blob: a56cd80183b9b62f4a1f1ea74095576e042cb993 [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 Dukhan69c3f2c2019-11-06 12:30:01 -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_unit_test", "xnnpack_visibility")
10
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",
56 "src/channel-pad-nc.c",
57 "src/channel-shuffle-nc.c",
58 "src/clamp-nc.c",
59 "src/convolution-nchw.c",
60 "src/convolution-nhwc.c",
61 "src/deconvolution-nhwc.c",
62 "src/fully-connected-nc.c",
63 "src/global-average-pooling-ncw.c",
64 "src/global-average-pooling-nwc.c",
65 "src/hardswish-nc.c",
66 "src/leaky-relu-nc.c",
67 "src/max-pooling-nhwc.c",
68 "src/multiply-nd.c",
69 "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",
Erich Elsen563df5f2019-10-23 08:02:21 -070086 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen0cc2c532019-10-15 04:44:18 -070087 "src/f32-dwconv-spchw/3x3p1-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -070088 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080089 "src/f32-dwconv-spchw/5x5p2-scalar.c",
Erich Elsen38709a62019-11-08 11:58:45 -080090 "src/f32-dwconv-spchw/5x5s2p2-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -080091 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
92 "src/f32-dwconv/gen/up1x25-scalar.c",
93 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
94 "src/f32-dwconv/gen/up1x4-scalar.c",
95 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
96 "src/f32-dwconv/gen/up1x9-scalar.c",
97 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
98 "src/f32-dwconv/gen/up2x25-scalar.c",
99 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
100 "src/f32-dwconv/gen/up2x4-scalar.c",
101 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
102 "src/f32-dwconv/gen/up2x9-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -0700103 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700104 "src/f32-gavgpool/mp7p7q-scalar.c",
105 "src/f32-gavgpool/up7-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800106 "src/f32-gemm/gen-inc/1x4-scalar.c",
107 "src/f32-gemm/gen-inc/2x4-scalar.c",
108 "src/f32-gemm/gen-inc/4x4-scalar.c",
109 "src/f32-gemm/gen/1x4-scalar.c",
110 "src/f32-gemm/gen/2x4-scalar.c",
111 "src/f32-gemm/gen/4x2-scalar.c",
112 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700113 "src/f32-hswish/scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800114 "src/f32-igemm/gen/1x4-scalar.c",
115 "src/f32-igemm/gen/2x4-scalar.c",
116 "src/f32-igemm/gen/4x2-scalar.c",
117 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-maxpool/9p8x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700119 "src/f32-pavgpool/mp9p8q-scalar.c",
120 "src/f32-pavgpool/up9-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800121 "src/f32-ppmm/gen/2x4-scalar.c",
122 "src/f32-ppmm/gen/3x3-scalar.c",
123 "src/f32-ppmm/gen/4x2-scalar.c",
124 "src/f32-ppmm/gen/4x4-scalar.c",
125 "src/f32-prelu/gen/scalar-2x1.c",
126 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700127 "src/f32-rmax/scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800128 "src/f32-spmm/gen/1x1-scalar-pipelined.c",
129 "src/f32-spmm/gen/1x1-scalar.c",
130 "src/f32-spmm/gen/2x1-scalar-pipelined.c",
131 "src/f32-spmm/gen/2x1-scalar.c",
132 "src/f32-spmm/gen/4x1-scalar-pipelined.c",
133 "src/f32-spmm/gen/4x1-scalar.c",
134 "src/f32-spmm/gen/8x1-scalar-pipelined.c",
135 "src/f32-spmm/gen/8x1-scalar.c",
136 "src/f32-spmm/gen/8x2-scalar.c",
137 "src/f32-spmm/gen/8x4-scalar.c",
138 "src/f32-vbinary/gen/vadd-scalar-x1.c",
139 "src/f32-vbinary/gen/vadd-scalar-x2.c",
140 "src/f32-vbinary/gen/vadd-scalar-x4.c",
141 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
142 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
143 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
144 "src/f32-vbinary/gen/vmul-scalar-x1.c",
145 "src/f32-vbinary/gen/vmul-scalar-x2.c",
146 "src/f32-vbinary/gen/vmul-scalar-x4.c",
147 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
148 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
149 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
150 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
151 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
152 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
153 "src/f32-vbinary/gen/vsub-scalar-x1.c",
154 "src/f32-vbinary/gen/vsub-scalar-x2.c",
155 "src/f32-vbinary/gen/vsub-scalar-x4.c",
156 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
157 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
158 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
159 "src/f32-vmulcaddc/gen/c1-scalar-2x.c",
160 "src/f32-vmulcaddc/gen/c2-scalar-2x.c",
161 "src/f32-vmulcaddc/gen/c4-scalar-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700162 "src/q8-avgpool/mp9p8q-scalar.c",
163 "src/q8-avgpool/up9-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700164 "src/q8-dwconv/up1x9-scalar.c",
165 "src/q8-gavgpool/mp7p7q-scalar.c",
166 "src/q8-gavgpool/up7-scalar.c",
167 "src/q8-gemm/2x2-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800168 "src/q8-igemm/2x2-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700169 "src/q8-vadd/scalar.c",
170 "src/u8-clamp/scalar.c",
171 "src/u8-lut32norm/scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800172 "src/u8-maxpool/9p8x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700173 "src/u8-rmax/scalar.c",
174 "src/x32-packx/x2-scalar.c",
175 "src/x32-packx/x3-scalar.c",
176 "src/x32-packx/x4-scalar.c",
177 "src/x32-pad/x2-scalar.c",
178 "src/x32-unpool/scalar.c",
179 "src/x32-zip/x2-scalar.c",
180 "src/x32-zip/x3-scalar.c",
181 "src/x32-zip/x4-scalar.c",
182 "src/x32-zip/xm-scalar.c",
183 "src/x8-lut/scalar.c",
184 "src/x8-zip/x2-scalar.c",
185 "src/x8-zip/x3-scalar.c",
186 "src/x8-zip/x4-scalar.c",
187 "src/x8-zip/xm-scalar.c",
188]
189
190PSIMD_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800191 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800192 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800193 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700194 "src/f32-avgpool/mp9p8q-psimd.c",
195 "src/f32-avgpool/up9-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800196 "src/f32-bilinear/gen/psimd-c4.c",
197 "src/f32-bilinear/gen/psimd-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700198 "src/f32-clamp/psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800199 "src/f32-dwconv/gen/up4x25-psimd-acc2.c",
200 "src/f32-dwconv/gen/up4x25-psimd.c",
201 "src/f32-dwconv/gen/up4x4-psimd-acc2.c",
202 "src/f32-dwconv/gen/up4x4-psimd.c",
203 "src/f32-dwconv/gen/up4x9-psimd-acc2.c",
204 "src/f32-dwconv/gen/up4x9-psimd.c",
205 "src/f32-dwconv/gen/up8x25-psimd-acc2.c",
206 "src/f32-dwconv/gen/up8x25-psimd.c",
207 "src/f32-dwconv/gen/up8x4-psimd-acc2.c",
208 "src/f32-dwconv/gen/up8x4-psimd.c",
209 "src/f32-dwconv/gen/up8x9-psimd-acc2.c",
210 "src/f32-dwconv/gen/up8x9-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700211 "src/f32-gavgpool/mp7p7q-psimd.c",
212 "src/f32-gavgpool/up7-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800213 "src/f32-gemm/gen/1x8-psimd-loadsplat.c",
214 "src/f32-gemm/gen/1x8-psimd-splat.c",
215 "src/f32-gemm/gen/1x8s4-psimd.c",
216 "src/f32-gemm/gen/4x8-psimd-loadsplat.c",
217 "src/f32-gemm/gen/4x8-psimd-splat.c",
218 "src/f32-gemm/gen/4x8s4-psimd.c",
219 "src/f32-gemm/gen/6x8-psimd-loadsplat.c",
220 "src/f32-gemm/gen/6x8-psimd-splat.c",
221 "src/f32-gemm/gen/6x8s4-psimd.c",
222 "src/f32-gemm/gen-inc/1x8-psimd-loadsplat.c",
223 "src/f32-gemm/gen-inc/1x8-psimd-splat.c",
224 "src/f32-gemm/gen-inc/1x8s4-psimd.c",
225 "src/f32-gemm/gen-inc/4x8-psimd-loadsplat.c",
226 "src/f32-gemm/gen-inc/4x8-psimd-splat.c",
227 "src/f32-gemm/gen-inc/4x8s4-psimd.c",
228 "src/f32-gemm/gen-inc/6x8-psimd-loadsplat.c",
229 "src/f32-gemm/gen-inc/6x8-psimd-splat.c",
230 "src/f32-gemm/gen-inc/6x8s4-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700231 "src/f32-hswish/psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800232 "src/f32-igemm/gen/1x8-psimd-loadsplat.c",
233 "src/f32-igemm/gen/1x8-psimd-splat.c",
234 "src/f32-igemm/gen/1x8s4-psimd.c",
235 "src/f32-igemm/gen/4x2c4-psimd.c",
236 "src/f32-igemm/gen/4x8-psimd-loadsplat.c",
237 "src/f32-igemm/gen/4x8-psimd-splat.c",
238 "src/f32-igemm/gen/4x8s4-psimd.c",
239 "src/f32-igemm/gen/6x8-psimd-loadsplat.c",
240 "src/f32-igemm/gen/6x8-psimd-splat.c",
241 "src/f32-igemm/gen/6x8s4-psimd.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800242 "src/f32-maxpool/9p8x-psimd-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700243 "src/f32-pavgpool/mp9p8q-psimd.c",
244 "src/f32-pavgpool/up9-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800245 "src/f32-ppmm/gen/4x8-psimd.c",
246 "src/f32-prelu/gen/psimd-2x4.c",
247 "src/f32-prelu/gen/psimd-2x8.c",
248 "src/f32-vbinary/gen/vadd-psimd-x4.c",
249 "src/f32-vbinary/gen/vadd-psimd-x8.c",
250 "src/f32-vbinary/gen/vaddc-psimd-x4.c",
251 "src/f32-vbinary/gen/vaddc-psimd-x8.c",
252 "src/f32-vbinary/gen/vmul-psimd-x4.c",
253 "src/f32-vbinary/gen/vmul-psimd-x8.c",
254 "src/f32-vbinary/gen/vmulc-psimd-x4.c",
255 "src/f32-vbinary/gen/vmulc-psimd-x8.c",
256 "src/f32-vbinary/gen/vrsubc-psimd-x4.c",
257 "src/f32-vbinary/gen/vrsubc-psimd-x8.c",
258 "src/f32-vbinary/gen/vsub-psimd-x4.c",
259 "src/f32-vbinary/gen/vsub-psimd-x8.c",
260 "src/f32-vbinary/gen/vsubc-psimd-x4.c",
261 "src/f32-vbinary/gen/vsubc-psimd-x8.c",
262 "src/f32-vmulcaddc/gen/c4-psimd-2x.c",
263 "src/f32-vmulcaddc/gen/c8-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700264 "src/x32-packx/x4-psimd.c",
265 "src/x32-pad/x2-psimd.c",
266 "src/x32-unpool/psimd.c",
267 "src/x32-zip/x2-psimd.c",
268 "src/x32-zip/x3-psimd.c",
269 "src/x32-zip/x4-psimd.c",
270 "src/x32-zip/xm-psimd.c",
271]
272
273# ISA-specific micro-kernels
274NEON_UKERNELS = [
