blob: ba9477a1b5f82a076abab4d978fd0bc00d25ba7e [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 = [
53 "src/add.c",
54 "src/argmax-pooling.c",
55 "src/average-pooling.c",
56 "src/channel-pad.c",
57 "src/channel-shuffle.c",
58 "src/clamp.c",
59 "src/convolution-spnchw.c",
60 "src/convolution.c",
61 "src/deconvolution.c",
62 "src/fully-connected.c",
63 "src/global-average-pooling-spnchw.c",
64 "src/global-average-pooling.c",
65 "src/hardswish.c",
66 "src/leaky-relu.c",
67 "src/max-pooling.c",
Marat Dukhanca2733c2019-11-15 23:21:17 -080068 "src/multiply.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070069 "src/prelu.c",
Marat Dukhan69722492019-11-11 19:55:50 -080070 "src/resize-bilinear.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070071 "src/sigmoid.c",
72 "src/softargmax.c",
73 "src/unpooling.c",
74]
75
76SCALAR_UKERNELS = [
77 "src/f32-argmaxpool/mp9p8q-scalar.c",
78 "src/f32-argmaxpool/up4-scalar.c",
79 "src/f32-argmaxpool/up9-scalar.c",
80 "src/f32-avgpool/mp9p8q-scalar.c",
81 "src/f32-avgpool/up9-scalar.c",
Marat Dukhan35dacfb2019-11-07 19:18:16 -080082 "src/f32-bilinear/scalar-c1.c",
83 "src/f32-bilinear/scalar-c2.c",
84 "src/f32-bilinear/scalar-c4.c",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -080085 "src/f32-binop/vadd-scalar-x1.c",
86 "src/f32-binop/vadd-scalar-x2.c",
87 "src/f32-binop/vadd-scalar-x4.c",
88 "src/f32-binop/vaddc-scalar-x1.c",
89 "src/f32-binop/vaddc-scalar-x2.c",
90 "src/f32-binop/vaddc-scalar-x4.c",
91 "src/f32-binop/vmul-scalar-x1.c",
92 "src/f32-binop/vmul-scalar-x2.c",
93 "src/f32-binop/vmul-scalar-x4.c",
94 "src/f32-binop/vmulc-scalar-x1.c",
95 "src/f32-binop/vmulc-scalar-x2.c",
96 "src/f32-binop/vmulc-scalar-x4.c",
97 "src/f32-binop/vsub-scalar-x1.c",
98 "src/f32-binop/vsub-scalar-x2.c",
99 "src/f32-binop/vsub-scalar-x4.c",
100 "src/f32-binop/vsubc-scalar-x1.c",
101 "src/f32-binop/vsubc-scalar-x2.c",
102 "src/f32-binop/vsubc-scalar-x4.c",
103 "src/f32-binop/vrsubc-scalar-x1.c",
104 "src/f32-binop/vrsubc-scalar-x2.c",
105 "src/f32-binop/vrsubc-scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700106 "src/f32-clamp/scalar.c",
Erich Elsen563df5f2019-10-23 08:02:21 -0700107 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700108 "src/f32-igemm/1x4-scalar.c",
109 "src/f32-igemm/2x4-scalar.c",
110 "src/f32-igemm/4x2-scalar.c",
111 "src/f32-igemm/4x4-scalar.c",
112 "src/f32-dwconv/up1x25-scalar.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800113 "src/f32-dwconv/up1x25-scalar-acc2.c",
114 "src/f32-dwconv/up2x25-scalar.c",
115 "src/f32-dwconv/up2x25-scalar-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700116 "src/f32-dwconv/up1x4-scalar.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800117 "src/f32-dwconv/up1x4-scalar-acc2.c",
118 "src/f32-dwconv/up2x4-scalar.c",
119 "src/f32-dwconv/up2x4-scalar-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700120 "src/f32-dwconv/up1x9-scalar.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800121 "src/f32-dwconv/up1x9-scalar-acc2.c",
122 "src/f32-dwconv/up2x9-scalar.c",
123 "src/f32-dwconv/up2x9-scalar-acc2.c",
Erich Elsen0cc2c532019-10-15 04:44:18 -0700124 "src/f32-dwconv-spchw/3x3p1-scalar.c",
Erich Elsen38709a62019-11-08 11:58:45 -0800125 "src/f32-dwconv-spchw/5x5p2-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -0700126 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Erich Elsen38709a62019-11-08 11:58:45 -0800127 "src/f32-dwconv-spchw/5x5s2p2-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -0700128 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700129 "src/f32-gavgpool/mp7p7q-scalar.c",
130 "src/f32-gavgpool/up7-scalar.c",
131 "src/f32-gemm/1x4-scalar.c",
132 "src/f32-gemm/2x4-scalar.c",
133 "src/f32-gemm/4x2-scalar.c",
134 "src/f32-gemm/4x4-scalar.c",
135 "src/f32-gemminc/1x4-scalar.c",
136 "src/f32-gemminc/2x4-scalar.c",
137 "src/f32-gemminc/4x4-scalar.c",
138 "src/f32-hswish/scalar.c",
139 "src/f32-maxpool/9p8q-scalar.c",
140 "src/f32-pavgpool/mp9p8q-scalar.c",
141 "src/f32-pavgpool/up9-scalar.c",
142 "src/f32-ppmm/2x4-scalar.c",
143 "src/f32-ppmm/3x3-scalar.c",
144 "src/f32-ppmm/4x2-scalar.c",
145 "src/f32-ppmm/4x4-scalar.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800146 "src/f32-prelu/scalar-2x1.c",
147 "src/f32-prelu/scalar-2x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700148 "src/f32-rmax/scalar.c",
149 "src/f32-spmm/1x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700150 "src/f32-spmm/1x1-scalar.c",
151 "src/f32-spmm/2x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700152 "src/f32-spmm/2x1-scalar.c",
153 "src/f32-spmm/4x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700154 "src/f32-spmm/4x1-scalar.c",
155 "src/f32-spmm/8x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700156 "src/f32-spmm/8x1-scalar.c",
Erich Elsenc6afd9b2019-10-24 16:10:53 -0700157 "src/f32-spmm/8x2-scalar.c",
158 "src/f32-spmm/8x4-scalar.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800159 "src/f32-vmulcaddc/c1-scalar-2x.c",
