
160 "src/f32-vbinary/gen/vmin-scalar-x1.c",
161 "src/f32-vbinary/gen/vmin-scalar-x2.c",
162 "src/f32-vbinary/gen/vmin-scalar-x4.c",
163 "src/f32-vbinary/gen/vminc-scalar-x1.c",
164 "src/f32-vbinary/gen/vminc-scalar-x2.c",
165 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800166 "src/f32-vbinary/gen/vmul-scalar-x1.c",
167 "src/f32-vbinary/gen/vmul-scalar-x2.c",
168 "src/f32-vbinary/gen/vmul-scalar-x4.c",
169 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
170 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
171 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800172 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
173 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
174 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800175 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
176 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
177 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
178 "src/f32-vbinary/gen/vsub-scalar-x1.c",
179 "src/f32-vbinary/gen/vsub-scalar-x2.c",
180 "src/f32-vbinary/gen/vsub-scalar-x4.c",
181 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
182 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
183 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
184 "src/f32-vmulcaddc/gen/c1-scalar-2x.c",
185 "src/f32-vmulcaddc/gen/c2-scalar-2x.c",
186 "src/f32-vmulcaddc/gen/c4-scalar-2x.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800187 "src/math/expminus-scalar-lut2048-p1.c",
188 "src/math/expminus-scalar-lut64-p2.c",
189 "src/math/expminus-scalar-p5.c",
190 "src/math/sigmoid-scalar-lut2048-p1-div.c",
191 "src/math/sigmoid-scalar-lut64-p2-div.c",
192 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700193 "src/q8-avgpool/mp9p8q-scalar.c",
194 "src/q8-avgpool/up9-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700195 "src/q8-dwconv/up1x9-scalar.c",
196 "src/q8-gavgpool/mp7p7q-scalar.c",
197 "src/q8-gavgpool/up7-scalar.c",
198 "src/q8-gemm/2x2-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800199 "src/q8-igemm/2x2-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700200 "src/q8-vadd/scalar.c",
201 "src/u8-clamp/scalar.c",
202 "src/u8-lut32norm/scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800203 "src/u8-maxpool/9p8x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700204 "src/u8-rmax/scalar.c",
205 "src/x32-packx/x2-scalar.c",
206 "src/x32-packx/x3-scalar.c",
207 "src/x32-packx/x4-scalar.c",
208 "src/x32-pad/x2-scalar.c",
209 "src/x32-unpool/scalar.c",
210 "src/x32-zip/x2-scalar.c",
211 "src/x32-zip/x3-scalar.c",
212 "src/x32-zip/x4-scalar.c",
213 "src/x32-zip/xm-scalar.c",
214 "src/x8-lut/scalar.c",
215 "src/x8-zip/x2-scalar.c",
216 "src/x8-zip/x3-scalar.c",
217 "src/x8-zip/x4-scalar.c",
218 "src/x8-zip/xm-scalar.c",
219]
220
Marat Dukhan436ebe62019-12-04 15:10:12 -0800221WASM_UKERNELS = [
222 "src/f32-avgpool/mp9p8q-wasm.c",
223 "src/f32-avgpool/up9-wasm.c",
224 "src/f32-clamp/wasm.c",
225 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
226 "src/f32-dwconv/gen/up1x25-wasm.c",
227 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
228 "src/f32-dwconv/gen/up1x4-wasm.c",
229 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
230 "src/f32-dwconv/gen/up1x9-wasm.c",
231 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
232 "src/f32-dwconv/gen/up2x25-wasm.c",
233 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
234 "src/f32-dwconv/gen/up2x4-wasm.c",
235 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
236 "src/f32-dwconv/gen/up2x9-wasm.c",
237 "src/f32-gavgpool/mp7p7q-wasm.c",
238 "src/f32-gavgpool/up7-wasm.c",
239 "src/f32-gemm/gen-inc/1x4-wasm.c",
240 "src/f32-gemm/gen-inc/2x4-wasm.c",
241 "src/f32-gemm/gen-inc/4x4-wasm.c",
242 "src/f32-gemm/gen/1x4-wasm.c",
243 "src/f32-gemm/gen/2x4-wasm.c",
244 "src/f32-gemm/gen/4x2-wasm.c",
245 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800246 "src/f32-hswish/gen/wasm-x1.c",
247 "src/f32-hswish/gen/wasm-x2.c",
248 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800249 "src/f32-igemm/gen/1x4-wasm.c",
250 "src/f32-igemm/gen/2x4-wasm.c",
251 "src/f32-igemm/gen/4x2-wasm.c",
252 "src/f32-igemm/gen/4x4-wasm.c",
253 "src/f32-maxpool/9p8x-wasm-c1.c",
254 "src/f32-pavgpool/mp9p8q-wasm.c",
255 "src/f32-pavgpool/up9-wasm.c",
256 "src/f32-prelu/gen/wasm-2x1.c",
257 "src/f32-prelu/gen/wasm-2x4.c",
258 "src/f32-vbinary/gen/vadd-wasm-x1.c",
259 "src/f32-vbinary/gen/vadd-wasm-x2.c",
260 "src/f32-vbinary/gen/vadd-wasm-x4.c",
261 "src/f32-vbinary/gen/vaddc-wasm-x1.c",
262 "src/f32-vbinary/gen/vaddc-wasm-x2.c",
263 "src/f32-vbinary/gen/vaddc-wasm-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800264 "src/f32-vbinary/gen/vdiv-wasm-x1.c",
265 "src/f32-vbinary/gen/vdiv-wasm-x2.c",
266 "src/f32-vbinary/gen/vdiv-wasm-x4.c",
267 "src/f32-vbinary/gen/vdivc-wasm-x1.c",
268 "src/f32-vbinary/gen/vdivc-wasm-x2.c",
269 "src/f32-vbinary/gen/vdivc-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800270 "src/f32-vbinary/gen/vmax-wasm-x1.c",
271 "src/f32-vbinary/gen/vmax-wasm-x2.c",
272 "src/f32-vbinary/gen/vmax-wasm-x4.c",
273 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
274 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
275 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
276 "src/f32-vbinary/gen/vmin-wasm-x1.c",
277 "src/f32-vbinary/gen/vmin-wasm-x2.c",
278 "src/f32-vbinary/gen/vmin-wasm-x4.c",
279 "src/f32-vbinary/gen/vminc-wasm-x1.c",
280 "src/f32-vbinary/gen/vminc-wasm-x2.c",
281 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800282 "src/f32-vbinary/gen/vmul-wasm-x1.c",
283 "src/f32-vbinary/gen/vmul-wasm-x2.c",
284 "src/f32-vbinary/gen/vmul-wasm-x4.c",
285 "src/f32-vbinary/gen/vmulc-wasm-x1.c",
286 "src/f32-vbinary/gen/vmulc-wasm-x2.c",
287 "src/f32-vbinary/gen/vmulc-wasm-x4.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800288 "src/f32-vbinary/gen/vrdivc-wasm-x1.c",
289 "src/f32-vbinary/gen/vrdivc-wasm-x2.c",
290 "src/f32-vbinary/gen/vrdivc-wasm-x4.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800291 "src/f32-vbinary/gen/vrsubc-wasm-x1.c",
292 "src/f32-vbinary/gen/vrsubc-wasm-x2.c",
293 "src/f32-vbinary/gen/vrsubc-wasm-x4.c",
294 "src/f32-vbinary/gen/vsub-wasm-x1.c",
295 "src/f32-vbinary/gen/vsub-wasm-x2.c",
296 "src/f32-vbinary/gen/vsub-wasm-x4.c",
297 "src/f32-vbinary/gen/vsubc-wasm-x1.c",
298 "src/f32-vbinary/gen/vsubc-wasm-x2.c",
299 "src/f32-vbinary/gen/vsubc-wasm-x4.c",
300 "src/f32-vmulcaddc/gen/c1-wasm-2x.c",
301 "src/f32-vmulcaddc/gen/c2-wasm-2x.c",
302 "src/f32-vmulcaddc/gen/c4-wasm-2x.c",
303]
304
Marat Dukhan08c4a432019-10-03 09:29:21 -0700305PSIMD_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800306 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800307 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800308 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700309 "src/f32-avgpool/mp9p8q-psimd.c",
310 "src/f32-avgpool/up9-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800311 "src/f32-bilinear/gen/psimd-c4.c",
312 "src/f32-bilinear/gen/psimd-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700313 "src/f32-clamp/psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800314 "src/f32-dwconv/gen/up4x25-psimd-acc2.c",
315 "src/f32-dwconv/gen/up4x25-psimd.c",
316 "src/f32-dwconv/gen/up4x4-psimd-acc2.c",
317 "src/f32-dwconv/gen/up4x4-psimd.c",
318 "src/f32-dwconv/gen/up4x9-psimd-acc2.c",
319 "src/f32-dwconv/gen/up4x9-psimd.c",
320 "src/f32-dwconv/gen/up8x25-psimd-acc2.c",
321 "src/f32-dwconv/gen/up8x25-psimd.c",
322 "src/f32-dwconv/gen/up8x4-psimd-acc2.c",
323 "src/f32-dwconv/gen/up8x4-psimd.c",
324 "src/f32-dwconv/gen/up8x9-psimd-acc2.c",
325 "src/f32-dwconv/gen/up8x9-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700326 "src/f32-gavgpool/mp7p7q-psimd.c",
327 "src/f32-gavgpool/up7-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800328 "src/f32-gemm/gen/1x8-psimd-loadsplat.c",
329 "src/f32-gemm/gen/1x8-psimd-splat.c",
330 "src/f32-gemm/gen/1x8s4-psimd.c",
331 "src/f32-gemm/gen/4x8-psimd-loadsplat.c",
332 "src/f32-gemm/gen/4x8-psimd-splat.c",
333 "src/f32-gemm/gen/4x8s4-psimd.c",
334 "src/f32-gemm/gen/6x8-psimd-loadsplat.c",
335 "src/f32-gemm/gen/6x8-psimd-splat.c",
336 "src/f32-gemm/gen/6x8s4-psimd.c",
337 "src/f32-gemm/gen-inc/1x8-psimd-loadsplat.c",
338 "src/f32-gemm/gen-inc/1x8-psimd-splat.c",
339 "src/f32-gemm/gen-inc/1x8s4-psimd.c",
340 "src/f32-gemm/gen-inc/4x8-psimd-loadsplat.c",
341 "src/f32-gemm/gen-inc/4x8-psimd-splat.c",
342 "src/f32-gemm/gen-inc/4x8s4-psimd.c",
343 "src/f32-gemm/gen-inc/6x8-psimd-loadsplat.c",
344 "src/f32-gemm/gen-inc/6x8-psimd-splat.c",
345 "src/f32-gemm/gen-inc/6x8s4-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800346 "src/f32-hswish/gen/psimd-x4.c",
347 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800348 "src/f32-igemm/gen/1x8-psimd-loadsplat.c",
349 "src/f32-igemm/gen/1x8-psimd-splat.c",
350 "src/f32-igemm/gen/1x8s4-psimd.c",
351 "src/f32-igemm/gen/4x2c4-psimd.c",
352 "src/f32-igemm/gen/4x8-psimd-loadsplat.c",
353 "src/f32-igemm/gen/4x8-psimd-splat.c",
354 "src/f32-igemm/gen/4x8s4-psimd.c",
355 "src/f32-igemm/gen/6x8-psimd-loadsplat.c",
356 "src/f32-igemm/gen/6x8-psimd-splat.c",
357 "src/f32-igemm/gen/6x8s4-psimd.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800358 "src/f32-maxpool/9p8x-psimd-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700359 "src/f32-pavgpool/mp9p8q-psimd.c",
360 "src/f32-pavgpool/up9-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800361 "src/f32-ppmm/gen/4x8-psimd.c",
362 "src/f32-prelu/gen/psimd-2x4.c",
363 "src/f32-prelu/gen/psimd-2x8.c",
364 "src/f32-vbinary/gen/vadd-psimd-x4.c",
365 "src/f32-vbinary/gen/vadd-psimd-x8.c",
366 "src/f32-vbinary/gen/vaddc-psimd-x4.c",
367 "src/f32-vbinary/gen/vaddc-psimd-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800368 "src/f32-vbinary/gen/vdiv-psimd-x4.c",
369 "src/f32-vbinary/gen/vdiv-psimd-x8.c",
370 "src/f32-vbinary/gen/vdivc-psimd-x4.c",
371 "src/f32-vbinary/gen/vdivc-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800372 "src/f32-vbinary/gen/vmax-psimd-x4.c",
373 "src/f32-vbinary/gen/vmax-psimd-x8.c",
374 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
375 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
376 "src/f32-vbinary/gen/vmin-psimd-x4.c",
377 "src/f32-vbinary/gen/vmin-psimd-x8.c",
378 "src/f32-vbinary/gen/vminc-psimd-x4.c",
379 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800380 "src/f32-vbinary/gen/vmul-psimd-x4.c",
381 "src/f32-vbinary/gen/vmul-psimd-x8.c",
382 "src/f32-vbinary/gen/vmulc-psimd-x4.c",
383 "src/f32-vbinary/gen/vmulc-psimd-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800384 "src/f32-vbinary/gen/vrdivc-psimd-x4.c",
385 "src/f32-vbinary/gen/vrdivc-psimd-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800386 "src/f32-vbinary/gen/vrsubc-psimd-x4.c",
387 "src/f32-vbinary/gen/vrsubc-psimd-x8.c",
388 "src/f32-vbinary/gen/vsub-psimd-x4.c",
389 "src/f32-vbinary/gen/vsub-psimd-x8.c",
390 "src/f32-vbinary/gen/vsubc-psimd-x4.c",
391 "src/f32-vbinary/gen/vsubc-psimd-x8.c",
392 "src/f32-vmulcaddc/gen/c4-psimd-2x.c",
393 "src/f32-vmulcaddc/gen/c8-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700394 "src/x32-packx/x4-psimd.c",
395 "src/x32-pad/x2-psimd.c",
396 "src/x32-unpool/psimd.c",
397 "src/x32-zip/x2-psimd.c",
398 "src/x32-zip/x3-psimd.c",
399 "src/x32-zip/x4-psimd.c",
400 "src/x32-zip/xm-psimd.c",
401]
402
403# ISA-specific micro-kernels
404NEON_UKERNELS = [
405 "src/f32-avgpool/mp9p8q-neon.c",
406 "src/f32-avgpool/up9-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800407 "src/f32-bilinear/gen/neon-c4.c",
408 "src/f32-bilinear/gen/neon-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700409 "src/f32-clamp/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800410 "src/f32-dwconv/gen/up4x9-neon.c",
411 "src/f32-dwconv/gen/up4x9-neon-acc2.c",
412 "src/f32-dwconv/gen/up8x9-neon.c",
413 "src/f32-dwconv/gen/up8x9-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700414 "src/f32-gavgpool-spchw/neon-x4.c",
415 "src/f32-gavgpool/mp7p7q-neon.c",
416 "src/f32-gavgpool/up7-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800417 "src/f32-gemm/gen/1x8-neon-lane-ld64.c",
418 "src/f32-gemm/gen/4x2-neon-lane-ld64.c",
419 "src/f32-gemm/gen/4x8-neon-lane-ld128.c",
420 "src/f32-gemm/gen/4x8-neon-lane-ld64.c",
421 "src/f32-gemm/gen/5x8-neon-lane-ld64.c",
422 "src/f32-gemm/gen/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800423 "src/f32-gemm/gen/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800424 "src/f32-gemm/gen/1x8-neon-dup-ld64.c",
425 "src/f32-gemm/gen/4x8-neon-dup-ld128.c",
426 "src/f32-gemm/gen/4x8-neon-dup-ld64.c",
427 "src/f32-gemm/gen/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800428 "src/f32-gemm/gen/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800429 "src/f32-gemm/gen/1x8s4-neon.c",
430 "src/f32-gemm/gen/4x8s4-neon.c",
431 "src/f32-gemm/gen/6x8s4-neon.c",
432 "src/f32-gemm/gen/8x8s4-neon.c",
433 "src/f32-gemm/gen-inc/1x8-neon-lane-ld64.c",