275 "src/f32-avgpool/mp9p8q-neon.c",
276 "src/f32-avgpool/up9-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800277 "src/f32-bilinear/gen/neon-c4.c",
278 "src/f32-bilinear/gen/neon-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700279 "src/f32-clamp/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800280 "src/f32-dwconv/gen/up4x9-neon.c",
281 "src/f32-dwconv/gen/up4x9-neon-acc2.c",
282 "src/f32-dwconv/gen/up8x9-neon.c",
283 "src/f32-dwconv/gen/up8x9-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700284 "src/f32-gavgpool-spchw/neon-x4.c",
285 "src/f32-gavgpool/mp7p7q-neon.c",
286 "src/f32-gavgpool/up7-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800287 "src/f32-gemm/gen/1x8-neon-lane-ld64.c",
288 "src/f32-gemm/gen/4x2-neon-lane-ld64.c",
289 "src/f32-gemm/gen/4x8-neon-lane-ld128.c",
290 "src/f32-gemm/gen/4x8-neon-lane-ld64.c",
291 "src/f32-gemm/gen/5x8-neon-lane-ld64.c",
292 "src/f32-gemm/gen/6x8-neon-lane-ld64.c",
293 "src/f32-gemm/gen/1x8-neon-dup-ld64.c",
294 "src/f32-gemm/gen/4x8-neon-dup-ld128.c",
295 "src/f32-gemm/gen/4x8-neon-dup-ld64.c",
296 "src/f32-gemm/gen/6x8-neon-dup-ld64.c",
297 "src/f32-gemm/gen/1x8s4-neon.c",
298 "src/f32-gemm/gen/4x8s4-neon.c",
299 "src/f32-gemm/gen/6x8s4-neon.c",
300 "src/f32-gemm/gen/8x8s4-neon.c",
301 "src/f32-gemm/gen-inc/1x8-neon-lane-ld64.c",
302 "src/f32-gemm/gen-inc/4x8-neon-lane-ld128.c",
303 "src/f32-gemm/gen-inc/4x8-neon-lane-ld64.c",
304 "src/f32-gemm/gen-inc/5x8-neon-lane-ld64.c",
305 "src/f32-gemm/gen-inc/6x8-neon-lane-ld64.c",
306 "src/f32-gemm/gen-inc/1x8-neon-dup-ld64.c",
307 "src/f32-gemm/gen-inc/4x8-neon-dup-ld128.c",
308 "src/f32-gemm/gen-inc/4x8-neon-dup-ld64.c",
309 "src/f32-gemm/gen-inc/6x8-neon-dup-ld64.c",
310 "src/f32-gemm/gen-inc/1x8s4-neon.c",
311 "src/f32-gemm/gen-inc/4x8s4-neon.c",
312 "src/f32-gemm/gen-inc/6x8s4-neon.c",
313 "src/f32-gemm/gen-inc/8x8s4-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700314 "src/f32-hswish/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800315 "src/f32-igemm/gen/1x8-neon-lane-ld64.c",
316 "src/f32-igemm/gen/4x2-neon-lane-ld64.c",
317 "src/f32-igemm/gen/4x4-neon-lane-ld64.c",
318 "src/f32-igemm/gen/4x8-neon-lane-ld128.c",
319 "src/f32-igemm/gen/4x8-neon-lane-ld64.c",
320 "src/f32-igemm/gen/6x8-neon-lane-ld64.c",
321 "src/f32-igemm/gen/1x8-neon-dup-ld64.c",
322 "src/f32-igemm/gen/4x8-neon-dup-ld128.c",
323 "src/f32-igemm/gen/4x8-neon-dup-ld64.c",
324 "src/f32-igemm/gen/6x8-neon-dup-ld64.c",
325 "src/f32-igemm/gen/1x8s4-neon.c",
326 "src/f32-igemm/gen/4x8s4-neon.c",
327 "src/f32-igemm/gen/6x8s4-neon.c",
328 "src/f32-igemm/gen/8x8s4-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700329 "src/f32-pavgpool/mp9p8q-neon.c",
330 "src/f32-pavgpool/up9-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800331 "src/f32-ppmm/gen/4x8-neon.c",
332 "src/f32-ppmm/gen/8x8-neon.c",
333 "src/f32-prelu/gen/neon-2x4.c",
334 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700335 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800336 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
337 "src/f32-vbinary/gen/vadd-neon-x4.c",
338 "src/f32-vbinary/gen/vadd-neon-x8.c",
339 "src/f32-vbinary/gen/vaddc-neon-x4.c",
340 "src/f32-vbinary/gen/vaddc-neon-x8.c",
341 "src/f32-vbinary/gen/vmul-neon-x4.c",
342 "src/f32-vbinary/gen/vmul-neon-x8.c",
343 "src/f32-vbinary/gen/vmulc-neon-x4.c",
344 "src/f32-vbinary/gen/vmulc-neon-x8.c",
345 "src/f32-vbinary/gen/vrsubc-neon-x4.c",
346 "src/f32-vbinary/gen/vrsubc-neon-x8.c",
347 "src/f32-vbinary/gen/vsub-neon-x4.c",
348 "src/f32-vbinary/gen/vsub-neon-x8.c",
349 "src/f32-vbinary/gen/vsubc-neon-x4.c",
350 "src/f32-vbinary/gen/vsubc-neon-x8.c",
351 "src/f32-vmulcaddc/gen/c4-neon-2x.c",
352 "src/f32-vmulcaddc/gen/c8-neon-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700353 "src/q8-avgpool/mp9p8q-neon.c",
354 "src/q8-avgpool/up9-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700355 "src/q8-dwconv/up8x9-neon.c",
356 "src/q8-gavgpool/mp7p7q-neon.c",
357 "src/q8-gavgpool/up7-neon.c",
358 "src/q8-gemm/4x8-neon.c",
359 "src/q8-gemm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800360 "src/q8-igemm/4x8-neon.c",
361 "src/q8-igemm/8x8-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700362 "src/q8-vadd/neon.c",
363 "src/u8-clamp/neon.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800364 "src/u8-maxpool/9p8x-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700365 "src/u8-rmax/neon.c",
366 "src/x32-packx/x4-neon-st4.c",
367 "src/x32-pad/x2-neon.c",
368 "src/x32-zip/x2-neon.c",
369 "src/x32-zip/x3-neon.c",
370 "src/x32-zip/x4-neon.c",
371 "src/x32-zip/xm-neon.c",
372 "src/x8-zip/x2-neon.c",
373 "src/x8-zip/x3-neon.c",
374 "src/x8-zip/x4-neon.c",
375 "src/x8-zip/xm-neon.c",
376]
377
378NEONFMA_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800379 "src/f32-bilinear/gen/neonfma-c4.c",
380 "src/f32-bilinear/gen/neonfma-c8.c",
381 "src/f32-igemm/gen/1x8-neonfma-dup-ld64.c",
382 "src/f32-igemm/gen/4x8-neonfma-dup-ld128.c",
383 "src/f32-igemm/gen/4x8-neonfma-dup-ld64.c",
384 "src/f32-igemm/gen/6x8-neonfma-dup-ld64.c",
385 "src/f32-igemm/gen/1x8s4-neonfma.c",
386 "src/f32-igemm/gen/4x8s4-neonfma.c",
387 "src/f32-igemm/gen/6x8s4-neonfma.c",
388 "src/f32-igemm/gen/8x8s4-neonfma.c",
389 "src/f32-dwconv/gen/up4x9-neonfma.c",
390 "src/f32-dwconv/gen/up4x9-neonfma-acc2.c",
391 "src/f32-dwconv/gen/up8x9-neonfma.c",
392 "src/f32-dwconv/gen/up8x9-neonfma-acc2.c",
393 "src/f32-gemm/gen/1x8-neonfma-dup-ld64.c",
394 "src/f32-gemm/gen/4x8-neonfma-dup-ld128.c",
395 "src/f32-gemm/gen/4x8-neonfma-dup-ld64.c",
396 "src/f32-gemm/gen/6x8-neonfma-dup-ld64.c",
397 "src/f32-gemm/gen/1x8s4-neonfma.c",
398 "src/f32-gemm/gen/4x8s4-neonfma.c",
399 "src/f32-gemm/gen/6x8s4-neonfma.c",
400 "src/f32-gemm/gen/8x8s4-neonfma.c",
401 "src/f32-gemm/gen-inc/1x8-neonfma-dup-ld64.c",
402 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld128.c",
403 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld64.c",
404 "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld64.c",
405 "src/f32-gemm/gen-inc/1x8s4-neonfma.c",
406 "src/f32-gemm/gen-inc/4x8s4-neonfma.c",
407 "src/f32-gemm/gen-inc/6x8s4-neonfma.c",
408 "src/f32-gemm/gen-inc/8x8s4-neonfma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700409 "src/f32-hswish/neonfma.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800410 "src/f32-ppmm/gen/4x8-neonfma.c",
411 "src/f32-ppmm/gen/8x8-neonfma.c",
412 "src/f32-sigmoid/gen/neonfma-p5-nr2fma-x16.c",
413 "src/f32-vmulcaddc/gen/c4-neonfma-2x.c",
414 "src/f32-vmulcaddc/gen/c8-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800415 "src/math/exp-neonfma-lut64-p2.c",
416 "src/math/exp-neonfma-p5.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800417 "src/math/expminus-neonfma-p5.c",
Marat Dukhan22aae132019-11-22 17:10:29 -0800418 "src/math/sigmoid-neonfma-p5-nr1recps1fma.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -0800419 "src/math/sigmoid-neonfma-p5-nr2fma.c",
Marat Dukhan22aae132019-11-22 17:10:29 -0800420 "src/math/sigmoid-neonfma-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700421]
422
423AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800424 "src/f32-gemm/gen/1x8-neonfma-lane-ld64.c",
425 "src/f32-gemm/gen/4x2-neonfma-lane-ld64.c",
426 "src/f32-gemm/gen/4x8-neonfma-lane-ld128.c",
427 "src/f32-gemm/gen/4x8-neonfma-lane-ld64.c",
428 "src/f32-gemm/gen/5x8-neonfma-lane-ld64.c",
429 "src/f32-gemm/gen/6x8-neonfma-lane-ld64.c",
430 "src/f32-gemm/gen-inc/1x8-neonfma-lane-ld64.c",
431 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld128.c",
432 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld64.c",
433 "src/f32-gemm/gen-inc/5x8-neonfma-lane-ld64.c",
434 "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld64.c",
435 "src/f32-igemm/gen/1x8-neonfma-lane-ld64.c",
436 "src/f32-igemm/gen/4x2-neonfma-lane-ld64.c",
437 "src/f32-igemm/gen/4x4-neonfma-lane-ld64.c",
438 "src/f32-igemm/gen/4x8-neonfma-lane-ld128.c",
439 "src/f32-igemm/gen/4x8-neonfma-lane-ld64.c",
440 "src/f32-igemm/gen/6x8-neonfma-lane-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700441 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
442 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
443 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
444 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
445 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
446 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
447 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800448 "src/f32-spmm/gen/12x1-neonfma.c",
449 "src/f32-spmm/gen/12x2-neonfma.c",
450 "src/f32-spmm/gen/12x4-neonfma.c",
451 "src/f32-spmm/gen/16x1-neonfma-pipelined.c",
452 "src/f32-spmm/gen/16x1-neonfma-unroll2.c",
453 "src/f32-spmm/gen/16x1-neonfma.c",
454 "src/f32-spmm/gen/16x2-neonfma.c",
455 "src/f32-spmm/gen/16x4-neonfma.c",
456 "src/f32-spmm/gen/4x1-neonfma-pipelined.c",
457 "src/f32-spmm/gen/4x1-neonfma-unroll2.c",
458 "src/f32-spmm/gen/4x1-neonfma.c",
459 "src/f32-spmm/gen/4x2-neonfma.c",
460 "src/f32-spmm/gen/4x4-neonfma.c",
461 "src/f32-spmm/gen/8x1-neonfma-pipelined.c",
462 "src/f32-spmm/gen/8x1-neonfma-unroll2.c",
463 "src/f32-spmm/gen/8x1-neonfma.c",
464 "src/f32-spmm/gen/8x2-neonfma.c",
465 "src/f32-spmm/gen/8x4-neonfma.c",
Marat Dukhan22aae132019-11-22 17:10:29 -0800466 "src/math/sigmoid-neonfma-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700467]
468
469AARCH64_NEONFP16ARITH_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800470 "src/f16-gemm/gen/4x8-neonfp16arith-ld64.c",
471 "src/f16-gemm/gen/6x8-neonfp16arith-ld64.c",
472 "src/f16-gemm/gen/8x8-neonfp16arith-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700473]
474
475SSE_UKERNELS = [
476 "src/f32-avgpool/mp9p8q-sse.c",