160 "src/f32-vmulcaddc/c2-scalar-2x.c",
161 "src/f32-vmulcaddc/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",
164 "src/q8-igemm/2x2-scalar.c",
165 "src/q8-dwconv/up1x9-scalar.c",
166 "src/q8-gavgpool/mp7p7q-scalar.c",
167 "src/q8-gavgpool/up7-scalar.c",
168 "src/q8-gemm/2x2-scalar.c",
169 "src/q8-vadd/scalar.c",
170 "src/u8-clamp/scalar.c",
171 "src/u8-lut32norm/scalar.c",
172 "src/u8-maxpool/9p8q-scalar.c",
173 "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 = [
191 "src/f32-argmaxpool/mp9p8q-psimd.c",
192 "src/f32-argmaxpool/up4-psimd.c",
193 "src/f32-argmaxpool/up9-psimd.c",
194 "src/f32-avgpool/mp9p8q-psimd.c",
195 "src/f32-avgpool/up9-psimd.c",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800196 "src/f32-bilinear/psimd-c4.c",
197 "src/f32-bilinear/psimd-c8.c",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -0800198 "src/f32-binop/vadd-psimd-x4.c",
199 "src/f32-binop/vadd-psimd-x8.c",
200 "src/f32-binop/vaddc-psimd-x4.c",
201 "src/f32-binop/vaddc-psimd-x8.c",
202 "src/f32-binop/vmul-psimd-x4.c",
203 "src/f32-binop/vmul-psimd-x8.c",
204 "src/f32-binop/vmulc-psimd-x4.c",
205 "src/f32-binop/vmulc-psimd-x8.c",
206 "src/f32-binop/vrsubc-psimd-x4.c",
207 "src/f32-binop/vrsubc-psimd-x8.c",
208 "src/f32-binop/vsub-psimd-x4.c",
209 "src/f32-binop/vsub-psimd-x8.c",
210 "src/f32-binop/vsubc-psimd-x4.c",
211 "src/f32-binop/vsubc-psimd-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700212 "src/f32-clamp/psimd.c",
213 "src/f32-igemm/1x8-psimd-loadsplat.c",
214 "src/f32-igemm/1x8-psimd-splat.c",
215 "src/f32-igemm/1x8s4-psimd.c",
216 "src/f32-igemm/4x2c4-psimd.c",
217 "src/f32-igemm/4x8-psimd-loadsplat.c",
218 "src/f32-igemm/4x8-psimd-splat.c",
219 "src/f32-igemm/4x8s4-psimd.c",
220 "src/f32-igemm/6x8-psimd-loadsplat.c",
221 "src/f32-igemm/6x8-psimd-splat.c",
222 "src/f32-igemm/6x8s4-psimd.c",
223 "src/f32-dwconv/up4x25-psimd.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800224 "src/f32-dwconv/up4x25-psimd-acc2.c",
225 "src/f32-dwconv/up8x25-psimd.c",
226 "src/f32-dwconv/up8x25-psimd-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700227 "src/f32-dwconv/up4x4-psimd.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800228 "src/f32-dwconv/up4x4-psimd-acc2.c",
229 "src/f32-dwconv/up8x4-psimd.c",
230 "src/f32-dwconv/up8x4-psimd-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700231 "src/f32-dwconv/up4x9-psimd.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800232 "src/f32-dwconv/up4x9-psimd-acc2.c",
233 "src/f32-dwconv/up8x9-psimd.c",
234 "src/f32-dwconv/up8x9-psimd-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700235 "src/f32-gavgpool/mp7p7q-psimd.c",
236 "src/f32-gavgpool/up7-psimd.c",
237 "src/f32-gemm/1x8-psimd-loadsplat.c",
238 "src/f32-gemm/1x8-psimd-splat.c",
239 "src/f32-gemm/1x8s4-psimd.c",
240 "src/f32-gemm/4x8-psimd-loadsplat.c",
241 "src/f32-gemm/4x8-psimd-splat.c",
242 "src/f32-gemm/4x8s4-psimd.c",
243 "src/f32-gemm/6x8-psimd-loadsplat.c",
244 "src/f32-gemm/6x8-psimd-splat.c",
245 "src/f32-gemm/6x8s4-psimd.c",
246 "src/f32-gemminc/1x8-psimd-loadsplat.c",
247 "src/f32-gemminc/1x8-psimd-splat.c",
248 "src/f32-gemminc/1x8s4-psimd.c",
249 "src/f32-gemminc/4x8-psimd-loadsplat.c",
250 "src/f32-gemminc/4x8-psimd-splat.c",
251 "src/f32-gemminc/4x8s4-psimd.c",
252 "src/f32-gemminc/6x8-psimd-loadsplat.c",
253 "src/f32-gemminc/6x8-psimd-splat.c",
254 "src/f32-gemminc/6x8s4-psimd.c",
255 "src/f32-hswish/psimd.c",
256 "src/f32-maxpool/9p8q-psimd.c",
257 "src/f32-pavgpool/mp9p8q-psimd.c",
258 "src/f32-pavgpool/up9-psimd.c",
259 "src/f32-ppmm/4x8-psimd.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800260 "src/f32-prelu/psimd-2x4.c",
261 "src/f32-prelu/psimd-2x8.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800262 "src/f32-vmulcaddc/c4-psimd-2x.c",
263 "src/f32-vmulcaddc/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 Dukhan35dacfb2019-11-07 19:18:16 -0800277 "src/f32-bilinear/neon-c4.c",
278 "src/f32-bilinear/neon-c8.c",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -0800279 "src/f32-binop/vadd-neon-x4.c",
280 "src/f32-binop/vadd-neon-x8.c",
281 "src/f32-binop/vaddc-neon-x4.c",
282 "src/f32-binop/vaddc-neon-x8.c",
283 "src/f32-binop/vmul-neon-x4.c",
284 "src/f32-binop/vmul-neon-x8.c",
285 "src/f32-binop/vmulc-neon-x4.c",
286 "src/f32-binop/vmulc-neon-x8.c",
287 "src/f32-binop/vrsubc-neon-x4.c",
288 "src/f32-binop/vrsubc-neon-x8.c",
289 "src/f32-binop/vsub-neon-x4.c",
290 "src/f32-binop/vsub-neon-x8.c",
291 "src/f32-binop/vsubc-neon-x4.c",
292 "src/f32-binop/vsubc-neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700293 "src/f32-clamp/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700294 "src/f32-dwconv/up4x9-neon.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800295 "src/f32-dwconv/up4x9-neon-acc2.c",
296 "src/f32-dwconv/up8x9-neon.c",
297 "src/f32-dwconv/up8x9-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700298 "src/f32-gavgpool-spchw/neon-x4.c",
299 "src/f32-gavgpool/mp7p7q-neon.c",
300 "src/f32-gavgpool/up7-neon.c",