434 "src/f32-gemm/gen-inc/4x8-neon-lane-ld128.c",
435 "src/f32-gemm/gen-inc/4x8-neon-lane-ld64.c",
436 "src/f32-gemm/gen-inc/5x8-neon-lane-ld64.c",
437 "src/f32-gemm/gen-inc/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800438 "src/f32-gemm/gen-inc/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800439 "src/f32-gemm/gen-inc/1x8-neon-dup-ld64.c",
440 "src/f32-gemm/gen-inc/4x8-neon-dup-ld128.c",
441 "src/f32-gemm/gen-inc/4x8-neon-dup-ld64.c",
442 "src/f32-gemm/gen-inc/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800443 "src/f32-gemm/gen-inc/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800444 "src/f32-gemm/gen-inc/1x8s4-neon.c",
445 "src/f32-gemm/gen-inc/4x8s4-neon.c",
446 "src/f32-gemm/gen-inc/6x8s4-neon.c",
447 "src/f32-gemm/gen-inc/8x8s4-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800448 "src/f32-hswish/gen/neon-x4.c",
449 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800450 "src/f32-igemm/gen/1x8-neon-lane-ld64.c",
451 "src/f32-igemm/gen/4x2-neon-lane-ld64.c",
452 "src/f32-igemm/gen/4x4-neon-lane-ld64.c",
453 "src/f32-igemm/gen/4x8-neon-lane-ld128.c",
454 "src/f32-igemm/gen/4x8-neon-lane-ld64.c",
455 "src/f32-igemm/gen/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800456 "src/f32-igemm/gen/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800457 "src/f32-igemm/gen/1x8-neon-dup-ld64.c",
458 "src/f32-igemm/gen/4x8-neon-dup-ld128.c",
459 "src/f32-igemm/gen/4x8-neon-dup-ld64.c",
460 "src/f32-igemm/gen/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800461 "src/f32-igemm/gen/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800462 "src/f32-igemm/gen/1x8s4-neon.c",
463 "src/f32-igemm/gen/4x8s4-neon.c",
464 "src/f32-igemm/gen/6x8s4-neon.c",
465 "src/f32-igemm/gen/8x8s4-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700466 "src/f32-pavgpool/mp9p8q-neon.c",
467 "src/f32-pavgpool/up9-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800468 "src/f32-ppmm/gen/4x8-neon.c",
469 "src/f32-ppmm/gen/8x8-neon.c",
470 "src/f32-prelu/gen/neon-2x4.c",
471 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700472 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800473 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
474 "src/f32-vbinary/gen/vadd-neon-x4.c",
475 "src/f32-vbinary/gen/vadd-neon-x8.c",
476 "src/f32-vbinary/gen/vaddc-neon-x4.c",
477 "src/f32-vbinary/gen/vaddc-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800478 "src/f32-vbinary/gen/vmax-neon-x4.c",
479 "src/f32-vbinary/gen/vmax-neon-x8.c",
480 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
481 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
482 "src/f32-vbinary/gen/vmin-neon-x4.c",
483 "src/f32-vbinary/gen/vmin-neon-x8.c",
484 "src/f32-vbinary/gen/vminc-neon-x4.c",
485 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800486 "src/f32-vbinary/gen/vmul-neon-x4.c",
487 "src/f32-vbinary/gen/vmul-neon-x8.c",
488 "src/f32-vbinary/gen/vmulc-neon-x4.c",
489 "src/f32-vbinary/gen/vmulc-neon-x8.c",
490 "src/f32-vbinary/gen/vrsubc-neon-x4.c",
491 "src/f32-vbinary/gen/vrsubc-neon-x8.c",
492 "src/f32-vbinary/gen/vsub-neon-x4.c",
493 "src/f32-vbinary/gen/vsub-neon-x8.c",
494 "src/f32-vbinary/gen/vsubc-neon-x4.c",
495 "src/f32-vbinary/gen/vsubc-neon-x8.c",
496 "src/f32-vmulcaddc/gen/c4-neon-2x.c",
497 "src/f32-vmulcaddc/gen/c8-neon-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700498 "src/q8-avgpool/mp9p8q-neon.c",
499 "src/q8-avgpool/up9-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700500 "src/q8-dwconv/up8x9-neon.c",
501 "src/q8-gavgpool/mp7p7q-neon.c",
502 "src/q8-gavgpool/up7-neon.c",
503 "src/q8-gemm/4x8-neon.c",
504 "src/q8-gemm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800505 "src/q8-igemm/4x8-neon.c",
506 "src/q8-igemm/8x8-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700507 "src/q8-vadd/neon.c",
508 "src/u8-clamp/neon.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800509 "src/u8-maxpool/9p8x-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700510 "src/u8-rmax/neon.c",
511 "src/x32-packx/x4-neon-st4.c",
512 "src/x32-pad/x2-neon.c",
513 "src/x32-zip/x2-neon.c",
514 "src/x32-zip/x3-neon.c",
515 "src/x32-zip/x4-neon.c",
516 "src/x32-zip/xm-neon.c",
517 "src/x8-zip/x2-neon.c",
518 "src/x8-zip/x3-neon.c",
519 "src/x8-zip/x4-neon.c",
520 "src/x8-zip/xm-neon.c",
521]
522
523NEONFMA_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800524 "src/f32-bilinear/gen/neonfma-c4.c",
525 "src/f32-bilinear/gen/neonfma-c8.c",
526 "src/f32-igemm/gen/1x8-neonfma-dup-ld64.c",
527 "src/f32-igemm/gen/4x8-neonfma-dup-ld128.c",
528 "src/f32-igemm/gen/4x8-neonfma-dup-ld64.c",
529 "src/f32-igemm/gen/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800530 "src/f32-igemm/gen/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800531 "src/f32-igemm/gen/1x8s4-neonfma.c",
532 "src/f32-igemm/gen/4x8s4-neonfma.c",
533 "src/f32-igemm/gen/6x8s4-neonfma.c",
534 "src/f32-igemm/gen/8x8s4-neonfma.c",
535 "src/f32-dwconv/gen/up4x9-neonfma.c",
536 "src/f32-dwconv/gen/up4x9-neonfma-acc2.c",
537 "src/f32-dwconv/gen/up8x9-neonfma.c",
538 "src/f32-dwconv/gen/up8x9-neonfma-acc2.c",
539 "src/f32-gemm/gen/1x8-neonfma-dup-ld64.c",
540 "src/f32-gemm/gen/4x8-neonfma-dup-ld128.c",
541 "src/f32-gemm/gen/4x8-neonfma-dup-ld64.c",
542 "src/f32-gemm/gen/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800543 "src/f32-gemm/gen/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800544 "src/f32-gemm/gen/1x8s4-neonfma.c",
545 "src/f32-gemm/gen/4x8s4-neonfma.c",
546 "src/f32-gemm/gen/6x8s4-neonfma.c",
547 "src/f32-gemm/gen/8x8s4-neonfma.c",
548 "src/f32-gemm/gen-inc/1x8-neonfma-dup-ld64.c",
549 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld128.c",
550 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld64.c",
551 "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800552 "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800553 "src/f32-gemm/gen-inc/1x8s4-neonfma.c",
554 "src/f32-gemm/gen-inc/4x8s4-neonfma.c",
555 "src/f32-gemm/gen-inc/6x8s4-neonfma.c",
556 "src/f32-gemm/gen-inc/8x8s4-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800557 "src/f32-hswish/gen/neonfma-x4.c",
558 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800559 "src/f32-ppmm/gen/4x8-neonfma.c",
560 "src/f32-ppmm/gen/8x8-neonfma.c",
561 "src/f32-sigmoid/gen/neonfma-p5-nr2fma-x16.c",
562 "src/f32-vmulcaddc/gen/c4-neonfma-2x.c",
563 "src/f32-vmulcaddc/gen/c8-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800564 "src/math/exp-neonfma-lut64-p2.c",
565 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800566 "src/math/expminus-neonfma-lut2048-p1.c",
567 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800568 "src/math/expminus-neonfma-p5.c",
Marat Dukhan91f8d862019-11-27 12:25:42 -0800569 "src/math/sigmoid-neonfma-lut2048-p1-nr1recps1fma.c",
570 "src/math/sigmoid-neonfma-lut2048-p1-nr2fma.c",
571 "src/math/sigmoid-neonfma-lut2048-p1-nr2recps.c",
Marat Dukhan22aae132019-11-22 17:10:29 -0800572 "src/math/sigmoid-neonfma-p5-nr1recps1fma.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -0800573 "src/math/sigmoid-neonfma-p5-nr2fma.c",
Marat Dukhan22aae132019-11-22 17:10:29 -0800574 "src/math/sigmoid-neonfma-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700575]
576
577AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan77ca6302019-12-06 12:48:15 -0800578 "src/f32-vbinary/gen/vdiv-neon-x4.c",
579 "src/f32-vbinary/gen/vdiv-neon-x8.c",
580 "src/f32-vbinary/gen/vdivc-neon-x4.c",
581 "src/f32-vbinary/gen/vdivc-neon-x8.c",
582 "src/f32-vbinary/gen/vrdivc-neon-x4.c",
583 "src/f32-vbinary/gen/vrdivc-neon-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800584 "src/f32-gemm/gen/1x8-neonfma-lane-ld64.c",
585 "src/f32-gemm/gen/4x2-neonfma-lane-ld64.c",
586 "src/f32-gemm/gen/4x8-neonfma-lane-ld128.c",
587 "src/f32-gemm/gen/4x8-neonfma-lane-ld64.c",
588 "src/f32-gemm/gen/5x8-neonfma-lane-ld64.c",
589 "src/f32-gemm/gen/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800590 "src/f32-gemm/gen/6x8-neonfma-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800591 "src/f32-gemm/gen-inc/1x8-neonfma-lane-ld64.c",
592 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld128.c",
593 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld64.c",
594 "src/f32-gemm/gen-inc/5x8-neonfma-lane-ld64.c",
595 "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800596 "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800597 "src/f32-igemm/gen/1x8-neonfma-lane-ld64.c",
598 "src/f32-igemm/gen/4x2-neonfma-lane-ld64.c",
599 "src/f32-igemm/gen/4x4-neonfma-lane-ld64.c",
600 "src/f32-igemm/gen/4x8-neonfma-lane-ld128.c",
601 "src/f32-igemm/gen/4x8-neonfma-lane-ld64.c",
602 "src/f32-igemm/gen/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800603 "src/f32-igemm/gen/6x8-neonfma-lane-ld128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700604 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
605 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
606 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
607 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
608 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
609 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
610 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800611 "src/f32-spmm/gen/12x1-neonfma.c",
612 "src/f32-spmm/gen/12x2-neonfma.c",
613 "src/f32-spmm/gen/12x4-neonfma.c",
614 "src/f32-spmm/gen/16x1-neonfma-pipelined.c",
615 "src/f32-spmm/gen/16x1-neonfma-unroll2.c",
616 "src/f32-spmm/gen/16x1-neonfma.c",
617 "src/f32-spmm/gen/16x2-neonfma.c",
618 "src/f32-spmm/gen/16x4-neonfma.c",
619 "src/f32-spmm/gen/4x1-neonfma-pipelined.c",
620 "src/f32-spmm/gen/4x1-neonfma-unroll2.c",
621 "src/f32-spmm/gen/4x1-neonfma.c",
622 "src/f32-spmm/gen/4x2-neonfma.c",
623 "src/f32-spmm/gen/4x4-neonfma.c",
624 "src/f32-spmm/gen/8x1-neonfma-pipelined.c",
625 "src/f32-spmm/gen/8x1-neonfma-unroll2.c",
626 "src/f32-spmm/gen/8x1-neonfma.c",
627 "src/f32-spmm/gen/8x2-neonfma.c",
628 "src/f32-spmm/gen/8x4-neonfma.c",
Marat Dukhan22aae132019-11-22 17:10:29 -0800629 "src/math/sigmoid-neonfma-p5-div.c",
Marat Dukhan91f8d862019-11-27 12:25:42 -0800630 "src/math/sigmoid-neonfma-lut2048-p1-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700631]
632
633AARCH64_NEONFP16ARITH_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800634 "src/f16-gemm/gen/4x8-neonfp16arith-ld64.c",
635 "src/f16-gemm/gen/6x8-neonfp16arith-ld64.c",
636 "src/f16-gemm/gen/8x8-neonfp16arith-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700637]
638
639SSE_UKERNELS = [
640 "src/f32-avgpool/mp9p8q-sse.c",
641 "src/f32-avgpool/up9-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800642 "src/f32-bilinear/gen/sse-c4.c",
643 "src/f32-bilinear/gen/sse-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700644 "src/f32-clamp/sse.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800645 "src/f32-dwconv-spchw/3x3p1-sse.c",
646 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800647 "src/f32-dwconv/gen/up4x25-sse-acc2.c",
648 "src/f32-dwconv/gen/up4x25-sse.c",
649 "src/f32-dwconv/gen/up4x4-sse-acc2.c",
650 "src/f32-dwconv/gen/up4x4-sse.c",
651 "src/f32-dwconv/gen/up4x9-sse-acc2.c",
652 "src/f32-dwconv/gen/up4x9-sse.c",
653 "src/f32-dwconv/gen/up8x25-sse-acc2.c",
654 "src/f32-dwconv/gen/up8x25-sse.c",
655 "src/f32-dwconv/gen/up8x4-sse-acc2.c",
656 "src/f32-dwconv/gen/up8x4-sse.c",
657 "src/f32-dwconv/gen/up8x9-sse-acc2.c",
658 "src/f32-dwconv/gen/up8x9-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700659 "src/f32-gavgpool-spchw/sse-x4.c",
660 "src/f32-gavgpool/mp7p7q-sse.c",
661 "src/f32-gavgpool/up7-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800662 "src/f32-gemm/gen/1x8-sse-dup.c",
663 "src/f32-gemm/gen/1x8-sse-load1.c",
664 "src/f32-gemm/gen/1x8s4-sse.c",
665 "src/f32-gemm/gen/4x8-sse-dup.c",
666 "src/f32-gemm/gen/4x8-sse-load1.c",
667 "src/f32-gemm/gen/4x8s4-sse.c",
668 "src/f32-gemm/gen-inc/1x8-sse-dup.c",
669 "src/f32-gemm/gen-inc/1x8-sse-load1.c",
670 "src/f32-gemm/gen-inc/1x8s4-sse.c",
671 "src/f32-gemm/gen-inc/4x8-sse-dup.c",
672 "src/f32-gemm/gen-inc/4x8-sse-load1.c",
673 "src/f32-gemm/gen-inc/4x8s4-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800674 "src/f32-hswish/gen/sse-x4.c",
675 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800676 "src/f32-igemm/gen/1x8-sse-dup.c",
677 "src/f32-igemm/gen/1x8-sse-load1.c",
678 "src/f32-igemm/gen/1x8s4-sse.c",
679 "src/f32-igemm/gen/4x2c4-sse.c",
680 "src/f32-igemm/gen/4x8-sse-dup.c",
681 "src/f32-igemm/gen/4x8-sse-load1.c",
682 "src/f32-igemm/gen/4x8s4-sse.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800683 "src/f32-maxpool/9p8x-sse-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700684 "src/f32-pavgpool/mp9p8q-sse.c",
685 "src/f32-pavgpool/up9-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800686 "src/f32-ppmm/gen/4x8-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700687 "src/f32-rmax/sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800688 "src/f32-spmm/gen/4x1-sse.c",