477 "src/f32-avgpool/up9-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800478 "src/f32-bilinear/gen/sse-c4.c",
479 "src/f32-bilinear/gen/sse-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700480 "src/f32-clamp/sse.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800481 "src/f32-dwconv-spchw/3x3p1-sse.c",
482 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800483 "src/f32-dwconv/gen/up4x25-sse-acc2.c",
484 "src/f32-dwconv/gen/up4x25-sse.c",
485 "src/f32-dwconv/gen/up4x4-sse-acc2.c",
486 "src/f32-dwconv/gen/up4x4-sse.c",
487 "src/f32-dwconv/gen/up4x9-sse-acc2.c",
488 "src/f32-dwconv/gen/up4x9-sse.c",
489 "src/f32-dwconv/gen/up8x25-sse-acc2.c",
490 "src/f32-dwconv/gen/up8x25-sse.c",
491 "src/f32-dwconv/gen/up8x4-sse-acc2.c",
492 "src/f32-dwconv/gen/up8x4-sse.c",
493 "src/f32-dwconv/gen/up8x9-sse-acc2.c",
494 "src/f32-dwconv/gen/up8x9-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700495 "src/f32-gavgpool-spchw/sse-x4.c",
496 "src/f32-gavgpool/mp7p7q-sse.c",
497 "src/f32-gavgpool/up7-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800498 "src/f32-gemm/gen/1x8-sse-dup.c",
499 "src/f32-gemm/gen/1x8-sse-load1.c",
500 "src/f32-gemm/gen/1x8s4-sse.c",
501 "src/f32-gemm/gen/4x8-sse-dup.c",
502 "src/f32-gemm/gen/4x8-sse-load1.c",
503 "src/f32-gemm/gen/4x8s4-sse.c",
504 "src/f32-gemm/gen-inc/1x8-sse-dup.c",
505 "src/f32-gemm/gen-inc/1x8-sse-load1.c",
506 "src/f32-gemm/gen-inc/1x8s4-sse.c",
507 "src/f32-gemm/gen-inc/4x8-sse-dup.c",
508 "src/f32-gemm/gen-inc/4x8-sse-load1.c",
509 "src/f32-gemm/gen-inc/4x8s4-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700510 "src/f32-hswish/sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800511 "src/f32-igemm/gen/1x8-sse-dup.c",
512 "src/f32-igemm/gen/1x8-sse-load1.c",
513 "src/f32-igemm/gen/1x8s4-sse.c",
514 "src/f32-igemm/gen/4x2c4-sse.c",
515 "src/f32-igemm/gen/4x8-sse-dup.c",
516 "src/f32-igemm/gen/4x8-sse-load1.c",
517 "src/f32-igemm/gen/4x8s4-sse.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800518 "src/f32-maxpool/9p8x-sse-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700519 "src/f32-pavgpool/mp9p8q-sse.c",
520 "src/f32-pavgpool/up9-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800521 "src/f32-ppmm/gen/4x8-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700522 "src/f32-rmax/sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800523 "src/f32-spmm/gen/4x1-sse.c",
524 "src/f32-spmm/gen/8x1-sse.c",
525 "src/f32-vbinary/gen/vadd-sse-x4.c",
526 "src/f32-vbinary/gen/vadd-sse-x8.c",
527 "src/f32-vbinary/gen/vaddc-sse-x4.c",
528 "src/f32-vbinary/gen/vaddc-sse-x8.c",
529 "src/f32-vbinary/gen/vmul-sse-x4.c",
530 "src/f32-vbinary/gen/vmul-sse-x8.c",
531 "src/f32-vbinary/gen/vmulc-sse-x4.c",
532 "src/f32-vbinary/gen/vmulc-sse-x8.c",
533 "src/f32-vbinary/gen/vrsubc-sse-x4.c",
534 "src/f32-vbinary/gen/vrsubc-sse-x8.c",
535 "src/f32-vbinary/gen/vsub-sse-x4.c",
536 "src/f32-vbinary/gen/vsub-sse-x8.c",
537 "src/f32-vbinary/gen/vsubc-sse-x4.c",
538 "src/f32-vbinary/gen/vsubc-sse-x8.c",
539 "src/f32-vmulcaddc/gen/c4-sse-2x.c",
540 "src/f32-vmulcaddc/gen/c8-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700541 "src/x32-packx/x4-sse.c",
542]
543
544SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800545 "src/f32-argmaxpool/9p8x-sse2-c4.c",
546 "src/f32-argmaxpool/4x-sse2-c4.c",
547 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800548 "src/f32-prelu/gen/sse2-2x4.c",
549 "src/f32-prelu/gen/sse2-2x8.c",
550 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
551 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700552 "src/q8-avgpool/mp9p8q-sse2.c",
553 "src/q8-avgpool/up9-sse2.c",
554 "src/q8-igemm/4x4c2-sse2.c",
555 "src/q8-dwconv/up8x9-sse2.c",
556 "src/q8-gavgpool/mp7p7q-sse2.c",
557 "src/q8-gavgpool/up7-sse2.c",
558 "src/q8-gemm/2x4c8-sse2.c",
559 "src/q8-gemm/4x4c2-sse2.c",
560 "src/q8-vadd/sse2.c",
561 "src/u8-clamp/sse2.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800562 "src/u8-maxpool/9p8x-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700563 "src/u8-rmax/sse2.c",
564 "src/x32-pad/x2-sse2.c",
565 "src/x32-zip/x2-sse2.c",
566 "src/x32-zip/x3-sse2.c",
567 "src/x32-zip/x4-sse2.c",
568 "src/x32-zip/xm-sse2.c",
569 "src/x8-zip/x2-sse2.c",
570 "src/x8-zip/x3-sse2.c",
571 "src/x8-zip/x4-sse2.c",
572 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -0800573 "src/math/exp-sse2-p5.c",
574 "src/math/expminus-sse2-p5.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -0800575 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700576]
577
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800578SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800579 "src/f32-prelu/gen/sse41-2x4.c",
580 "src/f32-prelu/gen/sse41-2x8.c",
581 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
582 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800583]
584
Marat Dukhan08c4a432019-10-03 09:29:21 -0700585AVX_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800586 "src/f32-dwconv/gen/up16x4-avx-acc2.c",
587 "src/f32-dwconv/gen/up16x4-avx.c",
588 "src/f32-dwconv/gen/up8x4-avx-acc2.c",
589 "src/f32-dwconv/gen/up8x4-avx.c",
590 "src/f32-dwconv/gen/up16x9-avx-acc2.c",
591 "src/f32-dwconv/gen/up16x9-avx.c",
592 "src/f32-dwconv/gen/up8x9-avx-acc2.c",
593 "src/f32-dwconv/gen/up8x9-avx.c",
594 "src/f32-dwconv/gen/up16x25-avx-acc2.c",
595 "src/f32-dwconv/gen/up16x25-avx.c",
596 "src/f32-dwconv/gen/up8x25-avx-acc2.c",
597 "src/f32-dwconv/gen/up8x25-avx.c",
598 "src/f32-gemm/gen/1x8-avx-broadcast.c",
599 "src/f32-gemm/gen/4x8-avx-broadcast.c",
600 "src/f32-gemm/gen/5x8-avx-broadcast.c",
601 "src/f32-gemm/gen/6x8-avx-broadcast.c",
602 "src/f32-gemm/gen/7x8-avx-broadcast.c",
603 "src/f32-gemm/gen-inc/1x8-avx-broadcast.c",
604 "src/f32-gemm/gen-inc/4x8-avx-broadcast.c",
605 "src/f32-gemm/gen-inc/5x8-avx-broadcast.c",
606 "src/f32-gemm/gen-inc/6x8-avx-broadcast.c",
607 "src/f32-gemm/gen-inc/7x8-avx-broadcast.c",
608 "src/f32-igemm/gen/1x8-avx-broadcast.c",
609 "src/f32-igemm/gen/4x8-avx-broadcast.c",
610 "src/f32-igemm/gen/5x8-avx-broadcast.c",
611 "src/f32-igemm/gen/6x8-avx-broadcast.c",
612 "src/f32-igemm/gen/7x8-avx-broadcast.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700613 "src/f32-rmax/avx.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700614 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700615]
616
Marat Dukhanfda12b82019-11-21 12:27:59 -0800617FMA3_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800618 "src/f32-dwconv/gen/up16x4-fma3-acc2.c",
619 "src/f32-dwconv/gen/up16x4-fma3.c",
620 "src/f32-dwconv/gen/up8x4-fma3-acc2.c",
621 "src/f32-dwconv/gen/up8x4-fma3.c",
622 "src/f32-dwconv/gen/up16x9-fma3-acc2.c",
623 "src/f32-dwconv/gen/up16x9-fma3.c",
624 "src/f32-dwconv/gen/up8x9-fma3-acc2.c",
625 "src/f32-dwconv/gen/up8x9-fma3.c",
626 "src/f32-dwconv/gen/up16x25-fma3-acc2.c",
627 "src/f32-dwconv/gen/up16x25-fma3.c",
628 "src/f32-dwconv/gen/up8x25-fma3-acc2.c",
629 "src/f32-dwconv/gen/up8x25-fma3.c",
630 "src/f32-gemm/gen/1x8-fma3-broadcast.c",
631 "src/f32-gemm/gen/4x8-fma3-broadcast.c",
632 "src/f32-gemm/gen/5x8-fma3-broadcast.c",
633 "src/f32-gemm/gen/6x8-fma3-broadcast.c",
634 "src/f32-gemm/gen/7x8-fma3-broadcast.c",
635 "src/f32-gemm/gen/8x8-fma3-broadcast.c",
636 "src/f32-gemm/gen-inc/1x8-fma3-broadcast.c",
637 "src/f32-gemm/gen-inc/4x8-fma3-broadcast.c",
638 "src/f32-gemm/gen-inc/5x8-fma3-broadcast.c",
639 "src/f32-gemm/gen-inc/6x8-fma3-broadcast.c",
640 "src/f32-gemm/gen-inc/7x8-fma3-broadcast.c",
641 "src/f32-gemm/gen-inc/8x8-fma3-broadcast.c",
642 "src/f32-igemm/gen/1x8-fma3-broadcast.c",
643 "src/f32-igemm/gen/4x8-fma3-broadcast.c",
644 "src/f32-igemm/gen/5x8-fma3-broadcast.c",
645 "src/f32-igemm/gen/6x8-fma3-broadcast.c",
646 "src/f32-igemm/gen/7x8-fma3-broadcast.c",
647 "src/f32-igemm/gen/8x8-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -0800648]
649
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700650AVX2_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700651 "src/f32-raddexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700652 "src/f32-raddextexp/avx2-p5-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700653 "src/f32-raddstoreexpminusmax/avx2-p5-unroll64.c",
654 "src/f32-vscaleexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700655 "src/f32-vscaleextexp/avx2-p5-unroll64.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700656 "src/math/exp-avx2-p5.c",
657 "src/math/exp-avx2-perm-p3.c",
658 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -0700659 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700660 "src/math/extexp-avx2-p5.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700661]
662
Marat Dukhan08c4a432019-10-03 09:29:21 -0700663AVX512F_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700664 "src/f32-raddexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700665 "src/f32-raddextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700666 "src/f32-raddstoreexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700667 "src/f32-rmax/avx512f.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700668 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700669 "src/f32-vscaleexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700670 "src/f32-vscaleextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700671 "src/math/exp-avx512f-p5-scalef.c",
672 "src/math/exp-avx512f-p5.c",
673 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -0700674 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700675 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700676]
677
678AARCH32_ASM_UKERNELS = [
679 "src/q8-dwconv/up8x9-aarch32-neon.S",
680]
681
682AARCH64_ASM_UKERNELS = [
683 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
684 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800685 "src/f32-gemm/gen/1x12-aarch64-neonfma-cortex-a53.S",