301 "src/f32-gemm/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700302 "src/f32-gemm/4x2-neon-ld64.c",
303 "src/f32-gemm/4x8-neon-ld128.c",
304 "src/f32-gemm/4x8-neon-ld64.c",
305 "src/f32-gemm/5x8-neon-ld64.c",
306 "src/f32-gemm/6x8-neon-ld64.c",
307 "src/f32-gemminc/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700308 "src/f32-gemminc/4x8-neon-ld128.c",
309 "src/f32-gemminc/4x8-neon-ld64.c",
310 "src/f32-gemminc/5x8-neon-ld64.c",
311 "src/f32-gemminc/6x8-neon-ld64.c",
312 "src/f32-hswish/neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800313 "src/f32-igemm/1x8-neon-ld64.c",
314 "src/f32-igemm/4x2-neon-ld64.c",
315 "src/f32-igemm/4x4-neon-ld64.c",
316 "src/f32-igemm/4x8-neon-ld128.c",
317 "src/f32-igemm/4x8-neon-ld64.c",
318 "src/f32-igemm/6x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700319 "src/f32-pavgpool/mp9p8q-neon.c",
320 "src/f32-pavgpool/up9-neon.c",
321 "src/f32-ppmm/4x8-neon.c",
322 "src/f32-ppmm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800323 "src/f32-prelu/neon-2x4.c",
324 "src/f32-prelu/neon-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700325 "src/f32-rmax/neon.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800326 "src/f32-vmulcaddc/c4-neon-2x.c",
327 "src/f32-vmulcaddc/c8-neon-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700328 "src/q8-avgpool/mp9p8q-neon.c",
329 "src/q8-avgpool/up9-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700330 "src/q8-dwconv/up8x9-neon.c",
331 "src/q8-gavgpool/mp7p7q-neon.c",
332 "src/q8-gavgpool/up7-neon.c",
333 "src/q8-gemm/4x8-neon.c",
334 "src/q8-gemm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800335 "src/q8-igemm/4x8-neon.c",
336 "src/q8-igemm/8x8-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700337 "src/q8-vadd/neon.c",
338 "src/u8-clamp/neon.c",
339 "src/u8-maxpool/9p8q-neon.c",
340 "src/u8-rmax/neon.c",
341 "src/x32-packx/x4-neon-st4.c",
342 "src/x32-pad/x2-neon.c",
343 "src/x32-zip/x2-neon.c",
344 "src/x32-zip/x3-neon.c",
345 "src/x32-zip/x4-neon.c",
346 "src/x32-zip/xm-neon.c",
347 "src/x8-zip/x2-neon.c",
348 "src/x8-zip/x3-neon.c",
349 "src/x8-zip/x4-neon.c",
350 "src/x8-zip/xm-neon.c",
351]
352
353NEONFMA_UKERNELS = [
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800354 "src/f32-bilinear/neonfma-c4.c",
355 "src/f32-bilinear/neonfma-c8.c",
Frank Barcharddb45b6a2019-10-09 16:42:45 -0700356 "src/f32-igemm/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700357 "src/f32-igemm/4x2-neonfma-ld64.c",
358 "src/f32-igemm/4x4-neonfma-ld64.c",
359 "src/f32-igemm/4x8-neonfma-ld128.c",
360 "src/f32-igemm/4x8-neonfma-ld64.c",
361 "src/f32-igemm/6x8-neonfma-ld64.c",
362 "src/f32-dwconv/up4x9-neonfma.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800363 "src/f32-dwconv/up4x9-neonfma-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700364 "src/f32-dwconv/up8x9-neonfma.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800365 "src/f32-dwconv/up8x9-neonfma-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700366 "src/f32-gemm/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700367 "src/f32-gemm/4x2-neonfma-ld64.c",
368 "src/f32-gemm/4x8-neonfma-ld128.c",
369 "src/f32-gemm/4x8-neonfma-ld64.c",
370 "src/f32-gemm/5x8-neonfma-ld64.c",
371 "src/f32-gemm/6x8-neonfma-ld64.c",
372 "src/f32-gemminc/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700373 "src/f32-gemminc/4x8-neonfma-ld128.c",
374 "src/f32-gemminc/4x8-neonfma-ld64.c",
375 "src/f32-gemminc/5x8-neonfma-ld64.c",
376 "src/f32-gemminc/6x8-neonfma-ld64.c",
377 "src/f32-hswish/neonfma.c",
378 "src/f32-ppmm/4x8-neonfma.c",
379 "src/f32-ppmm/8x8-neonfma.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800380 "src/f32-sigmoid/neonfma-p5-x16.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800381 "src/f32-vmulcaddc/c4-neonfma-2x.c",
382 "src/f32-vmulcaddc/c8-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800383 "src/math/exp-neonfma-lut64-p2.c",
384 "src/math/exp-neonfma-p5.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800385 "src/math/expminus-neonfma-p5.c",
386 "src/math/sigmoid-neonfma-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700387]
388
389AARCH64_NEONFMA_UKERNELS = [
390 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
391 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
392 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
393 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
394 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
395 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
396 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
397 "src/f32-spmm/12x1-neonfma.c",
398 "src/f32-spmm/12x2-neonfma.c",
399 "src/f32-spmm/12x4-neonfma.c",
400 "src/f32-spmm/16x1-neonfma-pipelined.c",
401 "src/f32-spmm/16x1-neonfma-unroll2.c",
402 "src/f32-spmm/16x1-neonfma.c",
403 "src/f32-spmm/16x2-neonfma.c",
404 "src/f32-spmm/16x4-neonfma.c",
405 "src/f32-spmm/4x1-neonfma-pipelined.c",
406 "src/f32-spmm/4x1-neonfma-unroll2.c",
407 "src/f32-spmm/4x1-neonfma.c",
408 "src/f32-spmm/4x2-neonfma.c",