689 "src/f32-spmm/gen/8x1-sse.c",
690 "src/f32-vbinary/gen/vadd-sse-x4.c",
691 "src/f32-vbinary/gen/vadd-sse-x8.c",
692 "src/f32-vbinary/gen/vaddc-sse-x4.c",
693 "src/f32-vbinary/gen/vaddc-sse-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800694 "src/f32-vbinary/gen/vdiv-sse-x4.c",
695 "src/f32-vbinary/gen/vdiv-sse-x8.c",
696 "src/f32-vbinary/gen/vdivc-sse-x4.c",
697 "src/f32-vbinary/gen/vdivc-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800698 "src/f32-vbinary/gen/vmax-sse-x4.c",
699 "src/f32-vbinary/gen/vmax-sse-x8.c",
700 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
701 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
702 "src/f32-vbinary/gen/vmin-sse-x4.c",
703 "src/f32-vbinary/gen/vmin-sse-x8.c",
704 "src/f32-vbinary/gen/vminc-sse-x4.c",
705 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800706 "src/f32-vbinary/gen/vmul-sse-x4.c",
707 "src/f32-vbinary/gen/vmul-sse-x8.c",
708 "src/f32-vbinary/gen/vmulc-sse-x4.c",
709 "src/f32-vbinary/gen/vmulc-sse-x8.c",
Marat Dukhan77ca6302019-12-06 12:48:15 -0800710 "src/f32-vbinary/gen/vrdivc-sse-x4.c",
711 "src/f32-vbinary/gen/vrdivc-sse-x8.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800712 "src/f32-vbinary/gen/vrsubc-sse-x4.c",
713 "src/f32-vbinary/gen/vrsubc-sse-x8.c",
714 "src/f32-vbinary/gen/vsub-sse-x4.c",
715 "src/f32-vbinary/gen/vsub-sse-x8.c",
716 "src/f32-vbinary/gen/vsubc-sse-x4.c",
717 "src/f32-vbinary/gen/vsubc-sse-x8.c",
718 "src/f32-vmulcaddc/gen/c4-sse-2x.c",
719 "src/f32-vmulcaddc/gen/c8-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700720 "src/x32-packx/x4-sse.c",
721]
722
723SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800724 "src/f32-argmaxpool/9p8x-sse2-c4.c",
725 "src/f32-argmaxpool/4x-sse2-c4.c",
726 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800727 "src/f32-prelu/gen/sse2-2x4.c",
728 "src/f32-prelu/gen/sse2-2x8.c",
729 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
730 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700731 "src/q8-avgpool/mp9p8q-sse2.c",
732 "src/q8-avgpool/up9-sse2.c",
733 "src/q8-igemm/4x4c2-sse2.c",
734 "src/q8-dwconv/up8x9-sse2.c",
735 "src/q8-gavgpool/mp7p7q-sse2.c",
736 "src/q8-gavgpool/up7-sse2.c",
737 "src/q8-gemm/2x4c8-sse2.c",
738 "src/q8-gemm/4x4c2-sse2.c",
739 "src/q8-vadd/sse2.c",
740 "src/u8-clamp/sse2.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800741 "src/u8-maxpool/9p8x-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700742 "src/u8-rmax/sse2.c",
743 "src/x32-pad/x2-sse2.c",
744 "src/x32-zip/x2-sse2.c",
745 "src/x32-zip/x3-sse2.c",
746 "src/x32-zip/x4-sse2.c",
747 "src/x32-zip/xm-sse2.c",
748 "src/x8-zip/x2-sse2.c",
749 "src/x8-zip/x3-sse2.c",
750 "src/x8-zip/x4-sse2.c",
751 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -0800752 "src/math/exp-sse2-p5.c",
753 "src/math/expminus-sse2-p5.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -0800754 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700755]
756
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800757SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800758 "src/f32-prelu/gen/sse41-2x4.c",
759 "src/f32-prelu/gen/sse41-2x8.c",
760 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
761 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800762]
763
Marat Dukhan08c4a432019-10-03 09:29:21 -0700764AVX_UKERNELS = [
Marat Dukhane2c3f292019-11-27 15:40:54 -0800765 "src/f32-clamp/avx.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800766 "src/f32-dwconv/gen/up16x4-avx-acc2.c",
767 "src/f32-dwconv/gen/up16x4-avx.c",
768 "src/f32-dwconv/gen/up8x4-avx-acc2.c",
769 "src/f32-dwconv/gen/up8x4-avx.c",
770 "src/f32-dwconv/gen/up16x9-avx-acc2.c",
771 "src/f32-dwconv/gen/up16x9-avx.c",
772 "src/f32-dwconv/gen/up8x9-avx-acc2.c",
773 "src/f32-dwconv/gen/up8x9-avx.c",
774 "src/f32-dwconv/gen/up16x25-avx-acc2.c",
775 "src/f32-dwconv/gen/up16x25-avx.c",
776 "src/f32-dwconv/gen/up8x25-avx-acc2.c",
777 "src/f32-dwconv/gen/up8x25-avx.c",
778 "src/f32-gemm/gen/1x8-avx-broadcast.c",
779 "src/f32-gemm/gen/4x8-avx-broadcast.c",
780 "src/f32-gemm/gen/5x8-avx-broadcast.c",
781 "src/f32-gemm/gen/6x8-avx-broadcast.c",
782 "src/f32-gemm/gen/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -0800783 "src/f32-gemm/gen/1x16-avx-broadcast.c",
784 "src/f32-gemm/gen/3x16-avx-broadcast.c",
785 "src/f32-gemm/gen/4x16-avx-broadcast.c",
786 "src/f32-gemm/gen/5x16-avx-broadcast.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800787 "src/f32-gemm/gen-inc/1x8-avx-broadcast.c",
788 "src/f32-gemm/gen-inc/4x8-avx-broadcast.c",
789 "src/f32-gemm/gen-inc/5x8-avx-broadcast.c",
790 "src/f32-gemm/gen-inc/6x8-avx-broadcast.c",
791 "src/f32-gemm/gen-inc/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -0800792 "src/f32-gemm/gen-inc/1x16-avx-broadcast.c",
793 "src/f32-gemm/gen-inc/3x16-avx-broadcast.c",
794 "src/f32-gemm/gen-inc/4x16-avx-broadcast.c",
795 "src/f32-gemm/gen-inc/5x16-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800796 "src/f32-hswish/gen/avx-x8.c",
797 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800798 "src/f32-igemm/gen/1x8-avx-broadcast.c",
799 "src/f32-igemm/gen/4x8-avx-broadcast.c",
800 "src/f32-igemm/gen/5x8-avx-broadcast.c",
801 "src/f32-igemm/gen/6x8-avx-broadcast.c",
802 "src/f32-igemm/gen/7x8-avx-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -0800803 "src/f32-igemm/gen/1x16-avx-broadcast.c",
804 "src/f32-igemm/gen/3x16-avx-broadcast.c",
805 "src/f32-igemm/gen/4x16-avx-broadcast.c",
806 "src/f32-igemm/gen/5x16-avx-broadcast.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700807 "src/f32-rmax/avx.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -0800808 "src/f32-vbinary/gen/vadd-avx-x8.c",
809 "src/f32-vbinary/gen/vadd-avx-x16.c",
810 "src/f32-vbinary/gen/vaddc-avx-x8.c",
811 "src/f32-vbinary/gen/vaddc-avx-x16.c",
812 "src/f32-vbinary/gen/vdiv-avx-x8.c",
813 "src/f32-vbinary/gen/vdiv-avx-x16.c",
814 "src/f32-vbinary/gen/vdivc-avx-x8.c",
815 "src/f32-vbinary/gen/vdivc-avx-x16.c",
816 "src/f32-vbinary/gen/vmax-avx-x8.c",
817 "src/f32-vbinary/gen/vmax-avx-x16.c",
818 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
819 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
820 "src/f32-vbinary/gen/vmin-avx-x8.c",
821 "src/f32-vbinary/gen/vmin-avx-x16.c",
822 "src/f32-vbinary/gen/vminc-avx-x8.c",
823 "src/f32-vbinary/gen/vminc-avx-x16.c",
824 "src/f32-vbinary/gen/vmul-avx-x8.c",
825 "src/f32-vbinary/gen/vmul-avx-x16.c",
826 "src/f32-vbinary/gen/vmulc-avx-x8.c",
827 "src/f32-vbinary/gen/vmulc-avx-x16.c",
828 "src/f32-vbinary/gen/vrdivc-avx-x8.c",
829 "src/f32-vbinary/gen/vrdivc-avx-x16.c",
830 "src/f32-vbinary/gen/vrsubc-avx-x8.c",
831 "src/f32-vbinary/gen/vrsubc-avx-x16.c",
832 "src/f32-vbinary/gen/vsub-avx-x8.c",
833 "src/f32-vbinary/gen/vsub-avx-x16.c",
834 "src/f32-vbinary/gen/vsubc-avx-x8.c",
835 "src/f32-vbinary/gen/vsubc-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700836 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700837]
838
Marat Dukhanfda12b82019-11-21 12:27:59 -0800839FMA3_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800840 "src/f32-dwconv/gen/up16x4-fma3-acc2.c",
841 "src/f32-dwconv/gen/up16x4-fma3.c",
842 "src/f32-dwconv/gen/up8x4-fma3-acc2.c",
843 "src/f32-dwconv/gen/up8x4-fma3.c",
844 "src/f32-dwconv/gen/up16x9-fma3-acc2.c",
845 "src/f32-dwconv/gen/up16x9-fma3.c",
846 "src/f32-dwconv/gen/up8x9-fma3-acc2.c",
847 "src/f32-dwconv/gen/up8x9-fma3.c",
848 "src/f32-dwconv/gen/up16x25-fma3-acc2.c",
849 "src/f32-dwconv/gen/up16x25-fma3.c",
850 "src/f32-dwconv/gen/up8x25-fma3-acc2.c",
851 "src/f32-dwconv/gen/up8x25-fma3.c",
852 "src/f32-gemm/gen/1x8-fma3-broadcast.c",
853 "src/f32-gemm/gen/4x8-fma3-broadcast.c",
854 "src/f32-gemm/gen/5x8-fma3-broadcast.c",
855 "src/f32-gemm/gen/6x8-fma3-broadcast.c",
856 "src/f32-gemm/gen/7x8-fma3-broadcast.c",
857 "src/f32-gemm/gen/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -0800858 "src/f32-gemm/gen/1x16-fma3-broadcast.c",
859 "src/f32-gemm/gen/3x16-fma3-broadcast.c",
860 "src/f32-gemm/gen/4x16-fma3-broadcast.c",
861 "src/f32-gemm/gen/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -0800862 "src/f32-gemm/gen/1x16s4-fma3-broadcast.c",
863 "src/f32-gemm/gen/3x16s4-fma3-broadcast.c",
864 "src/f32-gemm/gen/4x16s4-fma3-broadcast.c",
865 "src/f32-gemm/gen/5x16s4-fma3-broadcast.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800866 "src/f32-gemm/gen-inc/1x8-fma3-broadcast.c",
867 "src/f32-gemm/gen-inc/4x8-fma3-broadcast.c",
868 "src/f32-gemm/gen-inc/5x8-fma3-broadcast.c",
869 "src/f32-gemm/gen-inc/6x8-fma3-broadcast.c",
870 "src/f32-gemm/gen-inc/7x8-fma3-broadcast.c",
871 "src/f32-gemm/gen-inc/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -0800872 "src/f32-gemm/gen-inc/1x16-fma3-broadcast.c",
873 "src/f32-gemm/gen-inc/3x16-fma3-broadcast.c",
874 "src/f32-gemm/gen-inc/4x16-fma3-broadcast.c",
875 "src/f32-gemm/gen-inc/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -0800876 "src/f32-gemm/gen-inc/1x16s4-fma3-broadcast.c",
877 "src/f32-gemm/gen-inc/3x16s4-fma3-broadcast.c",
878 "src/f32-gemm/gen-inc/4x16s4-fma3-broadcast.c",
879 "src/f32-gemm/gen-inc/5x16s4-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800880 "src/f32-hswish/gen/fma3-x8.c",
881 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800882 "src/f32-igemm/gen/1x8-fma3-broadcast.c",
883 "src/f32-igemm/gen/4x8-fma3-broadcast.c",
884 "src/f32-igemm/gen/5x8-fma3-broadcast.c",
885 "src/f32-igemm/gen/6x8-fma3-broadcast.c",
886 "src/f32-igemm/gen/7x8-fma3-broadcast.c",
887 "src/f32-igemm/gen/8x8-fma3-broadcast.c",
Marat Dukhaneccfd712019-12-08 16:49:27 -0800888 "src/f32-igemm/gen/1x16-fma3-broadcast.c",
889 "src/f32-igemm/gen/3x16-fma3-broadcast.c",
890 "src/f32-igemm/gen/4x16-fma3-broadcast.c",
891 "src/f32-igemm/gen/5x16-fma3-broadcast.c",
Ashkan Aliabadid94b8562019-12-10 11:33:51 -0800892 "src/f32-igemm/gen/1x16s4-fma3-broadcast.c",
893 "src/f32-igemm/gen/3x16s4-fma3-broadcast.c",
894 "src/f32-igemm/gen/4x16s4-fma3-broadcast.c",
895 "src/f32-igemm/gen/5x16s4-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -0800896]
897
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700898AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -0800899 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
900 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
901 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
902 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
903 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
904 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
905 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
906 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
907 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
908 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
909 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
910 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
911 "src/f32-raddextexp/gen/avx2-p5-x64.c",
912 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
913 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
914 "src/f32-raddextexp/gen/avx2-p5-x72.c",
915 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
916 "src/f32-raddextexp/gen/avx2-p5-x80.c",
917 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
918 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
919 "src/f32-raddextexp/gen/avx2-p5-x96.c",
920 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
921 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
922 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
923 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
924 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
925 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
926 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
927 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
928 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
929 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
930 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
931 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
932 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
933 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
934 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
935 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
936 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
937 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
938 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
939 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
940 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
941 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
942 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
943 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
944 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
945 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
946 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