686 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a53.S",
687 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
688 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
689 "src/f32-gemm/gen/4x12-aarch64-neonfma-cortex-a53.S",
690 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a53.S",
691 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
692 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
693 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld128.S",
694 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld64.S",
695 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
696 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a53.S",
697 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
698 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a73.S",
699 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
700 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld128.S",
701 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld64.S",
702 "src/f32-gemm/gen-inc/1x12-aarch64-neonfma-cortex-a53.S",
703 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a53.S",
704 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a57.S",
705 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a75.S",
706 "src/f32-gemm/gen-inc/4x12-aarch64-neonfma-cortex-a53.S",
707 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a53.S",
708 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a57.S",
709 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a75.S",
710 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld128.S",
711 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld64.S",
712 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a75.S",
713 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a53.S",
714 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a57.S",
715 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a73.S",
716 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a75.S",
717 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld128.S",
718 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700719 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700720 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700721 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S",
722 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S",
723 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700724 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700725 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
726 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700727 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700728 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
729 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
730 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
731]
732
733INTERNAL_MICROKERNEL_HDRS = [
734 "src/xnnpack/argmaxpool.h",
735 "src/xnnpack/avgpool.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800736 "src/xnnpack/bilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700737 "src/xnnpack/clamp.h",
738 "src/xnnpack/common.h",
739 "src/xnnpack/conv.h",
740 "src/xnnpack/dwconv.h",
741 "src/xnnpack/gavgpool.h",
742 "src/xnnpack/gemm.h",
743 "src/xnnpack/hswish.h",
744 "src/xnnpack/igemm.h",
745 "src/xnnpack/lut.h",
746 "src/xnnpack/math.h",
747 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -0800748 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700749 "src/xnnpack/packx.h",
750 "src/xnnpack/pad.h",
751 "src/xnnpack/params.h",
752 "src/xnnpack/pavgpool.h",
753 "src/xnnpack/ppmm.h",
754 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700755 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700756 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700757 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700758 "src/xnnpack/rmax.h",
759 "src/xnnpack/scalar-utils.h",
760 "src/xnnpack/spmm.h",
761 "src/xnnpack/unpool.h",
762 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800763 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700764 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700765 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700766 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700767 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800768 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700769 "src/xnnpack/zip.h",
770]
771
772INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
773 "include/xnnpack.h",
774 "src/xnnpack/allocator.h",
775 "src/xnnpack/compute.h",
776 "src/xnnpack/im2col.h",
777 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700778 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700779 "src/xnnpack/operator.h",
780 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700781 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700782 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700783 "src/xnnpack/requantization.h",
784]
785
786ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
787 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700788]
789
790MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700791 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700792 "include/xnnpack.h",
793]
794
795MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
796 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700797 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700798 "src/xnnpack/requantization.h",
799 "include/xnnpack.h",
800]
801
802OPERATOR_TEST_PARAMS_HDRS = [
803 "src/xnnpack/params.h",
804 "src/xnnpack/common.h",
805]
806
807WEIGHTS_PACK_HDRS = [
808 "src/xnnpack/pack.h",
809 "src/xnnpack/operator.h",
810 "src/xnnpack/compute.h",
811]
812
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700813LOGGING_COPTS = select({
814 # No logging in optimized mode
815 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
816 # Full logging in debug mode
817 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
818 # Error-only logging in default (fastbuild) mode
819 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
820})
821
822LOGGING_HDRS = [
823 "src/xnnpack/log.h",
824]
825
Marat Dukhan08c4a432019-10-03 09:29:21 -0700826xnnpack_cc_library(
827 name = "scalar_ukernels",
828 srcs = SCALAR_UKERNELS,
829 hdrs = INTERNAL_HDRS,
830 aarch32_copts = ["-marm"],
831 copts = xnnpack_std_copts(),
832 deps = [
833 "@FP16",
834 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -0800835 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700836 ],
837)
838
839xnnpack_cc_library(
840 name = "psimd_ukernels",
841 srcs = PSIMD_UKERNELS,
842 hdrs = INTERNAL_HDRS,
843 aarch32_copts = [
844 "-marm",
845 "-mfpu=neon",
846 ],
847 copts = xnnpack_std_copts(),
848 optimized_copts = [
849 "-O3",
850 "-ffast-math",
851 ],
852 deps = [
853 "@FP16",
854 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -0800855 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700856 ],
857)
858
859xnnpack_cc_library(
860 name = "neon_ukernels",
861 hdrs = INTERNAL_HDRS,
862 aarch32_copts = [
863 "-marm",
864 "-mfpu=neon",
865 ],
866 aarch32_srcs = NEON_UKERNELS,
867 aarch64_srcs = NEON_UKERNELS,
868 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -0800869 deps = [
870 "@FP16",
871 "@pthreadpool",
872 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700873)
874
875xnnpack_cc_library(
876 name = "neonfma_ukernels",
877 hdrs = INTERNAL_HDRS,
878 aarch32_copts = [
879 "-marm",
880 "-mfpu=neon-vfpv4",
881 ],
882 aarch32_srcs = NEONFMA_UKERNELS,
883 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
884 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -0800885 deps = [
886 "@FP16",
887 "@pthreadpool",
888 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700889)
890
891xnnpack_cc_library(
892 name = "neonfp16arith_ukernels",
893 hdrs = INTERNAL_HDRS,
894 aarch64_copts = ["-march=armv8.2-a+fp16"],
895 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
896 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -0800897 deps = [
898 "@FP16",
899 "@pthreadpool",
900 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700901)
902
903xnnpack_cc_library(
904 name = "sse2_ukernels",
905 hdrs = INTERNAL_HDRS,
906 copts = xnnpack_std_copts(),
907 x86_copts = ["-msse2"],
908 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -0800909 deps = [
910 "@FP16",
911 "@pthreadpool",
912 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700913)
914
915xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800916 name = "sse41_ukernels",
917 hdrs = INTERNAL_HDRS,
918 copts = xnnpack_std_copts(),
919 x86_copts = ["-msse4.1"],
920 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -0800921 deps = [
922 "@FP16",
923 "@pthreadpool",
924 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800925)
926
927xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700928 name = "avx_ukernels",
929 hdrs = INTERNAL_HDRS,
930 copts = xnnpack_std_copts(),
931 x86_copts = ["-mavx"],
932 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -0800933 deps = [
934 "@FP16",
935 "@pthreadpool",
936 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700937)
938
939xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -0800940 name = "fma3_ukernels",
941 hdrs = INTERNAL_HDRS,
942 copts = xnnpack_std_copts(),
943 x86_copts = [
944 "-mfma",
945 ],
946 x86_srcs = FMA3_UKERNELS,
947 deps = [
948 "@FP16",
949 "@pthreadpool",
950 ],
951)
952
953xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700954 name = "avx2_ukernels",
955 hdrs = INTERNAL_HDRS,
956 copts = xnnpack_std_copts(),
957 x86_copts = [
958 "-mfma",
959 "-mavx2",
960 ],
961 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -0800962 deps = [
963 "@FP16",
964 "@pthreadpool",
965 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700966)
967
968xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700969 name = "avx512f_ukernels",
970 hdrs = INTERNAL_HDRS,
971 copts = xnnpack_std_copts(),
972 x86_copts = ["-mavx512f"],
973 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -0800974 deps = [
975 "@FP16",
976 "@pthreadpool",
977 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700978)
979
980xnnpack_cc_library(
981 name = "asm_ukernels",
982 hdrs = ["src/xnnpack/assembly.h"],
983 aarch32_srcs = AARCH32_ASM_UKERNELS,
984 aarch64_srcs = AARCH64_ASM_UKERNELS,
985)
986
987xnnpack_aggregate_library(
988 name = "ukernels",
989 aarch32_deps = [
990 ":psimd_ukernels",
991 ":neon_ukernels",
992 ":neonfma_ukernels",
993 ":asm_ukernels",
994 ],
995 aarch64_deps = [
996 ":psimd_ukernels",
997 ":neon_ukernels",
998 ":neonfma_ukernels",
999 ":neonfp16arith_ukernels",
1000 ":asm_ukernels",
1001 ],
1002 generic_deps = [":scalar_ukernels"],
1003 wasmsimd_deps = [
1004 ":psimd_ukernels",
1005 ],
1006 x86_deps = [
1007 ":psimd_ukernels",
1008 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001009 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001010 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001011 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001012 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001013 ":avx512f_ukernels",
1014 ],
1015)
1016
1017xnnpack_cc_library(
1018 name = "im2col",
1019 srcs = ["src/im2col.c"],
1020 hdrs = [
1021 "src/xnnpack/common.h",
1022 "src/xnnpack/im2col.h",
1023 ],
1024 copts = xnnpack_std_copts(),
1025)
1026
1027xnnpack_cc_library(
1028 name = "indirection",
1029 srcs = ["src/indirection.c"],
1030 hdrs = INTERNAL_HDRS,
1031 copts = xnnpack_std_copts(),
1032 deps = [
1033 "@FP16",
1034 "@FXdiv",
1035 "@pthreadpool",
1036 ],
1037)
1038
1039xnnpack_cc_library(
1040 name = "operator_run",
1041 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001042 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
1043 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001044 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
1045 "-Wno-vla",
1046 ],
1047 deps = [
1048 "@FP16",
1049 "@FXdiv",
1050 "@clog",
1051 "@pthreadpool",
1052 ],
1053)
1054
1055cc_library(
1056 name = "enable_assembly",
1057 defines = select({
1058 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1059 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001060 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001061 }),
1062)
1063
Marat Dukhancf056b22019-10-07 10:26:29 -07001064xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001065 name = "operators",
1066 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001067 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001068 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07001069 ],
1070 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001071 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001072 "-Isrc",
1073 "-Iinclude",
1074 ] + select({
1075 ":debug_build": [],
1076 "//conditions:default": xnnpack_min_size_copts(),
1077 }),
Marat Dukhancf056b22019-10-07 10:26:29 -07001078 wasm_srcs = ["src/wasm-stubs.c"],
1079 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001080 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001081 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001082 "@FP16",
1083 "@FXdiv",
1084 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001085 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001086 ],
1087)
1088
1089cc_library(
1090 name = "XNNPACK",
1091 srcs = [
1092 "src/init.c",
1093 ],
1094 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1095 "-Isrc",
1096 "-Iinclude",
1097 ] + select({
1098 ":debug_build": [],
1099 "//conditions:default": xnnpack_min_size_copts(),
1100 }),
1101 includes = ["include"],
1102 linkstatic = True,
1103 textual_hdrs = ["include/xnnpack.h"],
1104 visibility = xnnpack_visibility(),
1105 deps = [
1106 ":enable_assembly",
1107 ":ukernels",
1108 ":operator_run",
1109 ":operators",
1110 "@clog",
1111 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07001112 ] + select({
1113 ":emscripten": [],
1114 "//conditions:default": ["@cpuinfo"],
1115 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001116)
1117
1118cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001119 name = "xnnpack_operators_nhwc_f32",
1120 srcs = [
1121 "src/init.c",
1122 ],
1123 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1124 "-Isrc",
1125 "-Iinclude",
1126 ] + select({
1127 ":debug_build": [],
1128 "//conditions:default": xnnpack_min_size_copts(),
1129 }),
1130 defines = [
1131 "XNN_NO_Q8_OPERATORS",
1132 "XNN_NO_U8_OPERATORS",
1133 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08001134 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001135 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001136 includes = ["include"],
1137 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001138 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001139 visibility = xnnpack_visibility(),
1140 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001141 ":enable_assembly",
1142 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001143 ":operator_run",
1144 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001145 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001146 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001147 ] + select({
1148 ":emscripten": [],
1149 "//conditions:default": ["@cpuinfo"],
1150 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001151)
1152
Marat Dukhancf056b22019-10-07 10:26:29 -07001153xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001154 name = "bench_utils",
1155 srcs = ["bench/utils.cc"],
1156 hdrs = ["bench/utils.h"],
1157 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08001158 deps = [
1159 "@com_google_benchmark//:benchmark",
1160 "@cpuinfo",
1161 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001162)
1163
Frank Barchard7e955972019-10-11 10:34:25 -07001164######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07001165
1166xnnpack_benchmark(
1167 name = "q8_gemm_bench",
1168 srcs = [
1169 "bench/gemm.h",
1170 "bench/q8-gemm.cc",
1171 "src/xnnpack/AlignedAllocator.h",
1172 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1173 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
1174 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
1175)
1176
1177xnnpack_benchmark(
1178 name = "f16_gemm_bench",
1179 srcs = [
1180 "bench/f16-gemm.cc",
1181 "bench/gemm.h",
1182 "src/xnnpack/AlignedAllocator.h",
1183 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1184 copts = ["-Wno-unused-function"],
1185 deps = MICROKERNEL_BENCHMARK_DEPS,
1186)
1187
1188xnnpack_benchmark(
1189 name = "f32_igemm_bench",
1190 srcs = [
1191 "bench/f32-igemm.cc",
1192 "bench/conv.h",
1193 "src/xnnpack/AlignedAllocator.h",
1194 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07001195 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001196)
1197
1198xnnpack_benchmark(
1199 name = "f32_conv_hwc_bench",
1200 srcs = [
1201 "bench/f32-conv-hwc.cc",
1202 "bench/dconv.h",
1203 "src/xnnpack/AlignedAllocator.h",
1204 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1205 copts = ["-Wno-unused-function"],
1206 deps = MICROKERNEL_BENCHMARK_DEPS,
1207)
1208
1209xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07001210 name = "f32_conv_hwc2spchw_bench",
1211 srcs = [
1212 "bench/f32-conv-hwc2spchw.cc",
1213 "bench/dconv.h",
1214 "src/xnnpack/AlignedAllocator.h",
1215 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1216 copts = ["-Wno-unused-function"],
1217 deps = MICROKERNEL_BENCHMARK_DEPS,
1218)
1219
1220xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001221 name = "f32_dwconv_bench",
1222 srcs = [
1223 "bench/f32-dwconv.cc",
1224 "bench/dwconv.h",
1225 "src/xnnpack/AlignedAllocator.h",
1226 ] + MICROKERNEL_BENCHMARK_HDRS,
1227 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1228)
1229
1230xnnpack_benchmark(
1231 name = "f32_dwconv_spchw_bench",
1232 srcs = [
1233 "bench/f32-dwconv-spchw.cc",
1234 "bench/dwconv.h",
1235 "src/xnnpack/AlignedAllocator.h",
1236 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1237 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1238)
1239
1240xnnpack_benchmark(
1241 name = "f32_gemm_bench",
1242 srcs = [
1243 "bench/f32-gemm.cc",
1244 "bench/gemm.h",
1245 "src/xnnpack/AlignedAllocator.h",
1246 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1247 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07001248 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001249)
1250
1251xnnpack_benchmark(
1252 name = "f32_rmax_bench",
1253 srcs = [
1254 "bench/f32-rmax.cc",
1255 "src/xnnpack/AlignedAllocator.h",
1256 ] + MICROKERNEL_BENCHMARK_HDRS,
1257 deps = MICROKERNEL_BENCHMARK_DEPS,
1258)
1259
1260xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08001261 name = "f32_sigmoid_bench",
1262 srcs = [
1263 "bench/f32-sigmoid.cc",
1264 "src/xnnpack/AlignedAllocator.h",
1265 ] + MICROKERNEL_BENCHMARK_HDRS,
1266 copts = ["-Wno-unused-function"],
1267 deps = MICROKERNEL_BENCHMARK_DEPS,
1268)
1269
1270xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001271 name = "f32_spmm_bench",
1272 srcs = [
1273 "bench/f32-spmm.cc",
1274 "bench/gemm.h",
1275 "src/xnnpack/AlignedAllocator.h",
1276 ] + MICROKERNEL_BENCHMARK_HDRS,
1277 copts = ["-Wno-unused-function"],
1278 deps = MICROKERNEL_BENCHMARK_DEPS,
1279)
1280
1281xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07001282 name = "f32_softargmax_bench",
1283 srcs = [
1284 "bench/f32-softargmax.cc",
1285 ] + MICROKERNEL_BENCHMARK_HDRS,
1286 copts = ["-Wno-unused-function"],
1287 deps = MICROKERNEL_BENCHMARK_DEPS,
1288)
1289
1290xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001291 name = "f32_im2col_gemm_bench",
1292 srcs = [
1293 "bench/f32-im2col-gemm.cc",