409 "src/f32-spmm/4x4-neonfma.c",
410 "src/f32-spmm/8x1-neonfma-pipelined.c",
411 "src/f32-spmm/8x1-neonfma-unroll2.c",
412 "src/f32-spmm/8x1-neonfma.c",
413 "src/f32-spmm/8x2-neonfma.c",
414 "src/f32-spmm/8x4-neonfma.c",
415]
416
417AARCH64_NEONFP16ARITH_UKERNELS = [
418 "src/f16-gemm/4x8-neonfp16arith-ld64.c",
419 "src/f16-gemm/6x8-neonfp16arith-ld64.c",
420 "src/f16-gemm/8x8-neonfp16arith-ld64.c",
421]
422
423SSE_UKERNELS = [
424 "src/f32-avgpool/mp9p8q-sse.c",
425 "src/f32-avgpool/up9-sse.c",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800426 "src/f32-bilinear/sse-c4.c",
427 "src/f32-bilinear/sse-c8.c",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -0800428 "src/f32-binop/vadd-sse-x4.c",
429 "src/f32-binop/vadd-sse-x8.c",
430 "src/f32-binop/vaddc-sse-x4.c",
431 "src/f32-binop/vaddc-sse-x8.c",
432 "src/f32-binop/vmul-sse-x4.c",
433 "src/f32-binop/vmul-sse-x8.c",
434 "src/f32-binop/vmulc-sse-x4.c",
435 "src/f32-binop/vmulc-sse-x8.c",
436 "src/f32-binop/vrsubc-sse-x4.c",
437 "src/f32-binop/vrsubc-sse-x8.c",
438 "src/f32-binop/vsub-sse-x4.c",
439 "src/f32-binop/vsub-sse-x8.c",
440 "src/f32-binop/vsubc-sse-x4.c",
441 "src/f32-binop/vsubc-sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700442 "src/f32-clamp/sse.c",
443 "src/f32-igemm/1x8-sse-dup.c",
444 "src/f32-igemm/1x8-sse-load1.c",
445 "src/f32-igemm/1x8s4-sse.c",
446 "src/f32-igemm/4x2c4-sse.c",
447 "src/f32-igemm/4x8-sse-dup.c",
448 "src/f32-igemm/4x8-sse-load1.c",
449 "src/f32-igemm/4x8s4-sse.c",
450 "src/f32-dwconv/up4x25-sse.c",
451 "src/f32-dwconv/up4x4-sse.c",
452 "src/f32-dwconv/up4x9-sse.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800453 "src/f32-dwconv/up4x25-sse-acc2.c",
454 "src/f32-dwconv/up4x4-sse-acc2.c",
455 "src/f32-dwconv/up4x9-sse-acc2.c",
456 "src/f32-dwconv/up8x25-sse.c",
457 "src/f32-dwconv/up8x4-sse.c",
458 "src/f32-dwconv/up8x9-sse.c",
459 "src/f32-dwconv/up8x25-sse-acc2.c",
460 "src/f32-dwconv/up8x4-sse-acc2.c",
461 "src/f32-dwconv/up8x9-sse-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700462 "src/f32-gavgpool-spchw/sse-x4.c",
463 "src/f32-gavgpool/mp7p7q-sse.c",
464 "src/f32-gavgpool/up7-sse.c",
465 "src/f32-gemm/1x8-sse-dup.c",
466 "src/f32-gemm/1x8-sse-load1.c",
467 "src/f32-gemm/1x8s4-sse.c",
468 "src/f32-gemm/4x8-sse-dup.c",
469 "src/f32-gemm/4x8-sse-load1.c",
470 "src/f32-gemm/4x8s4-sse.c",
471 "src/f32-gemminc/1x8-sse-dup.c",
472 "src/f32-gemminc/1x8-sse-load1.c",
473 "src/f32-gemminc/1x8s4-sse.c",
474 "src/f32-gemminc/4x8-sse-dup.c",
475 "src/f32-gemminc/4x8-sse-load1.c",
476 "src/f32-gemminc/4x8s4-sse.c",
477 "src/f32-hswish/sse.c",
478 "src/f32-maxpool/9p8q-sse.c",
479 "src/f32-pavgpool/mp9p8q-sse.c",
480 "src/f32-pavgpool/up9-sse.c",
481 "src/f32-dwconv-spchw/3x3p1-sse.c",
482 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
483 "src/f32-ppmm/4x8-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700484 "src/f32-rmax/sse.c",
485 "src/f32-spmm/4x1-sse.c",
486 "src/f32-spmm/8x1-sse.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800487 "src/f32-vmulcaddc/c4-sse-2x.c",
488 "src/f32-vmulcaddc/c8-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700489 "src/x32-packx/x4-sse.c",
490]
491
492SSE2_UKERNELS = [
493 "src/f32-argmaxpool/mp9p8q-sse2.c",
494 "src/f32-argmaxpool/up4-sse2.c",
495 "src/f32-argmaxpool/up9-sse2.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800496 "src/f32-prelu/sse2-2x4.c",
497 "src/f32-prelu/sse2-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700498 "src/q8-avgpool/mp9p8q-sse2.c",
499 "src/q8-avgpool/up9-sse2.c",
500 "src/q8-igemm/4x4c2-sse2.c",
501 "src/q8-dwconv/up8x9-sse2.c",
502 "src/q8-gavgpool/mp7p7q-sse2.c",
503 "src/q8-gavgpool/up7-sse2.c",
504 "src/q8-gemm/2x4c8-sse2.c",
505 "src/q8-gemm/4x4c2-sse2.c",
506 "src/q8-vadd/sse2.c",
507 "src/u8-clamp/sse2.c",
508 "src/u8-maxpool/9p8q-sse2.c",
509 "src/u8-rmax/sse2.c",
510 "src/x32-pad/x2-sse2.c",
511 "src/x32-zip/x2-sse2.c",
512 "src/x32-zip/x3-sse2.c",
513 "src/x32-zip/x4-sse2.c",
514 "src/x32-zip/xm-sse2.c",
515 "src/x8-zip/x2-sse2.c",
516 "src/x8-zip/x3-sse2.c",
517 "src/x8-zip/x4-sse2.c",
518 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -0800519 "src/math/exp-sse2-p5.c",
520 "src/math/expminus-sse2-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700521]
522
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800523SSE41_UKERNELS = [
524 "src/f32-prelu/sse41-2x4.c",
525 "src/f32-prelu/sse41-2x8.c",
526]
527
Marat Dukhan08c4a432019-10-03 09:29:21 -0700528AVX_UKERNELS = [
529 "src/f32-rmax/avx.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700530 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700531]
532
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700533AVX2_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700534 "src/f32-raddexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700535 "src/f32-raddextexp/avx2-p5-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700536 "src/f32-raddstoreexpminusmax/avx2-p5-unroll64.c",