947 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
948 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
949 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
950 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
951 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
952 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
953 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
954 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
955 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
956 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
957 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
958 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700959 "src/math/exp-avx2-p5.c",
960 "src/math/exp-avx2-perm-p3.c",
961 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -0700962 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700963 "src/math/extexp-avx2-p5.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700964]
965
Marat Dukhan08c4a432019-10-03 09:29:21 -0700966AVX512F_UKERNELS = [
Marat Dukhane2c3f292019-11-27 15:40:54 -0800967 "src/f32-clamp/avx512f.c",
Marat Dukhan479f87e2019-11-27 15:17:06 -0800968 "src/f32-dwconv/gen/up32x4-avx512f-acc2.c",
969 "src/f32-dwconv/gen/up32x4-avx512f.c",
970 "src/f32-dwconv/gen/up16x4-avx512f-acc2.c",
971 "src/f32-dwconv/gen/up16x4-avx512f.c",
972 "src/f32-dwconv/gen/up32x9-avx512f-acc2.c",
973 "src/f32-dwconv/gen/up32x9-avx512f.c",
974 "src/f32-dwconv/gen/up16x9-avx512f-acc2.c",
975 "src/f32-dwconv/gen/up16x9-avx512f.c",
976 "src/f32-dwconv/gen/up32x25-avx512f-acc2.c",
977 "src/f32-dwconv/gen/up32x25-avx512f.c",
978 "src/f32-dwconv/gen/up16x25-avx512f-acc2.c",
979 "src/f32-dwconv/gen/up16x25-avx512f.c",
Marat Dukhan0f349c42019-11-27 11:58:54 -0800980 "src/f32-gemm/gen/1x16-avx512f-broadcast.c",
981 "src/f32-gemm/gen/4x16-avx512f-broadcast.c",
982 "src/f32-gemm/gen/5x16-avx512f-broadcast.c",
983 "src/f32-gemm/gen/6x16-avx512f-broadcast.c",
984 "src/f32-gemm/gen/7x16-avx512f-broadcast.c",
985 "src/f32-gemm/gen/8x16-avx512f-broadcast.c",
986 "src/f32-gemm/gen-inc/1x16-avx512f-broadcast.c",
987 "src/f32-gemm/gen-inc/4x16-avx512f-broadcast.c",
988 "src/f32-gemm/gen-inc/5x16-avx512f-broadcast.c",
989 "src/f32-gemm/gen-inc/6x16-avx512f-broadcast.c",
990 "src/f32-gemm/gen-inc/7x16-avx512f-broadcast.c",
991 "src/f32-gemm/gen-inc/8x16-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800992 "src/f32-hswish/gen/avx512f-x16.c",
993 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan0f349c42019-11-27 11:58:54 -0800994 "src/f32-igemm/gen/1x16-avx512f-broadcast.c",
995 "src/f32-igemm/gen/4x16-avx512f-broadcast.c",
996 "src/f32-igemm/gen/5x16-avx512f-broadcast.c",
997 "src/f32-igemm/gen/6x16-avx512f-broadcast.c",
998 "src/f32-igemm/gen/7x16-avx512f-broadcast.c",
999 "src/f32-igemm/gen/8x16-avx512f-broadcast.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001000 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1001 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1002 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1003 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1004 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1005 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1006 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1007 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1008 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1009 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1010 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1011 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1012 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1013 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1014 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1015 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1016 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1017 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1018 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1019 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1020 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1021 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1022 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1023 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1024 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1025 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1026 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1027 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1028 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1029 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1030 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1031 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1032 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1033 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1034 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1035 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001036 "src/f32-rmax/avx512f.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001037 "src/f32-vbinary/gen/vadd-avx512f-x16.c",
1038 "src/f32-vbinary/gen/vadd-avx512f-x32.c",
1039 "src/f32-vbinary/gen/vaddc-avx512f-x16.c",
1040 "src/f32-vbinary/gen/vaddc-avx512f-x32.c",
1041 "src/f32-vbinary/gen/vdiv-avx512f-x16.c",
1042 "src/f32-vbinary/gen/vdiv-avx512f-x32.c",
1043 "src/f32-vbinary/gen/vdivc-avx512f-x16.c",
1044 "src/f32-vbinary/gen/vdivc-avx512f-x32.c",
1045 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1046 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1047 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1048 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1049 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1050 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1051 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1052 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
1053 "src/f32-vbinary/gen/vmul-avx512f-x16.c",
1054 "src/f32-vbinary/gen/vmul-avx512f-x32.c",
1055 "src/f32-vbinary/gen/vmulc-avx512f-x16.c",
1056 "src/f32-vbinary/gen/vmulc-avx512f-x32.c",
1057 "src/f32-vbinary/gen/vrdivc-avx512f-x16.c",
1058 "src/f32-vbinary/gen/vrdivc-avx512f-x32.c",
1059 "src/f32-vbinary/gen/vrsubc-avx512f-x16.c",
1060 "src/f32-vbinary/gen/vrsubc-avx512f-x32.c",
1061 "src/f32-vbinary/gen/vsub-avx512f-x16.c",
1062 "src/f32-vbinary/gen/vsub-avx512f-x32.c",
1063 "src/f32-vbinary/gen/vsubc-avx512f-x16.c",
1064 "src/f32-vbinary/gen/vsubc-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001065 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001066 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1067 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1068 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1069 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1070 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1071 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1072 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1073 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1074 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1075 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1076 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1077 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1078 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1079 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1080 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1081 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1082 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1083 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1084 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1085 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1086 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1087 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1088 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1089 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001090 "src/math/exp-avx512f-p5-scalef.c",
1091 "src/math/exp-avx512f-p5.c",
1092 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001093 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001094 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001095]
1096
1097AARCH32_ASM_UKERNELS = [
1098 "src/q8-dwconv/up8x9-aarch32-neon.S",
Frank Barchard13916042019-12-11 10:56:34 -08001099 "src/f32-gemm/4x8-aarch32-neon-cortex-a53.S",
Frank Barchardabf81542019-12-13 16:18:30 -08001100 "src/f32-gemm/gen/4x8-aarch32-neon-cortex-a75.S",
1101 "src/f32-gemm/gen/4x8-aarch32-neon-pld-cortex-a75.S",
Frank Barchard8b0f0262019-11-27 23:18:40 -08001102 "src/f32-gemm/4x8-aarch32-neon-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001103]
1104
1105AARCH64_ASM_UKERNELS = [
1106 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
1107 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001108 "src/f32-gemm/gen/1x12-aarch64-neonfma-cortex-a53.S",
1109 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a53.S",
1110 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
1111 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
1112 "src/f32-gemm/gen/4x12-aarch64-neonfma-cortex-a53.S",
1113 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a53.S",
1114 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
1115 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
1116 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld128.S",
1117 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld64.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001118 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001119 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
1120 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001121 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a73.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001122 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001123 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
1124 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld128.S",
1125 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld64.S",
1126 "src/f32-gemm/gen-inc/1x12-aarch64-neonfma-cortex-a53.S",
1127 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a53.S",
1128 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a57.S",
1129 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a75.S",
1130 "src/f32-gemm/gen-inc/4x12-aarch64-neonfma-cortex-a53.S",
1131 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a53.S",
1132 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a57.S",
1133 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a75.S",
1134 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld128.S",
1135 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld64.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001136 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001137 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a75.S",
1138 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001139 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a73.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001140 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a57.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001141 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a75.S",
1142 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld128.S",
1143 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001144 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -07001145 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001146 "src/f32-igemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
1147 "src/f32-igemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001148 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -07001149 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001150 "src/f32-igemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
1151 "src/f32-igemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
1152 "src/f32-igemm/gen/5x8-aarch64-neonfma-cortex-a57.S",
1153 "src/f32-igemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -07001154 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001155 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
Frank Barchard387c2d12019-12-16 19:14:07 -08001156 "src/f32-igemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
1157 "src/f32-igemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001158]
1159
1160INTERNAL_MICROKERNEL_HDRS = [
1161 "src/xnnpack/argmaxpool.h",
1162 "src/xnnpack/avgpool.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08001163 "src/xnnpack/bilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001164 "src/xnnpack/clamp.h",
1165 "src/xnnpack/common.h",
1166 "src/xnnpack/conv.h",
1167 "src/xnnpack/dwconv.h",
1168 "src/xnnpack/gavgpool.h",
1169 "src/xnnpack/gemm.h",
1170 "src/xnnpack/hswish.h",
1171 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001172 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001173 "src/xnnpack/lut.h",