1294 "bench/conv.h",
1295 "src/xnnpack/AlignedAllocator.h",
1296 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1297 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
1298)
1299
1300########################### Benchmarks for operators ###########################
1301
1302xnnpack_benchmark(
1303 name = "add_bench",
1304 srcs = ["bench/add.cc"],
1305 deps = OPERATOR_BENCHMARK_DEPS,
1306)
1307
1308xnnpack_benchmark(
1309 name = "average_pooling_bench",
1310 srcs = ["bench/average-pooling.cc"],
1311 deps = OPERATOR_BENCHMARK_DEPS,
1312)
1313
1314xnnpack_benchmark(
1315 name = "channel_shuffle_bench",
1316 srcs = ["bench/channel-shuffle.cc"],
1317 deps = OPERATOR_BENCHMARK_DEPS,
1318)
1319
1320xnnpack_benchmark(
1321 name = "convolution_bench",
1322 srcs = ["bench/convolution.cc"],
1323 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1324 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1325)
1326
1327xnnpack_benchmark(
1328 name = "deconvolution_bench",
1329 srcs = ["bench/deconvolution.cc"],
1330 copts = xnnpack_optional_tflite_copts(),
1331 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1332)
1333
1334xnnpack_benchmark(
1335 name = "global_average_pooling_bench",
1336 srcs = ["bench/global-average-pooling.cc"],
1337 deps = OPERATOR_BENCHMARK_DEPS,
1338)
1339
1340xnnpack_benchmark(
1341 name = "max_pooling_bench",
1342 srcs = ["bench/max-pooling.cc"],
1343 deps = OPERATOR_BENCHMARK_DEPS,
1344)
1345
1346xnnpack_benchmark(
1347 name = "sigmoid_bench",
1348 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08001349 copts = xnnpack_optional_tflite_copts(),
1350 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001351)
1352
1353xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07001354 name = "prelu_bench",
1355 srcs = ["bench/prelu.cc"],
1356 copts = xnnpack_optional_tflite_copts(),
1357 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1358)
1359
1360xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001361 name = "softargmax_bench",
1362 srcs = ["bench/softargmax.cc"],
1363 deps = OPERATOR_BENCHMARK_DEPS,
1364)
1365
Marat Dukhanc068bb62019-10-04 13:24:39 -07001366############################# End-to-end benchmarks ############################
1367
1368cc_library(
1369 name = "mobilenet_v1",
1370 srcs = ["models/mobilenet-v1.cc"],
1371 hdrs = ["models/models.h"],
1372 linkstatic = True,
1373 deps = [
1374 ":XNNPACK",
1375 "@pthreadpool",
1376 ],
1377)
1378
1379cc_library(
1380 name = "mobilenet_v2",
1381 srcs = ["models/mobilenet-v2.cc"],
1382 hdrs = ["models/models.h"],
1383 linkstatic = True,
1384 deps = [
1385 ":XNNPACK",
1386 "@pthreadpool",
1387 ],
1388)
1389
1390xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07001391 name = "f32_dwconv_e2e_bench",
1392 srcs = ["bench/f32-dwconv-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1393 copts = ["-Wno-unused-function"],
1394 deps = MICROKERNEL_BENCHMARK_DEPS + [
1395 ":XNNPACK",
1396 ":mobilenet_v1",
1397 ":mobilenet_v2",
1398 ],
1399)
1400
1401xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07001402 name = "f32_gemm_e2e_bench",
1403 srcs = ["bench/f32-gemm-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1404 copts = ["-Wno-unused-function"],
1405 deps = MICROKERNEL_BENCHMARK_DEPS + [
1406 ":XNNPACK",
1407 ":mobilenet_v1",
1408 ":mobilenet_v2",
1409 ],
1410)
1411
1412xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07001413 name = "end2end_bench",
1414 srcs = ["bench/end2end.cc"],
1415 deps = [
1416 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07001417 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07001418 ":mobilenet_v1",
1419 ":mobilenet_v2",
1420 "@pthreadpool",
1421 ],
1422)
1423
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001424#################### Accuracy evaluation for math functions ####################
1425
1426xnnpack_benchmark(
1427 name = "f32_exp_eval",
1428 srcs = [
1429 "eval/f32-exp.cc",
1430 "src/xnnpack/AlignedAllocator.h",
1431 ] + ACCURACY_EVAL_HDRS,
1432 deps = ACCURACY_EVAL_DEPS,
1433)
1434
Marat Dukhan515c9772019-10-17 18:07:57 -07001435xnnpack_benchmark(
1436 name = "f32_expminus_eval",
1437 srcs = [
1438 "eval/f32-expminus.cc",
1439 "src/xnnpack/AlignedAllocator.h",
1440 ] + ACCURACY_EVAL_HDRS,
1441 deps = ACCURACY_EVAL_DEPS,
1442)
1443
Marat Dukhan98ba4412019-10-23 02:14:28 -07001444xnnpack_benchmark(
1445 name = "f32_extexp_eval",
1446 srcs = [
1447 "eval/f32-extexp.cc",
1448 "src/xnnpack/AlignedAllocator.h",
1449 ] + ACCURACY_EVAL_HDRS,
1450 deps = ACCURACY_EVAL_DEPS,
1451)
1452
Marat Dukhan346a9e52019-11-15 09:06:30 -08001453xnnpack_benchmark(
1454 name = "f32_sigmoid_eval",
1455 srcs = [
1456 "eval/f32-sigmoid.cc",
1457 "src/xnnpack/AlignedAllocator.h",
1458 ] + ACCURACY_EVAL_HDRS,
1459 deps = ACCURACY_EVAL_DEPS,
1460)
1461
Marat Dukhan08c4a432019-10-03 09:29:21 -07001462######################### Unit tests for micro-kernels #########################
1463
1464xnnpack_unit_test(
1465 name = "f16_gemm_test",
1466 srcs = [
1467 "test/f16-gemm.cc",
1468 "test/gemm-microkernel-tester.h",
1469 "src/xnnpack/AlignedAllocator.h",
1470 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1471 deps = MICROKERNEL_TEST_DEPS,
1472)
1473
1474xnnpack_unit_test(
1475 name = "f32_argmaxpool_test",
1476 srcs = [
1477 "test/f32-argmaxpool.cc",
1478 "test/argmaxpool-microkernel-tester.h",
1479 "src/xnnpack/AlignedAllocator.h",
1480 ] + MICROKERNEL_TEST_HDRS,
1481 deps = MICROKERNEL_TEST_DEPS,
1482)
1483
1484xnnpack_unit_test(
1485 name = "f32_avgpool_test",
1486 srcs = [
1487 "test/f32-avgpool.cc",
1488 "test/avgpool-microkernel-tester.h",
1489 "src/xnnpack/AlignedAllocator.h",
1490 ] + MICROKERNEL_TEST_HDRS,
1491 deps = MICROKERNEL_TEST_DEPS,
1492)
1493
1494xnnpack_unit_test(
Marat Dukhan35dacfb2019-11-07 19:18:16 -08001495 name = "f32_bilinear_test",
1496 srcs = [
1497 "test/f32-bilinear.cc",
1498 "test/bilinear-microkernel-tester.h",
1499 "src/xnnpack/AlignedAllocator.h",
1500 ] + MICROKERNEL_TEST_HDRS,
1501 deps = MICROKERNEL_TEST_DEPS,
1502)
1503
1504xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001505 name = "f32_clamp_test",
1506 srcs = [
1507 "test/f32-clamp.cc",
1508 "test/clamp-microkernel-tester.h",
1509 ] + MICROKERNEL_TEST_HDRS,
1510 deps = MICROKERNEL_TEST_DEPS,
1511)
1512
1513xnnpack_unit_test(
1514 name = "f32_igemm_test",
1515 srcs = [
1516 "test/f32-igemm.cc",
1517 "test/gemm-microkernel-tester.h",
1518 "src/xnnpack/AlignedAllocator.h",
1519 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1520 deps = MICROKERNEL_TEST_DEPS,
1521)
1522
1523xnnpack_unit_test(
1524 name = "f32_conv_hwc_test",
1525 srcs = [
1526 "test/f32-conv-hwc.cc",
1527 "test/conv-hwc-microkernel-tester.h",
1528 "src/xnnpack/AlignedAllocator.h",
1529 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1530 deps = MICROKERNEL_TEST_DEPS,
1531)
1532
1533xnnpack_unit_test(
1534 name = "f32_conv_hwc2spchw_test",
1535 srcs = [
1536 "test/f32-conv-hwc2spchw.cc",
1537 "test/conv-hwc2spchw-microkernel-tester.h",
1538 "src/xnnpack/AlignedAllocator.h",
1539 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1540 deps = MICROKERNEL_TEST_DEPS,
1541)
1542
1543xnnpack_unit_test(
1544 name = "f32_dwconv_test",
1545 srcs = [
1546 "test/f32-dwconv.cc",
1547 "test/dwconv-microkernel-tester.h",
1548 "src/xnnpack/AlignedAllocator.h",
1549 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1550 deps = MICROKERNEL_TEST_DEPS,
1551)
1552
1553xnnpack_unit_test(
1554 name = "f32_dwconv_spchw_test",
1555 srcs = [
1556 "test/f32-dwconv-spchw.cc",
1557 "test/dwconv-spchw-microkernel-tester.h",
1558 "src/xnnpack/AlignedAllocator.h",
1559 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1560 deps = MICROKERNEL_TEST_DEPS,
1561)
1562
1563xnnpack_unit_test(
1564 name = "f32_gavgpool_test",
1565 srcs = [
1566 "test/f32-gavgpool.cc",
1567 "test/gavgpool-microkernel-tester.h",
1568 "src/xnnpack/AlignedAllocator.h",
1569 ] + MICROKERNEL_TEST_HDRS,
1570 deps = MICROKERNEL_TEST_DEPS,
1571)
1572
1573xnnpack_unit_test(
1574 name = "f32_gavgpool_spchw_test",
1575 srcs = [
1576 "test/f32-gavgpool-spchw.cc",
1577 "test/gavgpool-spchw-microkernel-tester.h",
1578 "src/xnnpack/AlignedAllocator.h",
1579 ] + MICROKERNEL_TEST_HDRS,
1580 deps = MICROKERNEL_TEST_DEPS,
1581)
1582
1583xnnpack_unit_test(
1584 name = "f32_gemm_test",
1585 srcs = [
1586 "test/f32-gemm.cc",
1587 "test/gemm-microkernel-tester.h",
1588 "src/xnnpack/AlignedAllocator.h",
1589 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1590 deps = MICROKERNEL_TEST_DEPS,
1591)
1592
1593xnnpack_unit_test(
1594 name = "f32_gemminc_test",
1595 srcs = [
1596 "test/f32-gemminc.cc",
1597 "test/gemm-microkernel-tester.h",
1598 "src/xnnpack/AlignedAllocator.h",
1599 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1600 deps = MICROKERNEL_TEST_DEPS,
1601)
1602
1603xnnpack_unit_test(
1604 name = "f32_hswish_test",
1605 srcs = [
1606 "test/f32-hswish.cc",
1607 "test/hswish-microkernel-tester.h",
1608 ] + MICROKERNEL_TEST_HDRS,
1609 deps = MICROKERNEL_TEST_DEPS,
1610)
1611
1612xnnpack_unit_test(
1613 name = "f32_maxpool_test",
1614 srcs = [
1615 "test/f32-maxpool.cc",
1616 "test/maxpool-microkernel-tester.h",
1617 ] + MICROKERNEL_TEST_HDRS,
1618 deps = MICROKERNEL_TEST_DEPS,
1619)
1620
1621xnnpack_unit_test(
1622 name = "f32_pavgpool_test",
1623 srcs = [
1624 "test/f32-pavgpool.cc",
1625 "test/avgpool-microkernel-tester.h",
1626 "src/xnnpack/AlignedAllocator.h",
1627 ] + MICROKERNEL_TEST_HDRS,
1628 deps = MICROKERNEL_TEST_DEPS,
1629)
1630
1631xnnpack_unit_test(
1632 name = "f32_ppmm_test",
1633 srcs = [
1634 "test/f32-ppmm.cc",
1635 "test/gemm-microkernel-tester.h",
1636 "src/xnnpack/AlignedAllocator.h",
1637 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1638 deps = MICROKERNEL_TEST_DEPS,
1639)
1640
1641xnnpack_unit_test(