537 "src/f32-vscaleexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700538 "src/f32-vscaleextexp/avx2-p5-unroll64.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700539 "src/math/exp-avx2-p5.c",
540 "src/math/exp-avx2-perm-p3.c",
541 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -0700542 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700543 "src/math/extexp-avx2-p5.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700544]
545
Marat Dukhan08c4a432019-10-03 09:29:21 -0700546AVX512F_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700547 "src/f32-raddexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700548 "src/f32-raddextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700549 "src/f32-raddstoreexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700550 "src/f32-rmax/avx512f.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700551 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700552 "src/f32-vscaleexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700553 "src/f32-vscaleextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700554 "src/math/exp-avx512f-p5-scalef.c",
555 "src/math/exp-avx512f-p5.c",
556 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -0700557 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700558 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700559]
560
561AARCH32_ASM_UKERNELS = [
562 "src/q8-dwconv/up8x9-aarch32-neon.S",
563]
564
565AARCH64_ASM_UKERNELS = [
566 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
567 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
568 "src/f32-gemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700569 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700570 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a57.S",
571 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a75.S",
572 "src/f32-gemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700573 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700574 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a57.S",
575 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a75.S",
576 "src/f32-gemm/4x8-aarch64-neonfma-ld128.S",
577 "src/f32-gemm/4x8-aarch64-neonfma-ld64.S",
578 "src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700579 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700580 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a57.S",
581 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S",
582 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S",
583 "src/f32-gemm/6x8-aarch64-neonfma-ld128.S",
584 "src/f32-gemm/6x8-aarch64-neonfma-ld64.S",
585 "src/f32-gemminc/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700586 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700587 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a57.S",
588 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a75.S",
589 "src/f32-gemminc/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700590 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700591 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a57.S",
592 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a75.S",
593 "src/f32-gemminc/4x8-aarch64-neonfma-ld128.S",
594 "src/f32-gemminc/4x8-aarch64-neonfma-ld64.S",
595 "src/f32-gemminc/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700596 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700597 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a57.S",
598 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a73.S",
599 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a75.S",
600 "src/f32-gemminc/6x8-aarch64-neonfma-ld128.S",
601 "src/f32-gemminc/6x8-aarch64-neonfma-ld64.S",
602 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700603 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700604 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S",
605 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S",
606 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700607 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700608 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
609 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700610 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700611 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
612 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
613 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
614]
615
616INTERNAL_MICROKERNEL_HDRS = [
617 "src/xnnpack/argmaxpool.h",
618 "src/xnnpack/avgpool.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800619 "src/xnnpack/bilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700620 "src/xnnpack/clamp.h",
621 "src/xnnpack/common.h",
622 "src/xnnpack/conv.h",
623 "src/xnnpack/dwconv.h",
624 "src/xnnpack/gavgpool.h",
625 "src/xnnpack/gemm.h",
626 "src/xnnpack/hswish.h",
627 "src/xnnpack/igemm.h",
628 "src/xnnpack/lut.h",
629 "src/xnnpack/math.h",
630 "src/xnnpack/maxpool.h",
631 "src/xnnpack/packx.h",
632 "src/xnnpack/pad.h",
633 "src/xnnpack/params.h",
634 "src/xnnpack/pavgpool.h",