1174 "src/xnnpack/math.h",
1175 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001176 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001177 "src/xnnpack/packx.h",
1178 "src/xnnpack/pad.h",
1179 "src/xnnpack/params.h",
1180 "src/xnnpack/pavgpool.h",
1181 "src/xnnpack/ppmm.h",
1182 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001183 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001184 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001185 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001186 "src/xnnpack/rmax.h",
1187 "src/xnnpack/scalar-utils.h",
1188 "src/xnnpack/spmm.h",
1189 "src/xnnpack/unpool.h",
1190 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001191 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001192 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001193 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001194 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001195 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001196 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001197 "src/xnnpack/zip.h",
1198]
1199
1200INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1201 "include/xnnpack.h",
1202 "src/xnnpack/allocator.h",
1203 "src/xnnpack/compute.h",
1204 "src/xnnpack/im2col.h",
1205 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001206 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001207 "src/xnnpack/operator.h",
1208 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001209 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001210 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001211 "src/xnnpack/requantization.h",
1212]
1213
1214ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1215 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001216]
1217
1218MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001219 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001220 "include/xnnpack.h",
1221]
1222
1223MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
1224 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001225 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001226 "src/xnnpack/requantization.h",
1227 "include/xnnpack.h",
1228]
1229
1230OPERATOR_TEST_PARAMS_HDRS = [
1231 "src/xnnpack/params.h",
1232 "src/xnnpack/common.h",
1233]
1234
1235WEIGHTS_PACK_HDRS = [
1236 "src/xnnpack/pack.h",
1237 "src/xnnpack/operator.h",
1238 "src/xnnpack/compute.h",
1239]
1240
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001241LOGGING_COPTS = select({
1242 # No logging in optimized mode
1243 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1244 # Full logging in debug mode
1245 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1246 # Error-only logging in default (fastbuild) mode
1247 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1248})
1249
1250LOGGING_HDRS = [
1251 "src/xnnpack/log.h",
1252]
1253
Marat Dukhan08c4a432019-10-03 09:29:21 -07001254xnnpack_cc_library(
1255 name = "scalar_ukernels",
1256 srcs = SCALAR_UKERNELS,
1257 hdrs = INTERNAL_HDRS,
1258 aarch32_copts = ["-marm"],
1259 copts = xnnpack_std_copts(),
1260 deps = [
1261 "@FP16",
1262 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001263 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001264 ],
1265)
1266
1267xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001268 name = "wasm_ukernels",
1269 hdrs = INTERNAL_HDRS,
1270 copts = xnnpack_std_copts(),
1271 wasm_srcs = WASM_UKERNELS,
1272 deps = [
1273 "@FP16",
1274 "@FXdiv",
1275 "@pthreadpool",
1276 ],
1277)
1278
1279xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001280 name = "psimd_ukernels",
1281 srcs = PSIMD_UKERNELS,
1282 hdrs = INTERNAL_HDRS,
1283 aarch32_copts = [
1284 "-marm",
1285 "-mfpu=neon",
1286 ],
1287 copts = xnnpack_std_copts(),
1288 optimized_copts = [
1289 "-O3",
1290 "-ffast-math",
1291 ],
1292 deps = [
1293 "@FP16",
1294 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001295 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001296 ],
1297)
1298
1299xnnpack_cc_library(
1300 name = "neon_ukernels",
1301 hdrs = INTERNAL_HDRS,
1302 aarch32_copts = [
1303 "-marm",
1304 "-mfpu=neon",
1305 ],
1306 aarch32_srcs = NEON_UKERNELS,
1307 aarch64_srcs = NEON_UKERNELS,
1308 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001309 deps = [
1310 "@FP16",
1311 "@pthreadpool",
1312 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001313)
1314
1315xnnpack_cc_library(
1316 name = "neonfma_ukernels",
1317 hdrs = INTERNAL_HDRS,
1318 aarch32_copts = [
1319 "-marm",
1320 "-mfpu=neon-vfpv4",
1321 ],
1322 aarch32_srcs = NEONFMA_UKERNELS,
1323 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
1324 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001325 deps = [
1326 "@FP16",
1327 "@pthreadpool",
1328 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001329)
1330
1331xnnpack_cc_library(
1332 name = "neonfp16arith_ukernels",
1333 hdrs = INTERNAL_HDRS,
1334 aarch64_copts = ["-march=armv8.2-a+fp16"],
1335 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
1336 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001337 deps = [
1338 "@FP16",
1339 "@pthreadpool",
1340 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001341)
1342
1343xnnpack_cc_library(
1344 name = "sse2_ukernels",
1345 hdrs = INTERNAL_HDRS,
1346 copts = xnnpack_std_copts(),
1347 x86_copts = ["-msse2"],
1348 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001349 deps = [
1350 "@FP16",
1351 "@pthreadpool",
1352 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001353)
1354
1355xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001356 name = "sse41_ukernels",
1357 hdrs = INTERNAL_HDRS,
1358 copts = xnnpack_std_copts(),
1359 x86_copts = ["-msse4.1"],
1360 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001361 deps = [
1362 "@FP16",
1363 "@pthreadpool",
1364 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001365)
1366
1367xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001368 name = "avx_ukernels",
1369 hdrs = INTERNAL_HDRS,
1370 copts = xnnpack_std_copts(),
1371 x86_copts = ["-mavx"],
1372 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001373 deps = [
1374 "@FP16",
1375 "@pthreadpool",
1376 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001377)
1378
1379xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001380 name = "fma3_ukernels",
1381 hdrs = INTERNAL_HDRS,
1382 copts = xnnpack_std_copts(),
1383 x86_copts = [
1384 "-mfma",
1385 ],
1386 x86_srcs = FMA3_UKERNELS,
1387 deps = [
1388 "@FP16",
1389 "@pthreadpool",
1390 ],
1391)
1392
1393xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001394 name = "avx2_ukernels",
1395 hdrs = INTERNAL_HDRS,
1396 copts = xnnpack_std_copts(),
1397 x86_copts = [
1398 "-mfma",
1399 "-mavx2",
1400 ],
1401 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001402 deps = [
1403 "@FP16",
1404 "@pthreadpool",
1405 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001406)
1407
1408xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001409 name = "avx512f_ukernels",
1410 hdrs = INTERNAL_HDRS,
1411 copts = xnnpack_std_copts(),
1412 x86_copts = ["-mavx512f"],
1413 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001414 deps = [
1415 "@FP16",
1416 "@pthreadpool",
1417 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001418)
1419
1420xnnpack_cc_library(
1421 name = "asm_ukernels",
1422 hdrs = ["src/xnnpack/assembly.h"],
1423 aarch32_srcs = AARCH32_ASM_UKERNELS,
1424 aarch64_srcs = AARCH64_ASM_UKERNELS,
1425)
1426
1427xnnpack_aggregate_library(
1428 name = "ukernels",
1429 aarch32_deps = [
1430 ":psimd_ukernels",
1431 ":neon_ukernels",
1432 ":neonfma_ukernels",
1433 ":asm_ukernels",
1434 ],
1435 aarch64_deps = [
1436 ":psimd_ukernels",
1437 ":neon_ukernels",
1438 ":neonfma_ukernels",
1439 ":neonfp16arith_ukernels",
1440 ":asm_ukernels",
1441 ],
1442 generic_deps = [":scalar_ukernels"],
Marat Dukhan436ebe62019-12-04 15:10:12 -08001443 wasm_deps = [
1444 ":wasm_ukernels",
1445 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001446 wasmsimd_deps = [
Marat Dukhan436ebe62019-12-04 15:10:12 -08001447 ":wasm_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001448 ":psimd_ukernels",
1449 ],
1450 x86_deps = [
1451 ":psimd_ukernels",
1452 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001453 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001454 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001455 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001456 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001457 ":avx512f_ukernels",
1458 ],
1459)
1460
1461xnnpack_cc_library(
1462 name = "im2col",
1463 srcs = ["src/im2col.c"],
1464 hdrs = [
1465 "src/xnnpack/common.h",
1466 "src/xnnpack/im2col.h",
1467 ],
1468 copts = xnnpack_std_copts(),
1469)
1470
1471xnnpack_cc_library(
1472 name = "indirection",
1473 srcs = ["src/indirection.c"],
1474 hdrs = INTERNAL_HDRS,
1475 copts = xnnpack_std_copts(),
1476 deps = [
1477 "@FP16",
1478 "@FXdiv",
1479 "@pthreadpool",
1480 ],
1481)
1482
1483xnnpack_cc_library(
1484 name = "operator_run",
1485 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001486 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
1487 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001488 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
1489 "-Wno-vla",
1490 ],
1491 deps = [
1492 "@FP16",
1493 "@FXdiv",
1494 "@clog",
1495 "@pthreadpool",
1496 ],
1497)
1498
1499cc_library(
1500 name = "enable_assembly",
1501 defines = select({
1502 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1503 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001504 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001505 }),
1506)
1507
Marat Dukhancf056b22019-10-07 10:26:29 -07001508xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001509 name = "operators",
1510 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001511 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001512 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07001513 ],
1514 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001515 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001516 "-Isrc",
1517 "-Iinclude",
1518 ] + select({
1519 ":debug_build": [],
1520 "//conditions:default": xnnpack_min_size_copts(),
1521 }),
Marat Dukhancf056b22019-10-07 10:26:29 -07001522 wasm_srcs = ["src/wasm-stubs.c"],
1523 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001524 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001525 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001526 "@FP16",
1527 "@FXdiv",
1528 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001529 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001530 ],
1531)
1532
1533cc_library(
1534 name = "XNNPACK",
1535 srcs = [
1536 "src/init.c",
1537 ],
1538 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1539 "-Isrc",
1540 "-Iinclude",
1541 ] + select({
1542 ":debug_build": [],
1543 "//conditions:default": xnnpack_min_size_copts(),
1544 }),
1545 includes = ["include"],
1546 linkstatic = True,
1547 textual_hdrs = ["include/xnnpack.h"],
1548 visibility = xnnpack_visibility(),
1549 deps = [
1550 ":enable_assembly",
1551 ":ukernels",
1552 ":operator_run",
1553 ":operators",
1554 "@clog",
1555 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07001556 ] + select({
1557 ":emscripten": [],
1558 "//conditions:default": ["@cpuinfo"],
1559 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001560)
1561
1562cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001563 name = "xnnpack_operators_nhwc_f32",
1564 srcs = [
1565 "src/init.c",
1566 ],
1567 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1568 "-Isrc",
1569 "-Iinclude",
1570 ] + select({
1571 ":debug_build": [],
1572 "//conditions:default": xnnpack_min_size_copts(),
1573 }),
1574 defines = [
1575 "XNN_NO_Q8_OPERATORS",
1576 "XNN_NO_U8_OPERATORS",
1577 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08001578 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001579 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001580 includes = ["include"],
1581 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001582 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001583 visibility = xnnpack_visibility(),
1584 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001585 ":enable_assembly",
1586 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001587 ":operator_run",
1588 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001589 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001590 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001591 ] + select({
1592 ":emscripten": [],
1593 "//conditions:default": ["@cpuinfo"],
1594 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001595)
1596
Marat Dukhancf056b22019-10-07 10:26:29 -07001597xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001598 name = "bench_utils",
1599 srcs = ["bench/utils.cc"],
1600 hdrs = ["bench/utils.h"],
1601 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08001602 deps = [