1642 name = "f32_prelu_test",
1643 srcs = [
1644 "test/f32-prelu.cc",
1645 "test/prelu-microkernel-tester.h",
1646 "src/xnnpack/AlignedAllocator.h",
1647 ] + MICROKERNEL_TEST_HDRS,
1648 deps = MICROKERNEL_TEST_DEPS,
1649)
1650
1651xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001652 name = "f32_raddexpminusmax_test",
1653 srcs = [
1654 "test/f32-raddexpminusmax.cc",
1655 "test/raddexpminusmax-microkernel-tester.h",
1656 ] + MICROKERNEL_TEST_HDRS,
1657 deps = MICROKERNEL_TEST_DEPS,
1658)
1659
1660xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001661 name = "f32_raddextexp_test",
1662 srcs = [
1663 "test/f32-raddextexp.cc",
1664 "test/raddextexp-microkernel-tester.h",
1665 ] + MICROKERNEL_TEST_HDRS,
1666 deps = MICROKERNEL_TEST_DEPS,
1667)
1668
1669xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001670 name = "f32_raddstoreexpminusmax_test",
1671 srcs = [
1672 "test/f32-raddstoreexpminusmax.cc",
1673 "test/raddstoreexpminusmax-microkernel-tester.h",
1674 ] + MICROKERNEL_TEST_HDRS,
1675 deps = MICROKERNEL_TEST_DEPS,
1676)
1677
1678xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001679 name = "f32_rmax_test",
1680 srcs = [
1681 "test/f32-rmax.cc",
1682 "test/rmax-microkernel-tester.h",
1683 ] + MICROKERNEL_TEST_HDRS,
1684 deps = MICROKERNEL_TEST_DEPS,
1685)
1686
1687xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08001688 name = "f32_sigmoid_test",
1689 srcs = [
1690 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001691 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001692 ] + MICROKERNEL_TEST_HDRS,
1693 deps = MICROKERNEL_TEST_DEPS,
1694)
1695
1696xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001697 name = "f32_spmm_test",
1698 srcs = [
1699 "test/f32-spmm.cc",
1700 "test/spmm-microkernel-tester.h",
1701 "src/xnnpack/AlignedAllocator.h",
1702 ] + MICROKERNEL_TEST_HDRS,
1703 deps = MICROKERNEL_TEST_DEPS,
1704)
1705
1706xnnpack_unit_test(
1707 name = "f32_vadd_test",
1708 srcs = [
1709 "test/f32-vadd.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001710 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001711 ] + MICROKERNEL_TEST_HDRS,
1712 deps = MICROKERNEL_TEST_DEPS,
1713)
1714
1715xnnpack_unit_test(
1716 name = "f32_vaddc_test",
1717 srcs = [
1718 "test/f32-vaddc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001719 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001720 ] + MICROKERNEL_TEST_HDRS,
1721 deps = MICROKERNEL_TEST_DEPS,
1722)
1723
1724xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001725 name = "f32_vmul_test",
1726 srcs = [
1727 "test/f32-vmul.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001728 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001729 ] + MICROKERNEL_TEST_HDRS,
1730 deps = MICROKERNEL_TEST_DEPS,
1731)
1732
1733xnnpack_unit_test(
1734 name = "f32_vmulc_test",
1735 srcs = [
1736 "test/f32-vmulc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001737 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001738 ] + MICROKERNEL_TEST_HDRS,
1739 deps = MICROKERNEL_TEST_DEPS,
1740)
1741
1742xnnpack_unit_test(
1743 name = "f32_vmulcaddc_test",
1744 srcs = [
1745 "test/f32-vmulcaddc.cc",
1746 "test/vmulcaddc-microkernel-tester.h",
1747 "src/xnnpack/AlignedAllocator.h",
1748 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1749 deps = MICROKERNEL_TEST_DEPS,
1750)
1751
1752xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001753 name = "f32_vscale_test",
1754 srcs = [
1755 "test/f32-vscale.cc",
1756 "test/vscale-microkernel-tester.h",
1757 ] + MICROKERNEL_TEST_HDRS,
1758 deps = MICROKERNEL_TEST_DEPS,
1759)
1760
1761xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001762 name = "f32_vscaleexpminusmax_test",
1763 srcs = [
1764 "test/f32-vscaleexpminusmax.cc",
1765 "test/vscaleexpminusmax-microkernel-tester.h",
1766 ] + MICROKERNEL_TEST_HDRS,
1767 deps = MICROKERNEL_TEST_DEPS,
1768)
1769
1770xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001771 name = "f32_vscaleextexp_test",
1772 srcs = [
1773 "test/f32-vscaleextexp.cc",
1774 "test/vscaleextexp-microkernel-tester.h",
1775 ] + MICROKERNEL_TEST_HDRS,
1776 deps = MICROKERNEL_TEST_DEPS,
1777)
1778
1779xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001780 name = "f32_vsub_test",
1781 srcs = [
1782 "test/f32-vsub.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001783 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001784 ] + MICROKERNEL_TEST_HDRS,
1785 deps = MICROKERNEL_TEST_DEPS,
1786)
1787
1788xnnpack_unit_test(
1789 name = "f32_vsubc_test",
1790 srcs = [
1791 "test/f32-vsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001792 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001793 ] + MICROKERNEL_TEST_HDRS,
1794 deps = MICROKERNEL_TEST_DEPS,
1795)
1796
1797xnnpack_unit_test(
1798 name = "f32_vrsubc_test",
1799 srcs = [
1800 "test/f32-vrsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001801 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001802 ] + MICROKERNEL_TEST_HDRS,
1803 deps = MICROKERNEL_TEST_DEPS,
1804)
1805
1806xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001807 name = "q8_avgpool_test",
1808 srcs = [
1809 "test/q8-avgpool.cc",
1810 "test/avgpool-microkernel-tester.h",
1811 "src/xnnpack/AlignedAllocator.h",
1812 ] + MICROKERNEL_TEST_HDRS,
1813 deps = MICROKERNEL_TEST_DEPS,
1814)
1815
1816xnnpack_unit_test(
1817 name = "q8_igemm_test",
1818 srcs = [
1819 "test/q8-igemm.cc",
1820 "test/gemm-microkernel-tester.h",
1821 "src/xnnpack/AlignedAllocator.h",
1822 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1823 deps = MICROKERNEL_TEST_DEPS,
1824)
1825
1826xnnpack_unit_test(
1827 name = "q8_dwconv_test",
1828 srcs = [
1829 "test/q8-dwconv.cc",
1830 "test/dwconv-microkernel-tester.h",
1831 "src/xnnpack/AlignedAllocator.h",
1832 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1833 deps = MICROKERNEL_TEST_DEPS,
1834)
1835
1836xnnpack_unit_test(
1837 name = "q8_gavgpool_test",
1838 srcs = [
1839 "test/q8-gavgpool.cc",
1840 "test/gavgpool-microkernel-tester.h",
1841 "src/xnnpack/AlignedAllocator.h",
1842 ] + MICROKERNEL_TEST_HDRS,
1843 deps = MICROKERNEL_TEST_DEPS,
1844)
1845
1846xnnpack_unit_test(
1847 name = "q8_gemm_test",
1848 srcs = [
1849 "test/q8-gemm.cc",
1850 "test/gemm-microkernel-tester.h",
1851 "src/xnnpack/AlignedAllocator.h",
1852 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1853 deps = MICROKERNEL_TEST_DEPS,
1854)
1855
1856xnnpack_unit_test(
1857 name = "q8_vadd_test",
1858 srcs = [
1859 "test/q8-vadd.cc",
1860 "test/vadd-microkernel-tester.h",
1861 ] + MICROKERNEL_TEST_HDRS,
1862 deps = MICROKERNEL_TEST_DEPS,
1863)
1864
1865xnnpack_unit_test(
1866 name = "u8_clamp_test",
1867 srcs = [
1868 "test/u8-clamp.cc",
1869 "test/clamp-microkernel-tester.h",
1870 ] + MICROKERNEL_TEST_HDRS,
1871 deps = MICROKERNEL_TEST_DEPS,
1872)
1873
1874xnnpack_unit_test(
1875 name = "u8_lut32norm_test",
1876 srcs = [
1877 "test/u8-lut32norm.cc",
1878 "test/lut-norm-microkernel-tester.h",
1879 ] + MICROKERNEL_TEST_HDRS,
1880 deps = MICROKERNEL_TEST_DEPS,
1881)
1882
1883xnnpack_unit_test(
1884 name = "u8_maxpool_test",
1885 srcs = [
1886 "test/u8-maxpool.cc",
1887 "test/maxpool-microkernel-tester.h",
1888 ] + MICROKERNEL_TEST_HDRS,
1889 deps = MICROKERNEL_TEST_DEPS,
1890)
1891
1892xnnpack_unit_test(
1893 name = "u8_rmax_test",
1894 srcs = [
1895 "test/u8-rmax.cc",
1896 "test/rmax-microkernel-tester.h",
1897 ] + MICROKERNEL_TEST_HDRS,
1898 deps = MICROKERNEL_TEST_DEPS,
1899)
1900
1901xnnpack_unit_test(
1902 name = "x32_packx_test",
1903 srcs = [
1904 "test/x32-packx.cc",
1905 "test/pack-microkernel-tester.h",
1906 "src/xnnpack/AlignedAllocator.h",
1907 ] + MICROKERNEL_TEST_HDRS,
1908 deps = MICROKERNEL_TEST_DEPS,
1909)
1910
1911xnnpack_unit_test(
1912 name = "x32_pad_test",
1913 srcs = [
1914 "test/x32-pad.cc",
1915 "test/pad-microkernel-tester.h",
1916 ] + MICROKERNEL_TEST_HDRS,
1917 deps = MICROKERNEL_TEST_DEPS,
1918)
1919
1920xnnpack_unit_test(
1921 name = "x32_unpool_test",
1922 srcs = [
1923 "test/x32-unpool.cc",
1924 "test/unpool-microkernel-tester.h",
1925 ] + MICROKERNEL_TEST_HDRS,
1926 deps = MICROKERNEL_TEST_DEPS,
1927)
1928
1929xnnpack_unit_test(
1930 name = "x32_zip_test",
1931 srcs = [
1932 "test/x32-zip.cc",
1933 "test/zip-microkernel-tester.h",
1934 ] + MICROKERNEL_TEST_HDRS,
1935 deps = MICROKERNEL_TEST_DEPS,
1936)
1937
1938xnnpack_unit_test(
1939 name = "x8_lut_test",
1940 srcs = [
1941 "test/x8-lut.cc",
1942 "test/lut-microkernel-tester.h",
1943 ] + MICROKERNEL_TEST_HDRS,
1944 deps = MICROKERNEL_TEST_DEPS,
1945)
1946
1947xnnpack_unit_test(
1948 name = "x8_zip_test",
1949 srcs = [
1950 "test/x8-zip.cc",
1951 "test/zip-microkernel-tester.h",
1952 ] + MICROKERNEL_TEST_HDRS,
1953 deps = MICROKERNEL_TEST_DEPS,
1954)
1955
1956########################### Size test for the library ##########################
1957
1958xnnpack_binary(
1959 name = "size_test",
1960 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001961 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001962)
1963
1964########################### Unit tests for operators ###########################
1965
1966xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001967 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001968 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001969 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001970 "test/add-operator-tester.h",
1971 ],
1972 deps = OPERATOR_TEST_DEPS,
1973)
1974
1975xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001976 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001977 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001978 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001979 "test/argmax-pooling-operator-tester.h",
1980 ] + OPERATOR_TEST_PARAMS_HDRS,
1981 deps = OPERATOR_TEST_DEPS,
1982)
1983