635 "src/xnnpack/ppmm.h",
636 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700637 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700638 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700639 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700640 "src/xnnpack/rmax.h",
641 "src/xnnpack/scalar-utils.h",
642 "src/xnnpack/spmm.h",
643 "src/xnnpack/unpool.h",
644 "src/xnnpack/vadd.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -0800645 "src/xnnpack/vbinop.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700646 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700647 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700648 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700649 "src/xnnpack/vscaleextexp.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800650 "src/xnnpack/vunop.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700651 "src/xnnpack/zip.h",
652]
653
654INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
655 "include/xnnpack.h",
656 "src/xnnpack/allocator.h",
657 "src/xnnpack/compute.h",
658 "src/xnnpack/im2col.h",
659 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700660 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700661 "src/xnnpack/operator.h",
662 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700663 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700664 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700665 "src/xnnpack/requantization.h",
666]
667
668ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
669 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700670]
671
672MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700673 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700674 "include/xnnpack.h",
675]
676
677MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
678 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700679 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700680 "src/xnnpack/requantization.h",
681 "include/xnnpack.h",
682]
683
684OPERATOR_TEST_PARAMS_HDRS = [
685 "src/xnnpack/params.h",
686 "src/xnnpack/common.h",
687]
688
689WEIGHTS_PACK_HDRS = [
690 "src/xnnpack/pack.h",
691 "src/xnnpack/operator.h",
692 "src/xnnpack/compute.h",
693]
694
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700695LOGGING_COPTS = select({
696 # No logging in optimized mode
697 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
698 # Full logging in debug mode
699 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
700 # Error-only logging in default (fastbuild) mode
701 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
702})
703
704LOGGING_HDRS = [
705 "src/xnnpack/log.h",
706]
707
Marat Dukhan08c4a432019-10-03 09:29:21 -0700708xnnpack_cc_library(
709 name = "scalar_ukernels",
710 srcs = SCALAR_UKERNELS,
711 hdrs = INTERNAL_HDRS,
712 aarch32_copts = ["-marm"],
713 copts = xnnpack_std_copts(),
714 deps = [
715 "@FP16",
716 "@FXdiv",
717 ],
718)
719
720xnnpack_cc_library(
721 name = "psimd_ukernels",
722 srcs = PSIMD_UKERNELS,
723 hdrs = INTERNAL_HDRS,
724 aarch32_copts = [
725 "-marm",
726 "-mfpu=neon",
727 ],
728 copts = xnnpack_std_copts(),
729 optimized_copts = [
730 "-O3",
731 "-ffast-math",
732 ],
733 deps = [
734 "@FP16",
735 "@psimd",
736 ],
737)
738
739xnnpack_cc_library(
740 name = "neon_ukernels",
741 hdrs = INTERNAL_HDRS,
742 aarch32_copts = [
743 "-marm",
744 "-mfpu=neon",
745 ],
746 aarch32_srcs = NEON_UKERNELS,
747 aarch64_srcs = NEON_UKERNELS,
748 copts = xnnpack_std_copts(),
749 deps = ["@FP16"],
750)
751
752xnnpack_cc_library(
753 name = "neonfma_ukernels",
754 hdrs = INTERNAL_HDRS,
755 aarch32_copts = [
756 "-marm",
757 "-mfpu=neon-vfpv4",
758 ],
759 aarch32_srcs = NEONFMA_UKERNELS,
760 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
761 copts = xnnpack_std_copts(),
762 deps = ["@FP16"],
763)
764
765xnnpack_cc_library(
766 name = "neonfp16arith_ukernels",
767 hdrs = INTERNAL_HDRS,
768 aarch64_copts = ["-march=armv8.2-a+fp16"],
769 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
770 copts = xnnpack_std_copts(),
771 deps = ["@FP16"],
772)
773
774xnnpack_cc_library(
775 name = "sse2_ukernels",
776 hdrs = INTERNAL_HDRS,
777 copts = xnnpack_std_copts(),
778 x86_copts = ["-msse2"],
779 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
780 deps = ["@FP16"],
781)
782
783xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800784 name = "sse41_ukernels",
785 hdrs = INTERNAL_HDRS,
786 copts = xnnpack_std_copts(),
787 x86_copts = ["-msse4.1"],
788 x86_srcs = SSE41_UKERNELS,
789 deps = ["@FP16"],
790)
791
792xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700793 name = "avx_ukernels",
794 hdrs = INTERNAL_HDRS,
795 copts = xnnpack_std_copts(),
796 x86_copts = ["-mavx"],
797 x86_srcs = AVX_UKERNELS,
798 deps = ["@FP16"],
799)
800
801xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700802 name = "avx2_ukernels",
803 hdrs = INTERNAL_HDRS,
804 copts = xnnpack_std_copts(),
805 x86_copts = [
806 "-mfma",
807 "-mavx2",
808 ],
809 x86_srcs = AVX2_UKERNELS,