1603 "@com_google_benchmark//:benchmark",
1604 "@cpuinfo",
1605 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001606)
1607
Frank Barchard7e955972019-10-11 10:34:25 -07001608######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07001609
1610xnnpack_benchmark(
1611 name = "q8_gemm_bench",
1612 srcs = [
1613 "bench/gemm.h",
1614 "bench/q8-gemm.cc",
1615 "src/xnnpack/AlignedAllocator.h",
1616 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1617 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
1618 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
1619)
1620
1621xnnpack_benchmark(
1622 name = "f16_gemm_bench",
1623 srcs = [
1624 "bench/f16-gemm.cc",
1625 "bench/gemm.h",
1626 "src/xnnpack/AlignedAllocator.h",
1627 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1628 copts = ["-Wno-unused-function"],
1629 deps = MICROKERNEL_BENCHMARK_DEPS,
1630)
1631
1632xnnpack_benchmark(
1633 name = "f32_igemm_bench",
1634 srcs = [
1635 "bench/f32-igemm.cc",
1636 "bench/conv.h",
1637 "src/xnnpack/AlignedAllocator.h",
1638 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07001639 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001640)
1641
1642xnnpack_benchmark(
1643 name = "f32_conv_hwc_bench",
1644 srcs = [
1645 "bench/f32-conv-hwc.cc",
1646 "bench/dconv.h",
1647 "src/xnnpack/AlignedAllocator.h",
1648 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1649 copts = ["-Wno-unused-function"],
1650 deps = MICROKERNEL_BENCHMARK_DEPS,
1651)
1652
1653xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07001654 name = "f32_conv_hwc2spchw_bench",
1655 srcs = [
1656 "bench/f32-conv-hwc2spchw.cc",
1657 "bench/dconv.h",
1658 "src/xnnpack/AlignedAllocator.h",
1659 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1660 copts = ["-Wno-unused-function"],
1661 deps = MICROKERNEL_BENCHMARK_DEPS,
1662)
1663
1664xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001665 name = "f32_dwconv_bench",
1666 srcs = [
1667 "bench/f32-dwconv.cc",
1668 "bench/dwconv.h",
1669 "src/xnnpack/AlignedAllocator.h",
1670 ] + MICROKERNEL_BENCHMARK_HDRS,
1671 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1672)
1673
1674xnnpack_benchmark(
1675 name = "f32_dwconv_spchw_bench",
1676 srcs = [
1677 "bench/f32-dwconv-spchw.cc",
1678 "bench/dwconv.h",
1679 "src/xnnpack/AlignedAllocator.h",
1680 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1681 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1682)
1683
1684xnnpack_benchmark(
1685 name = "f32_gemm_bench",
1686 srcs = [
1687 "bench/f32-gemm.cc",
1688 "bench/gemm.h",
1689 "src/xnnpack/AlignedAllocator.h",
1690 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1691 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07001692 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001693)
1694
1695xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001696 name = "f32_raddexpminusmax_bench",
1697 srcs = [
1698 "bench/f32-raddexpminusmax.cc",
1699 "src/xnnpack/AlignedAllocator.h",
1700 ] + MICROKERNEL_BENCHMARK_HDRS,
1701 deps = MICROKERNEL_BENCHMARK_DEPS,
1702)
1703
1704xnnpack_benchmark(
1705 name = "f32_raddextexp_bench",
1706 srcs = [
1707 "bench/f32-raddextexp.cc",
1708 "src/xnnpack/AlignedAllocator.h",
1709 ] + MICROKERNEL_BENCHMARK_HDRS,
1710 deps = MICROKERNEL_BENCHMARK_DEPS,
1711)
1712
1713xnnpack_benchmark(
1714 name = "f32_raddstoreexpminusmax_bench",
1715 srcs = [
1716 "bench/f32-raddstoreexpminusmax.cc",
1717 "src/xnnpack/AlignedAllocator.h",
1718 ] + MICROKERNEL_BENCHMARK_HDRS,
1719 deps = MICROKERNEL_BENCHMARK_DEPS,
1720)
1721
1722xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001723 name = "f32_rmax_bench",
1724 srcs = [
1725 "bench/f32-rmax.cc",
1726 "src/xnnpack/AlignedAllocator.h",
1727 ] + MICROKERNEL_BENCHMARK_HDRS,
1728 deps = MICROKERNEL_BENCHMARK_DEPS,
1729)
1730
1731xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08001732 name = "f32_sigmoid_bench",
1733 srcs = [
1734 "bench/f32-sigmoid.cc",
1735 "src/xnnpack/AlignedAllocator.h",
1736 ] + MICROKERNEL_BENCHMARK_HDRS,
1737 copts = ["-Wno-unused-function"],
1738 deps = MICROKERNEL_BENCHMARK_DEPS,
1739)
1740
1741xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001742 name = "f32_spmm_bench",
1743 srcs = [
1744 "bench/f32-spmm.cc",
1745 "bench/gemm.h",
1746 "src/xnnpack/AlignedAllocator.h",
1747 ] + MICROKERNEL_BENCHMARK_HDRS,
1748 copts = ["-Wno-unused-function"],
1749 deps = MICROKERNEL_BENCHMARK_DEPS,
1750)
1751
1752xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07001753 name = "f32_softargmax_bench",
1754 srcs = [
1755 "bench/f32-softargmax.cc",
1756 ] + MICROKERNEL_BENCHMARK_HDRS,
1757 copts = ["-Wno-unused-function"],
1758 deps = MICROKERNEL_BENCHMARK_DEPS,
1759)
1760
1761xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001762 name = "f32_vscaleexpminusmax_bench",
1763 srcs = [
1764 "bench/f32-vscaleexpminusmax.cc",
1765 "src/xnnpack/AlignedAllocator.h",
1766 ] + MICROKERNEL_BENCHMARK_HDRS,
1767 deps = MICROKERNEL_BENCHMARK_DEPS,
1768)
1769
1770xnnpack_benchmark(
1771 name = "f32_vscaleextexp_bench",
1772 srcs = [
1773 "bench/f32-vscaleextexp.cc",
1774 "src/xnnpack/AlignedAllocator.h",
1775 ] + MICROKERNEL_BENCHMARK_HDRS,
1776 deps = MICROKERNEL_BENCHMARK_DEPS,
1777)
1778
1779xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001780 name = "f32_im2col_gemm_bench",
1781 srcs = [
1782 "bench/f32-im2col-gemm.cc",
1783 "bench/conv.h",
1784 "src/xnnpack/AlignedAllocator.h",
1785 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1786 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
1787)
1788
1789########################### Benchmarks for operators ###########################
1790
1791xnnpack_benchmark(
1792 name = "add_bench",
1793 srcs = ["bench/add.cc"],
1794 deps = OPERATOR_BENCHMARK_DEPS,
1795)
1796
1797xnnpack_benchmark(
1798 name = "average_pooling_bench",
1799 srcs = ["bench/average-pooling.cc"],
1800 deps = OPERATOR_BENCHMARK_DEPS,
1801)
1802
1803xnnpack_benchmark(
1804 name = "channel_shuffle_bench",
1805 srcs = ["bench/channel-shuffle.cc"],
1806 deps = OPERATOR_BENCHMARK_DEPS,
1807)
1808
1809xnnpack_benchmark(
1810 name = "convolution_bench",
1811 srcs = ["bench/convolution.cc"],
1812 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1813 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1814)
1815
1816xnnpack_benchmark(
1817 name = "deconvolution_bench",
1818 srcs = ["bench/deconvolution.cc"],
1819 copts = xnnpack_optional_tflite_copts(),
1820 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1821)
1822
1823xnnpack_benchmark(
1824 name = "global_average_pooling_bench",
1825 srcs = ["bench/global-average-pooling.cc"],
1826 deps = OPERATOR_BENCHMARK_DEPS,
1827)
1828
1829xnnpack_benchmark(
1830 name = "max_pooling_bench",
1831 srcs = ["bench/max-pooling.cc"],
1832 deps = OPERATOR_BENCHMARK_DEPS,
1833)
1834
1835xnnpack_benchmark(
1836 name = "sigmoid_bench",
1837 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08001838 copts = xnnpack_optional_tflite_copts(),
1839 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001840)
1841
1842xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07001843 name = "prelu_bench",
1844 srcs = ["bench/prelu.cc"],
1845 copts = xnnpack_optional_tflite_copts(),
1846 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1847)
1848
1849xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001850 name = "softargmax_bench",
1851 srcs = ["bench/softargmax.cc"],
1852 deps = OPERATOR_BENCHMARK_DEPS,
1853)
1854
Marat Dukhanc068bb62019-10-04 13:24:39 -07001855############################# End-to-end benchmarks ############################
1856
1857cc_library(
1858 name = "mobilenet_v1",
1859 srcs = ["models/mobilenet-v1.cc"],
1860 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08001861 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07001862 linkstatic = True,
1863 deps = [
1864 ":XNNPACK",
1865 "@pthreadpool",
1866 ],
1867)
1868
1869cc_library(
1870 name = "mobilenet_v2",
1871 srcs = ["models/mobilenet-v2.cc"],
1872 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08001873 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07001874 linkstatic = True,
1875 deps = [
1876 ":XNNPACK",
1877 "@pthreadpool",
1878 ],
1879)
1880
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001881cc_library(
1882 name = "mobilenet_v3_large",
1883 srcs = ["models/mobilenet-v3-large.cc"],
1884 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08001885 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001886 linkstatic = True,
1887 deps = [
1888 ":XNNPACK",
1889 "@pthreadpool",
1890 ],
1891)
1892
1893cc_library(
1894 name = "mobilenet_v3_small",
1895 srcs = ["models/mobilenet-v3-small.cc"],
1896 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08001897 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001898 linkstatic = True,
1899 deps = [
1900 ":XNNPACK",
1901 "@pthreadpool",
1902 ],
1903)
1904
Marat Dukhanc068bb62019-10-04 13:24:39 -07001905xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07001906 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001907 srcs = [
1908 "bench/f32-dwconv-e2e.cc",
1909 "bench/end2end.h",
1910 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07001911 copts = ["-Wno-unused-function"],
1912 deps = MICROKERNEL_BENCHMARK_DEPS + [
1913 ":XNNPACK",
1914 ":mobilenet_v1",
1915 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001916 ":mobilenet_v3_large",
1917 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07001918 ],
1919)
1920
1921xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07001922 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001923 srcs = [
1924 "bench/f32-gemm-e2e.cc",
1925 "bench/end2end.h",
1926 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07001927 copts = ["-Wno-unused-function"],
1928 deps = MICROKERNEL_BENCHMARK_DEPS + [
1929 ":XNNPACK",
1930 ":mobilenet_v1",
1931 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001932 ":mobilenet_v3_large",
1933 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07001934 ],
1935)
1936
1937xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07001938 name = "end2end_bench",
1939 srcs = ["bench/end2end.cc"],
1940 deps = [
1941 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07001942 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07001943 ":mobilenet_v1",
1944 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08001945 ":mobilenet_v3_large",
1946 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07001947 "@pthreadpool",
1948 ],
1949)
1950
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001951#################### Accuracy evaluation for math functions ####################
1952
1953xnnpack_benchmark(
1954 name = "f32_exp_eval",
1955 srcs = [
1956 "eval/f32-exp.cc",
1957 "src/xnnpack/AlignedAllocator.h",
1958 ] + ACCURACY_EVAL_HDRS,
1959 deps = ACCURACY_EVAL_DEPS,
1960)
1961
Marat Dukhan515c9772019-10-17 18:07:57 -07001962xnnpack_benchmark(
1963 name = "f32_expminus_eval",
1964 srcs = [
1965 "eval/f32-expminus.cc",
1966 "src/xnnpack/AlignedAllocator.h",
1967 ] + ACCURACY_EVAL_HDRS,
1968 deps = ACCURACY_EVAL_DEPS,
1969)
1970
Marat Dukhan98ba4412019-10-23 02:14:28 -07001971xnnpack_benchmark(
1972 name = "f32_extexp_eval",
1973 srcs = [
1974 "eval/f32-extexp.cc",
1975 "src/xnnpack/AlignedAllocator.h",
1976 ] + ACCURACY_EVAL_HDRS,
1977 deps = ACCURACY_EVAL_DEPS,
1978)
1979
Marat Dukhan346a9e52019-11-15 09:06:30 -08001980xnnpack_benchmark(
1981 name = "f32_sigmoid_eval",
1982 srcs = [
1983 "eval/f32-sigmoid.cc",
1984 "src/xnnpack/AlignedAllocator.h",
1985 ] + ACCURACY_EVAL_HDRS,
1986 deps = ACCURACY_EVAL_DEPS,
1987)
1988
Marat Dukhan08c4a432019-10-03 09:29:21 -07001989######################### Unit tests for micro-kernels #########################
1990
1991xnnpack_unit_test(
1992 name = "f16_gemm_test",
1993 srcs = [
1994 "test/f16-gemm.cc",
1995 "test/gemm-microkernel-tester.h",
1996 "src/xnnpack/AlignedAllocator.h",
1997 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1998 deps = MICROKERNEL_TEST_DEPS,
1999)
2000
2001xnnpack_unit_test(
2002 name = "f32_argmaxpool_test",
2003 srcs = [
2004 "test/f32-argmaxpool.cc",
2005 "test/argmaxpool-microkernel-tester.h",
2006 "src/xnnpack/AlignedAllocator.h",
2007 ] + MICROKERNEL_TEST_HDRS,
2008 deps = MICROKERNEL_TEST_DEPS,
2009)
2010
2011xnnpack_unit_test(
2012 name = "f32_avgpool_test",
2013 srcs = [
2014 "test/f32-avgpool.cc",
2015 "test/avgpool-microkernel-tester.h",
2016 "src/xnnpack/AlignedAllocator.h",
2017 ] + MICROKERNEL_TEST_HDRS,
2018 deps = MICROKERNEL_TEST_DEPS,
2019)
2020
2021xnnpack_unit_test(
Marat Dukhan35dacfb2019-11-07 19:18:16 -08002022 name = "f32_bilinear_test",
2023 srcs = [
2024 "test/f32-bilinear.cc",
2025 "test/bilinear-microkernel-tester.h",
2026 "src/xnnpack/AlignedAllocator.h",
2027 ] + MICROKERNEL_TEST_HDRS,
2028 deps = MICROKERNEL_TEST_DEPS,
2029)
2030
2031xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002032 name = "f32_clamp_test",
2033 srcs = [
2034 "test/f32-clamp.cc",