1984xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001985 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001986 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001987 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001988 "test/average-pooling-operator-tester.h",
1989 ] + OPERATOR_TEST_PARAMS_HDRS,
1990 deps = OPERATOR_TEST_DEPS,
1991)
1992
1993xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001994 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001995 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001996 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001997 "test/channel-pad-operator-tester.h",
1998 ] + OPERATOR_TEST_PARAMS_HDRS,
1999 deps = OPERATOR_TEST_DEPS,
2000)
2001
2002xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002003 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002004 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002005 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002006 "test/channel-shuffle-operator-tester.h",
2007 ],
2008 deps = OPERATOR_TEST_DEPS,
2009)
2010
2011xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002012 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002013 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002014 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002015 "test/clamp-operator-tester.h",
2016 ],
2017 deps = OPERATOR_TEST_DEPS,
2018)
2019
2020xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002021 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002022 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002023 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002024 "test/convolution-operator-tester.h",
2025 ],
2026 deps = OPERATOR_TEST_DEPS,
2027)
2028
2029xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002030 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002031 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002032 "test/convolution-nchw.cc",
2033 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002034 ],
2035 deps = OPERATOR_TEST_DEPS,
2036)
2037
2038xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002039 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002040 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002041 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002042 "test/deconvolution-operator-tester.h",
2043 ] + OPERATOR_TEST_PARAMS_HDRS,
2044 deps = OPERATOR_TEST_DEPS,
2045)
2046
2047xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002048 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002049 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002050 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002051 "test/fully-connected-operator-tester.h",
2052 ],
2053 deps = OPERATOR_TEST_DEPS,
2054)
2055
2056xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002057 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002058 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002059 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002060 "test/global-average-pooling-operator-tester.h",
2061 ] + OPERATOR_TEST_PARAMS_HDRS,
2062 deps = OPERATOR_TEST_DEPS,
2063)
2064
2065xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002066 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002067 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002068 "test/global-average-pooling-ncw.cc",
2069 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002070 ],
2071 deps = OPERATOR_TEST_DEPS,
2072)
2073
2074xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002075 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002076 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002077 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002078 "test/hardswish-operator-tester.h",
2079 ],
2080 deps = OPERATOR_TEST_DEPS,
2081)
2082
2083xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002084 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002085 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002086 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002087 "test/leaky-relu-operator-tester.h",
2088 ],
2089 deps = OPERATOR_TEST_DEPS,
2090)
2091
2092xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002093 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002094 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002095 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002096 "test/max-pooling-operator-tester.h",
2097 ] + OPERATOR_TEST_PARAMS_HDRS,
2098 deps = OPERATOR_TEST_DEPS,
2099)
2100
2101xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002102 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08002103 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002104 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08002105 "test/multiply-operator-tester.h",
2106 ],
2107 deps = OPERATOR_TEST_DEPS,
2108)
2109
2110xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002111 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002112 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002113 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002114 "test/prelu-operator-tester.h",
2115 ] + OPERATOR_TEST_PARAMS_HDRS,
2116 deps = OPERATOR_TEST_DEPS,
2117)
2118
2119xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002120 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08002121 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002122 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08002123 "test/resize-bilinear-operator-tester.h",
2124 ] + OPERATOR_TEST_PARAMS_HDRS,
2125 deps = OPERATOR_TEST_DEPS,
2126)
2127
2128xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002129 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002130 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002131 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002132 "test/sigmoid-operator-tester.h",
2133 ],
2134 deps = OPERATOR_TEST_DEPS,
2135)
2136
2137xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002138 name = "softargmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002139 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002140 "test/softargmax-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002141 "test/softargmax-operator-tester.h",
2142 ],
2143 deps = OPERATOR_TEST_DEPS,
2144)
2145
2146xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002147 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002148 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002149 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002150 "test/unpooling-operator-tester.h",
2151 ],
2152 deps = OPERATOR_TEST_DEPS,
2153)
2154
2155############################# Build configurations #############################
2156
Marat Dukhanb8642352019-10-30 15:43:02 -07002157# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07002158config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07002159 name = "xnn_enable_assembly_explicit_true",
2160 define_values = {"xnn_enable_assembly": "true"},
2161)
2162
2163# Disables usage of assembly kernels.
2164config_setting(
2165 name = "xnn_enable_assembly_explicit_false",
2166 define_values = {"xnn_enable_assembly": "false"},
2167)
2168
2169# Builds with -c dbg
2170config_setting(
2171 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002172 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07002173 "compilation_mode": "dbg",
2174 },
2175)
2176
2177# Builds with -c opt
2178config_setting(
2179 name = "optimized_build",
2180 values = {
2181 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002182 },
2183)
2184
2185config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07002186 name = "linux_k8",
2187 values = {"cpu": "k8"},
2188)
2189
2190config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07002191 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07002192 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07002193)
2194
2195config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002196 name = "android",
2197 values = {"crosstool_top": "//external:android/crosstool"},
2198)
2199
2200config_setting(
2201 name = "android_armv7",
2202 values = {
2203 "crosstool_top": "//external:android/crosstool",
2204 "cpu": "armeabi-v7a",
2205 },
2206)
2207
2208config_setting(
2209 name = "android_arm64",
2210 values = {
2211 "crosstool_top": "//external:android/crosstool",
2212 "cpu": "arm64-v8a",
2213 },
2214)
2215
2216config_setting(
2217 name = "android_x86",
2218 values = {
2219 "crosstool_top": "//external:android/crosstool",
2220 "cpu": "x86",
2221 },
2222)
2223
2224config_setting(
2225 name = "android_x86_64",
2226 values = {
2227 "crosstool_top": "//external:android/crosstool",
2228 "cpu": "x86_64",
2229 },
2230)
2231
2232config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07002233 name = "macos_x86_64",
2234 values = {
2235 "apple_platform_type": "macos",
2236 "cpu": "darwin",
2237 },
2238)
2239
2240config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002241 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002242 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07002243)
2244
2245config_setting(
2246 name = "emscripten_wasm",
2247 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002248 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002249 "cpu": "wasm",
2250 },
2251)
2252
2253config_setting(
2254 name = "emscripten_wasmsimd",
2255 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002256 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002257 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07002258 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002259 },
2260)
2261
2262config_setting(
2263 name = "emscripten_asmjs",
2264 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002265 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002266 "cpu": "asmjs",
2267 },
2268)