810 deps = ["@FP16"],
811)
812
813xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700814 name = "avx512f_ukernels",
815 hdrs = INTERNAL_HDRS,
816 copts = xnnpack_std_copts(),
817 x86_copts = ["-mavx512f"],
818 x86_srcs = AVX512F_UKERNELS,
819 deps = ["@FP16"],
820)
821
822xnnpack_cc_library(
823 name = "asm_ukernels",
824 hdrs = ["src/xnnpack/assembly.h"],
825 aarch32_srcs = AARCH32_ASM_UKERNELS,
826 aarch64_srcs = AARCH64_ASM_UKERNELS,
827)
828
829xnnpack_aggregate_library(
830 name = "ukernels",
831 aarch32_deps = [
832 ":psimd_ukernels",
833 ":neon_ukernels",
834 ":neonfma_ukernels",
835 ":asm_ukernels",
836 ],
837 aarch64_deps = [
838 ":psimd_ukernels",
839 ":neon_ukernels",
840 ":neonfma_ukernels",
841 ":neonfp16arith_ukernels",
842 ":asm_ukernels",
843 ],
844 generic_deps = [":scalar_ukernels"],
845 wasmsimd_deps = [
846 ":psimd_ukernels",
847 ],
848 x86_deps = [
849 ":psimd_ukernels",
850 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800851 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700852 ":avx_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700853 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700854 ":avx512f_ukernels",
855 ],
856)
857
858xnnpack_cc_library(
859 name = "im2col",
860 srcs = ["src/im2col.c"],
861 hdrs = [
862 "src/xnnpack/common.h",
863 "src/xnnpack/im2col.h",
864 ],
865 copts = xnnpack_std_copts(),
866)
867
868xnnpack_cc_library(
869 name = "indirection",
870 srcs = ["src/indirection.c"],
871 hdrs = INTERNAL_HDRS,
872 copts = xnnpack_std_copts(),
873 deps = [
874 "@FP16",
875 "@FXdiv",
876 "@pthreadpool",
877 ],
878)
879
880xnnpack_cc_library(
881 name = "operator_run",
882 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700883 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
884 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700885 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
886 "-Wno-vla",
887 ],
888 deps = [
889 "@FP16",
890 "@FXdiv",
891 "@clog",
892 "@pthreadpool",
893 ],
894)
895
896cc_library(
897 name = "enable_assembly",
898 defines = select({
899 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
900 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -0700901 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700902 }),
903)
904
Marat Dukhancf056b22019-10-07 10:26:29 -0700905xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700906 name = "operators",
907 srcs = OPERATOR_SRCS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700908 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -0700909 ],
910 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700911 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700912 "-Isrc",
913 "-Iinclude",
914 ] + select({
915 ":debug_build": [],
916 "//conditions:default": xnnpack_min_size_copts(),
917 }),
Marat Dukhancf056b22019-10-07 10:26:29 -0700918 wasm_srcs = ["src/wasm-stubs.c"],
919 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700920 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700921 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700922 "@FP16",
923 "@FXdiv",
924 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700925 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700926 ],
927)
928
929cc_library(
930 name = "XNNPACK",
931 srcs = [
932 "src/init.c",
933 ],
934 copts = xnnpack_std_copts() + LOGGING_COPTS + [
935 "-Isrc",
936 "-Iinclude",
937 ] + select({
938 ":debug_build": [],
939 "//conditions:default": xnnpack_min_size_copts(),
940 }),
941 includes = ["include"],
942 linkstatic = True,
943 textual_hdrs = ["include/xnnpack.h"],
944 visibility = xnnpack_visibility(),
945 deps = [
946 ":enable_assembly",
947 ":ukernels",
948 ":operator_run",
949 ":operators",
950 "@clog",
951 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -0700952 ] + select({
953 ":emscripten": [],
954 "//conditions:default": ["@cpuinfo"],
955 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700956)
957
958cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700959 name = "xnnpack_operators_nhwc_f32",
960 srcs = [
961 "src/init.c",
962 ],
963 copts = xnnpack_std_copts() + LOGGING_COPTS + [
964 "-Isrc",
965 "-Iinclude",
966 ] + select({
967 ":debug_build": [],
968 "//conditions:default": xnnpack_min_size_copts(),
969 }),
970 defines = [
971 "XNN_NO_Q8_OPERATORS",
972 "XNN_NO_U8_OPERATORS",
973 "XNN_NO_X8_OPERATORS",
974 "XNN_NO_SPNCHW_OPERATORS",
975 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700976 includes = ["include"],
977 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700978 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700979 visibility = xnnpack_visibility(),
980 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700981 ":enable_assembly",
982 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700983 ":operator_run",
984 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700985 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700986 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700987 ] + select({
988 ":emscripten": [],