2035 "test/clamp-microkernel-tester.h",
2036 ] + MICROKERNEL_TEST_HDRS,
2037 deps = MICROKERNEL_TEST_DEPS,
2038)
2039
2040xnnpack_unit_test(
2041 name = "f32_igemm_test",
2042 srcs = [
2043 "test/f32-igemm.cc",
2044 "test/gemm-microkernel-tester.h",
2045 "src/xnnpack/AlignedAllocator.h",
2046 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2047 deps = MICROKERNEL_TEST_DEPS,
2048)
2049
2050xnnpack_unit_test(
2051 name = "f32_conv_hwc_test",
2052 srcs = [
2053 "test/f32-conv-hwc.cc",
2054 "test/conv-hwc-microkernel-tester.h",
2055 "src/xnnpack/AlignedAllocator.h",
2056 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2057 deps = MICROKERNEL_TEST_DEPS,
2058)
2059
2060xnnpack_unit_test(
2061 name = "f32_conv_hwc2spchw_test",
2062 srcs = [
2063 "test/f32-conv-hwc2spchw.cc",
2064 "test/conv-hwc2spchw-microkernel-tester.h",
2065 "src/xnnpack/AlignedAllocator.h",
2066 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2067 deps = MICROKERNEL_TEST_DEPS,
2068)
2069
2070xnnpack_unit_test(
2071 name = "f32_dwconv_test",
2072 srcs = [
2073 "test/f32-dwconv.cc",
2074 "test/dwconv-microkernel-tester.h",
2075 "src/xnnpack/AlignedAllocator.h",
2076 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2077 deps = MICROKERNEL_TEST_DEPS,
2078)
2079
2080xnnpack_unit_test(
2081 name = "f32_dwconv_spchw_test",
2082 srcs = [
2083 "test/f32-dwconv-spchw.cc",
2084 "test/dwconv-spchw-microkernel-tester.h",
2085 "src/xnnpack/AlignedAllocator.h",
2086 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2087 deps = MICROKERNEL_TEST_DEPS,
2088)
2089
2090xnnpack_unit_test(
2091 name = "f32_gavgpool_test",
2092 srcs = [
2093 "test/f32-gavgpool.cc",
2094 "test/gavgpool-microkernel-tester.h",
2095 "src/xnnpack/AlignedAllocator.h",
2096 ] + MICROKERNEL_TEST_HDRS,
2097 deps = MICROKERNEL_TEST_DEPS,
2098)
2099
2100xnnpack_unit_test(
2101 name = "f32_gavgpool_spchw_test",
2102 srcs = [
2103 "test/f32-gavgpool-spchw.cc",
2104 "test/gavgpool-spchw-microkernel-tester.h",
2105 "src/xnnpack/AlignedAllocator.h",
2106 ] + MICROKERNEL_TEST_HDRS,
2107 deps = MICROKERNEL_TEST_DEPS,
2108)
2109
2110xnnpack_unit_test(
2111 name = "f32_gemm_test",
2112 srcs = [
2113 "test/f32-gemm.cc",
2114 "test/gemm-microkernel-tester.h",
2115 "src/xnnpack/AlignedAllocator.h",
2116 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2117 deps = MICROKERNEL_TEST_DEPS,
2118)
2119
2120xnnpack_unit_test(
2121 name = "f32_gemminc_test",
2122 srcs = [
2123 "test/f32-gemminc.cc",
2124 "test/gemm-microkernel-tester.h",
2125 "src/xnnpack/AlignedAllocator.h",
2126 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2127 deps = MICROKERNEL_TEST_DEPS,
2128)
2129
2130xnnpack_unit_test(
2131 name = "f32_hswish_test",
2132 srcs = [
2133 "test/f32-hswish.cc",
2134 "test/hswish-microkernel-tester.h",
2135 ] + MICROKERNEL_TEST_HDRS,
2136 deps = MICROKERNEL_TEST_DEPS,
2137)
2138
2139xnnpack_unit_test(
2140 name = "f32_maxpool_test",
2141 srcs = [
2142 "test/f32-maxpool.cc",
2143 "test/maxpool-microkernel-tester.h",
2144 ] + MICROKERNEL_TEST_HDRS,
2145 deps = MICROKERNEL_TEST_DEPS,
2146)
2147
2148xnnpack_unit_test(
2149 name = "f32_pavgpool_test",
2150 srcs = [
2151 "test/f32-pavgpool.cc",
2152 "test/avgpool-microkernel-tester.h",
2153 "src/xnnpack/AlignedAllocator.h",
2154 ] + MICROKERNEL_TEST_HDRS,
2155 deps = MICROKERNEL_TEST_DEPS,
2156)
2157
2158xnnpack_unit_test(
2159 name = "f32_ppmm_test",
2160 srcs = [
2161 "test/f32-ppmm.cc",
2162 "test/gemm-microkernel-tester.h",
2163 "src/xnnpack/AlignedAllocator.h",
2164 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2165 deps = MICROKERNEL_TEST_DEPS,
2166)
2167
2168xnnpack_unit_test(
2169 name = "f32_prelu_test",
2170 srcs = [
2171 "test/f32-prelu.cc",
2172 "test/prelu-microkernel-tester.h",
2173 "src/xnnpack/AlignedAllocator.h",
2174 ] + MICROKERNEL_TEST_HDRS,
2175 deps = MICROKERNEL_TEST_DEPS,
2176)
2177
2178xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002179 name = "f32_raddexpminusmax_test",
2180 srcs = [
2181 "test/f32-raddexpminusmax.cc",
2182 "test/raddexpminusmax-microkernel-tester.h",
2183 ] + MICROKERNEL_TEST_HDRS,
2184 deps = MICROKERNEL_TEST_DEPS,
2185)
2186
2187xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002188 name = "f32_raddextexp_test",
2189 srcs = [
2190 "test/f32-raddextexp.cc",
2191 "test/raddextexp-microkernel-tester.h",
2192 ] + MICROKERNEL_TEST_HDRS,
2193 deps = MICROKERNEL_TEST_DEPS,
2194)
2195
2196xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002197 name = "f32_raddstoreexpminusmax_test",
2198 srcs = [
2199 "test/f32-raddstoreexpminusmax.cc",
2200 "test/raddstoreexpminusmax-microkernel-tester.h",
2201 ] + MICROKERNEL_TEST_HDRS,
2202 deps = MICROKERNEL_TEST_DEPS,
2203)
2204
2205xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002206 name = "f32_rmax_test",
2207 srcs = [
2208 "test/f32-rmax.cc",
2209 "test/rmax-microkernel-tester.h",
2210 ] + MICROKERNEL_TEST_HDRS,
2211 deps = MICROKERNEL_TEST_DEPS,
2212)
2213
2214xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08002215 name = "f32_sigmoid_test",
2216 srcs = [
2217 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002218 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08002219 ] + MICROKERNEL_TEST_HDRS,
2220 deps = MICROKERNEL_TEST_DEPS,
2221)
2222
2223xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002224 name = "f32_spmm_test",
2225 srcs = [
2226 "test/f32-spmm.cc",
2227 "test/spmm-microkernel-tester.h",
2228 "src/xnnpack/AlignedAllocator.h",
2229 ] + MICROKERNEL_TEST_HDRS,
2230 deps = MICROKERNEL_TEST_DEPS,
2231)
2232
2233xnnpack_unit_test(
2234 name = "f32_vadd_test",
2235 srcs = [
2236 "test/f32-vadd.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002237 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002238 ] + MICROKERNEL_TEST_HDRS,
2239 deps = MICROKERNEL_TEST_DEPS,
2240)
2241
2242xnnpack_unit_test(
2243 name = "f32_vaddc_test",
2244 srcs = [
2245 "test/f32-vaddc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002246 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002247 ] + MICROKERNEL_TEST_HDRS,
2248 deps = MICROKERNEL_TEST_DEPS,
2249)
2250
2251xnnpack_unit_test(
Marat Dukhan77ca6302019-12-06 12:48:15 -08002252 name = "f32_vdiv_test",
2253 srcs = [
2254 "test/f32-vdiv.cc",
2255 "test/vbinary-microkernel-tester.h",
2256 ] + MICROKERNEL_TEST_HDRS,
2257 deps = MICROKERNEL_TEST_DEPS,
2258)
2259
2260xnnpack_unit_test(
2261 name = "f32_vdivc_test",
2262 srcs = [
2263 "test/f32-vdivc.cc",
2264 "test/vbinaryc-microkernel-tester.h",
2265 ] + MICROKERNEL_TEST_HDRS,
2266 deps = MICROKERNEL_TEST_DEPS,
2267)
2268
2269xnnpack_unit_test(
2270 name = "f32_vrdivc_test",
2271 srcs = [
2272 "test/f32-vrdivc.cc",
2273 "test/vbinaryc-microkernel-tester.h",
2274 ] + MICROKERNEL_TEST_HDRS,
2275 deps = MICROKERNEL_TEST_DEPS,
2276)
2277
2278xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08002279 name = "f32_vmax_test",
2280 srcs = [
2281 "test/f32-vmax.cc",
2282 "test/vbinary-microkernel-tester.h",
2283 ] + MICROKERNEL_TEST_HDRS,
2284 deps = MICROKERNEL_TEST_DEPS,
2285)
2286
2287xnnpack_unit_test(
2288 name = "f32_vmaxc_test",
2289 srcs = [
2290 "test/f32-vmaxc.cc",
2291 "test/vbinaryc-microkernel-tester.h",
2292 ] + MICROKERNEL_TEST_HDRS,
2293 deps = MICROKERNEL_TEST_DEPS,
2294)
2295
2296xnnpack_unit_test(
2297 name = "f32_vmin_test",
2298 srcs = [
2299 "test/f32-vmin.cc",
2300 "test/vbinary-microkernel-tester.h",
2301 ] + MICROKERNEL_TEST_HDRS,
2302 deps = MICROKERNEL_TEST_DEPS,
2303)
2304
2305xnnpack_unit_test(
2306 name = "f32_vminc_test",
2307 srcs = [
2308 "test/f32-vminc.cc",
2309 "test/vbinaryc-microkernel-tester.h",
2310 ] + MICROKERNEL_TEST_HDRS,
2311 deps = MICROKERNEL_TEST_DEPS,
2312)
2313
2314xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002315 name = "f32_vmul_test",
2316 srcs = [
2317 "test/f32-vmul.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002318 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002319 ] + MICROKERNEL_TEST_HDRS,
2320 deps = MICROKERNEL_TEST_DEPS,
2321)
2322
2323xnnpack_unit_test(
2324 name = "f32_vmulc_test",
2325 srcs = [
2326 "test/f32-vmulc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002327 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002328 ] + MICROKERNEL_TEST_HDRS,
2329 deps = MICROKERNEL_TEST_DEPS,
2330)
2331
2332xnnpack_unit_test(
2333 name = "f32_vmulcaddc_test",
2334 srcs = [
2335 "test/f32-vmulcaddc.cc",
2336 "test/vmulcaddc-microkernel-tester.h",
2337 "src/xnnpack/AlignedAllocator.h",
2338 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2339 deps = MICROKERNEL_TEST_DEPS,
2340)
2341
2342xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002343 name = "f32_vscale_test",
2344 srcs = [
2345 "test/f32-vscale.cc",
2346 "test/vscale-microkernel-tester.h",
2347 ] + MICROKERNEL_TEST_HDRS,
2348 deps = MICROKERNEL_TEST_DEPS,
2349)
2350
2351xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002352 name = "f32_vscaleexpminusmax_test",
2353 srcs = [
2354 "test/f32-vscaleexpminusmax.cc",
2355 "test/vscaleexpminusmax-microkernel-tester.h",
2356 ] + MICROKERNEL_TEST_HDRS,
2357 deps = MICROKERNEL_TEST_DEPS,
2358)
2359
2360xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002361 name = "f32_vscaleextexp_test",
2362 srcs = [
2363 "test/f32-vscaleextexp.cc",
2364 "test/vscaleextexp-microkernel-tester.h",
2365 ] + MICROKERNEL_TEST_HDRS,
2366 deps = MICROKERNEL_TEST_DEPS,
2367)
2368
2369xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07002370 name = "f32_vsub_test",
2371 srcs = [
2372 "test/f32-vsub.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002373 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002374 ] + MICROKERNEL_TEST_HDRS,
2375 deps = MICROKERNEL_TEST_DEPS,
2376)
2377
2378xnnpack_unit_test(
2379 name = "f32_vsubc_test",
2380 srcs = [
2381 "test/f32-vsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002382 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08002383 ] + MICROKERNEL_TEST_HDRS,
2384 deps = MICROKERNEL_TEST_DEPS,
2385)
2386
2387xnnpack_unit_test(
2388 name = "f32_vrsubc_test",
2389 srcs = [
2390 "test/f32-vrsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002391 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002392 ] + MICROKERNEL_TEST_HDRS,
2393 deps = MICROKERNEL_TEST_DEPS,
2394)
2395
2396xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002397 name = "q8_avgpool_test",
2398 srcs = [
2399 "test/q8-avgpool.cc",
2400 "test/avgpool-microkernel-tester.h",
2401 "src/xnnpack/AlignedAllocator.h",
2402 ] + MICROKERNEL_TEST_HDRS,
2403 deps = MICROKERNEL_TEST_DEPS,
2404)
2405
2406xnnpack_unit_test(
2407 name = "q8_igemm_test",
2408 srcs = [
2409 "test/q8-igemm.cc",
2410 "test/gemm-microkernel-tester.h",
2411 "src/xnnpack/AlignedAllocator.h",
2412 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2413 deps = MICROKERNEL_TEST_DEPS,
2414)
2415
2416xnnpack_unit_test(
2417 name = "q8_dwconv_test",
2418 srcs = [
2419 "test/q8-dwconv.cc",
2420 "test/dwconv-microkernel-tester.h",
2421 "src/xnnpack/AlignedAllocator.h",
2422 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2423 deps = MICROKERNEL_TEST_DEPS,
2424)
2425
2426xnnpack_unit_test(
2427 name = "q8_gavgpool_test",
2428 srcs = [
2429 "test/q8-gavgpool.cc",
2430 "test/gavgpool-microkernel-tester.h",
2431 "src/xnnpack/AlignedAllocator.h",
2432 ] + MICROKERNEL_TEST_HDRS,
2433 deps = MICROKERNEL_TEST_DEPS,
2434)
2435
2436xnnpack_unit_test(
2437 name = "q8_gemm_test",
2438 srcs = [
2439 "test/q8-gemm.cc",
2440 "test/gemm-microkernel-tester.h",
2441 "src/xnnpack/AlignedAllocator.h",
2442 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
2443 deps = MICROKERNEL_TEST_DEPS,
2444)
2445
2446xnnpack_unit_test(
2447 name = "q8_vadd_test",
2448 srcs = [
2449 "test/q8-vadd.cc",
2450 "test/vadd-microkernel-tester.h",
2451 ] + MICROKERNEL_TEST_HDRS,
2452 deps = MICROKERNEL_TEST_DEPS,
2453)
2454
2455xnnpack_unit_test(
2456 name = "u8_clamp_test",
2457 srcs = [
2458 "test/u8-clamp.cc",
2459 "test/clamp-microkernel-tester.h",
2460 ] + MICROKERNEL_TEST_HDRS,
2461 deps = MICROKERNEL_TEST_DEPS,
2462)
2463
2464xnnpack_unit_test(
2465 name = "u8_lut32norm_test",
2466 srcs = [
2467 "test/u8-lut32norm.cc",
2468 "test/lut-norm-microkernel-tester.h",
2469 ] + MICROKERNEL_TEST_HDRS,
2470 deps = MICROKERNEL_TEST_DEPS,
2471)
2472
2473xnnpack_unit_test(
2474 name = "u8_maxpool_test",
2475 srcs = [
2476 "test/u8-maxpool.cc",
2477 "test/maxpool-microkernel-tester.h",
2478 ] + MICROKERNEL_TEST_HDRS,
2479 deps = MICROKERNEL_TEST_DEPS,
2480)
2481
2482xnnpack_unit_test(
2483 name = "u8_rmax_test",
2484 srcs = [
2485 "test/u8-rmax.cc",
2486 "test/rmax-microkernel-tester.h",
2487 ] + MICROKERNEL_TEST_HDRS,
2488 deps = MICROKERNEL_TEST_DEPS,
2489)
2490
2491xnnpack_unit_test(
2492 name = "x32_packx_test",
2493 srcs = [
2494 "test/x32-packx.cc",
2495 "test/pack-microkernel-tester.h",
2496 "src/xnnpack/AlignedAllocator.h",
2497 ] + MICROKERNEL_TEST_HDRS,
2498 deps = MICROKERNEL_TEST_DEPS,
2499)
2500
2501xnnpack_unit_test(
2502 name = "x32_pad_test",
2503 srcs = [
2504 "test/x32-pad.cc",
2505 "test/pad-microkernel-tester.h",
2506 ] + MICROKERNEL_TEST_HDRS,
2507 deps = MICROKERNEL_TEST_DEPS,
2508)
2509
2510xnnpack_unit_test(
2511 name = "x32_unpool_test",
2512 srcs = [
2513 "test/x32-unpool.cc",
2514 "test/unpool-microkernel-tester.h",
2515 ] + MICROKERNEL_TEST_HDRS,
2516 deps = MICROKERNEL_TEST_DEPS,
2517)
2518
2519xnnpack_unit_test(
2520 name = "x32_zip_test",
2521 srcs = [
2522 "test/x32-zip.cc",
2523 "test/zip-microkernel-tester.h",
2524 ] + MICROKERNEL_TEST_HDRS,
2525 deps = MICROKERNEL_TEST_DEPS,
2526)
2527
2528xnnpack_unit_test(
2529 name = "x8_lut_test",
2530 srcs = [
2531 "test/x8-lut.cc",
2532 "test/lut-microkernel-tester.h",
2533 ] + MICROKERNEL_TEST_HDRS,
2534 deps = MICROKERNEL_TEST_DEPS,
2535)
2536
2537xnnpack_unit_test(
2538 name = "x8_zip_test",
2539 srcs = [
2540 "test/x8-zip.cc",
2541 "test/zip-microkernel-tester.h",
2542 ] + MICROKERNEL_TEST_HDRS,
2543 deps = MICROKERNEL_TEST_DEPS,
2544)
2545
2546########################### Size test for the library ##########################
2547
2548xnnpack_binary(
2549 name = "size_test",
2550 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002551 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002552)
2553
2554########################### Unit tests for operators ###########################
2555
2556xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002557 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002558 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002559 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002560 "test/add-operator-tester.h",
2561 ],
2562 deps = OPERATOR_TEST_DEPS,
2563)
2564
2565xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08002566 name = "add_nd_test",
2567 srcs = [
2568 "test/add-nd.cc",
2569 "test/binary-elementwise-operator-tester.h",
2570 ],
2571 deps = OPERATOR_TEST_DEPS,
2572)
2573
2574xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002575 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002576 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002577 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002578 "test/argmax-pooling-operator-tester.h",
2579 ] + OPERATOR_TEST_PARAMS_HDRS,
2580 deps = OPERATOR_TEST_DEPS,
2581)
2582
2583xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002584 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002585 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002586 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002587 "test/average-pooling-operator-tester.h",
2588 ] + OPERATOR_TEST_PARAMS_HDRS,
2589 deps = OPERATOR_TEST_DEPS,
2590)
2591
2592xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002593 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002594 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002595 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002596 "test/channel-pad-operator-tester.h",
2597 ] + OPERATOR_TEST_PARAMS_HDRS,
2598 deps = OPERATOR_TEST_DEPS,
2599)
2600
2601xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002602 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002603 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002604 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002605 "test/channel-shuffle-operator-tester.h",
2606 ],
2607 deps = OPERATOR_TEST_DEPS,
2608)
2609
2610xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002611 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002612 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002613 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002614 "test/clamp-operator-tester.h",
2615 ],
2616 deps = OPERATOR_TEST_DEPS,
2617)
2618
2619xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002620 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002621 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002622 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002623 "test/convolution-operator-tester.h",
2624 ],
2625 deps = OPERATOR_TEST_DEPS,
2626)
2627
2628xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002629 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002630 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002631 "test/convolution-nchw.cc",
2632 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002633 ],
2634 deps = OPERATOR_TEST_DEPS,
2635)
2636
2637xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002638 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002639 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002640 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002641 "test/deconvolution-operator-tester.h",
2642 ] + OPERATOR_TEST_PARAMS_HDRS,
2643 deps = OPERATOR_TEST_DEPS,
2644)
2645
2646xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08002647 name = "divide_nd_test",
2648 srcs = [
2649 "test/binary-elementwise-operator-tester.h",
2650 "test/divide-nd.cc",
2651 ],
2652 deps = OPERATOR_TEST_DEPS,
2653)
2654
2655xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002656 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002657 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002658 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002659 "test/fully-connected-operator-tester.h",
2660 ],
2661 deps = OPERATOR_TEST_DEPS,
2662)
2663
2664xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002665 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002666 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002667 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002668 "test/global-average-pooling-operator-tester.h",
2669 ] + OPERATOR_TEST_PARAMS_HDRS,
2670 deps = OPERATOR_TEST_DEPS,
2671)
2672
2673xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002674 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002675 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002676 "test/global-average-pooling-ncw.cc",
2677 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002678 ],
2679 deps = OPERATOR_TEST_DEPS,
2680)
2681
2682xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002683 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002684 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002685 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002686 "test/hardswish-operator-tester.h",
2687 ],
2688 deps = OPERATOR_TEST_DEPS,
2689)
2690
2691xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002692 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002693 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002694 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002695 "test/leaky-relu-operator-tester.h",
2696 ],
2697 deps = OPERATOR_TEST_DEPS,
2698)
2699
2700xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002701 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002702 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002703 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002704 "test/max-pooling-operator-tester.h",
2705 ] + OPERATOR_TEST_PARAMS_HDRS,
2706 deps = OPERATOR_TEST_DEPS,
2707)
2708
2709xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08002710 name = "maximum_nd_test",
2711 srcs = [
2712 "test/binary-elementwise-operator-tester.h",
2713 "test/maximum-nd.cc",
2714 ],
2715 deps = OPERATOR_TEST_DEPS,
2716)
2717
2718xnnpack_unit_test(
2719 name = "minimum_nd_test",
2720 srcs = [
2721 "test/binary-elementwise-operator-tester.h",
2722 "test/minimum-nd.cc",
2723 ],
2724 deps = OPERATOR_TEST_DEPS,
2725)
2726
2727xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002728 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08002729 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08002730 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002731 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08002732 ],
2733 deps = OPERATOR_TEST_DEPS,
2734)
2735
2736xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002737 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002738 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002739 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002740 "test/prelu-operator-tester.h",
2741 ] + OPERATOR_TEST_PARAMS_HDRS,
2742 deps = OPERATOR_TEST_DEPS,
2743)
2744
2745xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002746 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08002747 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002748 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08002749 "test/resize-bilinear-operator-tester.h",
2750 ] + OPERATOR_TEST_PARAMS_HDRS,
2751 deps = OPERATOR_TEST_DEPS,
2752)
2753
2754xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002755 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002756 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002757 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002758 "test/sigmoid-operator-tester.h",
2759 ],
2760 deps = OPERATOR_TEST_DEPS,
2761)
2762
2763xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002764 name = "softargmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002765 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002766 "test/softargmax-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002767 "test/softargmax-operator-tester.h",
2768 ],
2769 deps = OPERATOR_TEST_DEPS,
2770)
2771
2772xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08002773 name = "subtract_nd_test",
2774 srcs = [
2775 "test/binary-elementwise-operator-tester.h",
2776 "test/subtract-nd.cc",
2777 ],
2778 deps = OPERATOR_TEST_DEPS,
2779)
2780
2781xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002782 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002783 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002784 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002785 "test/unpooling-operator-tester.h",
2786 ],
2787 deps = OPERATOR_TEST_DEPS,
2788)
2789
2790############################# Build configurations #############################
2791
Marat Dukhanb8642352019-10-30 15:43:02 -07002792# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07002793config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07002794 name = "xnn_enable_assembly_explicit_true",
2795 define_values = {"xnn_enable_assembly": "true"},
2796)
2797
2798# Disables usage of assembly kernels.
2799config_setting(
2800 name = "xnn_enable_assembly_explicit_false",
2801 define_values = {"xnn_enable_assembly": "false"},
2802)
2803
2804# Builds with -c dbg
2805config_setting(
2806 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002807 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07002808 "compilation_mode": "dbg",
2809 },
2810)
2811
2812# Builds with -c opt
2813config_setting(
2814 name = "optimized_build",
2815 values = {
2816 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002817 },
2818)
2819
2820config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07002821 name = "linux_k8",
2822 values = {"cpu": "k8"},
2823)
2824
2825config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07002826 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07002827 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07002828)
2829
2830config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002831 name = "android",
2832 values = {"crosstool_top": "//external:android/crosstool"},
2833)
2834
2835config_setting(
2836 name = "android_armv7",
2837 values = {
2838 "crosstool_top": "//external:android/crosstool",
2839 "cpu": "armeabi-v7a",
2840 },
2841)
2842
2843config_setting(
2844 name = "android_arm64",
2845 values = {
2846 "crosstool_top": "//external:android/crosstool",
2847 "cpu": "arm64-v8a",
2848 },
2849)
2850
2851config_setting(
2852 name = "android_x86",
2853 values = {
2854 "crosstool_top": "//external:android/crosstool",
2855 "cpu": "x86",
2856 },
2857)
2858
2859config_setting(
2860 name = "android_x86_64",
2861 values = {
2862 "crosstool_top": "//external:android/crosstool",
2863 "cpu": "x86_64",
2864 },
2865)
2866
2867config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07002868 name = "macos_x86_64",
2869 values = {
2870 "apple_platform_type": "macos",
2871 "cpu": "darwin",
2872 },
2873)
2874
2875config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002876 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002877 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07002878)
2879
2880config_setting(
2881 name = "emscripten_wasm",
2882 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002883 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002884 "cpu": "wasm",
2885 },
2886)
2887
2888config_setting(
2889 name = "emscripten_wasmsimd",
2890 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002891 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002892 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07002893 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002894 },
2895)
2896
2897config_setting(
2898 name = "emscripten_asmjs",
2899 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002900 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002901 "cpu": "asmjs",
2902 },
2903)