989 "//conditions:default": ["@cpuinfo"],
990 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700991)
992
Marat Dukhancf056b22019-10-07 10:26:29 -0700993xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700994 name = "bench_utils",
995 srcs = ["bench/utils.cc"],
996 hdrs = ["bench/utils.h"],
997 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -0800998 deps = [
999 "@com_google_benchmark//:benchmark",
1000 "@cpuinfo",
1001 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001002)
1003
Frank Barchard7e955972019-10-11 10:34:25 -07001004######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07001005
1006xnnpack_benchmark(
1007 name = "q8_gemm_bench",
1008 srcs = [
1009 "bench/gemm.h",
1010 "bench/q8-gemm.cc",
1011 "src/xnnpack/AlignedAllocator.h",
1012 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1013 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
1014 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
1015)
1016
1017xnnpack_benchmark(
1018 name = "f16_gemm_bench",
1019 srcs = [
1020 "bench/f16-gemm.cc",
1021 "bench/gemm.h",
1022 "src/xnnpack/AlignedAllocator.h",
1023 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1024 copts = ["-Wno-unused-function"],
1025 deps = MICROKERNEL_BENCHMARK_DEPS,
1026)
1027
1028xnnpack_benchmark(
1029 name = "f32_igemm_bench",
1030 srcs = [
1031 "bench/f32-igemm.cc",
1032 "bench/conv.h",
1033 "src/xnnpack/AlignedAllocator.h",
1034 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07001035 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001036)
1037
1038xnnpack_benchmark(
1039 name = "f32_conv_hwc_bench",
1040 srcs = [
1041 "bench/f32-conv-hwc.cc",
1042 "bench/dconv.h",
1043 "src/xnnpack/AlignedAllocator.h",
1044 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1045 copts = ["-Wno-unused-function"],
1046 deps = MICROKERNEL_BENCHMARK_DEPS,
1047)
1048
1049xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07001050 name = "f32_conv_hwc2spchw_bench",
1051 srcs = [
1052 "bench/f32-conv-hwc2spchw.cc",
1053 "bench/dconv.h",
1054 "src/xnnpack/AlignedAllocator.h",
1055 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1056 copts = ["-Wno-unused-function"],
1057 deps = MICROKERNEL_BENCHMARK_DEPS,
1058)
1059
1060xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001061 name = "f32_dwconv_bench",
1062 srcs = [
1063 "bench/f32-dwconv.cc",
1064 "bench/dwconv.h",
1065 "src/xnnpack/AlignedAllocator.h",
1066 ] + MICROKERNEL_BENCHMARK_HDRS,
1067 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1068)
1069
1070xnnpack_benchmark(
1071 name = "f32_dwconv_spchw_bench",
1072 srcs = [
1073 "bench/f32-dwconv-spchw.cc",
1074 "bench/dwconv.h",
1075 "src/xnnpack/AlignedAllocator.h",
1076 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1077 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1078)
1079
1080xnnpack_benchmark(
1081 name = "f32_gemm_bench",
1082 srcs = [
1083 "bench/f32-gemm.cc",
1084 "bench/gemm.h",
1085 "src/xnnpack/AlignedAllocator.h",
1086 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1087 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07001088 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001089)
1090
1091xnnpack_benchmark(
1092 name = "f32_rmax_bench",
1093 srcs = [
1094 "bench/f32-rmax.cc",
1095 "src/xnnpack/AlignedAllocator.h",
1096 ] + MICROKERNEL_BENCHMARK_HDRS,
1097 deps = MICROKERNEL_BENCHMARK_DEPS,
1098)
1099
1100xnnpack_benchmark(
1101 name = "f32_spmm_bench",
1102 srcs = [
1103 "bench/f32-spmm.cc",
1104 "bench/gemm.h",
1105 "src/xnnpack/AlignedAllocator.h",
1106 ] + MICROKERNEL_BENCHMARK_HDRS,
1107 copts = ["-Wno-unused-function"],
1108 deps = MICROKERNEL_BENCHMARK_DEPS,
1109)
1110
1111xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07001112 name = "f32_softargmax_bench",
1113 srcs = [
1114 "bench/f32-softargmax.cc",
1115 ] + MICROKERNEL_BENCHMARK_HDRS,
1116 copts = ["-Wno-unused-function"],
1117 deps = MICROKERNEL_BENCHMARK_DEPS,
1118)
1119
1120xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001121 name = "f32_im2col_gemm_bench",
1122 srcs = [
1123 "bench/f32-im2col-gemm.cc",
1124 "bench/conv.h",
1125 "src/xnnpack/AlignedAllocator.h",
1126 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1127 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
1128)
1129
1130########################### Benchmarks for operators ###########################
1131
1132xnnpack_benchmark(
1133 name = "add_bench",
1134 srcs = ["bench/add.cc"],
1135 deps = OPERATOR_BENCHMARK_DEPS,
1136)
1137
1138xnnpack_benchmark(
1139 name = "average_pooling_bench",
1140 srcs = ["bench/average-pooling.cc"],
1141 deps = OPERATOR_BENCHMARK_DEPS,
1142)
1143
1144xnnpack_benchmark(
1145 name = "channel_shuffle_bench",
1146 srcs = ["bench/channel-shuffle.cc"],
1147 deps = OPERATOR_BENCHMARK_DEPS,
1148)
1149
1150xnnpack_benchmark(
1151 name = "convolution_bench",
1152 srcs = ["bench/convolution.cc"],
1153 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1154 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1155)
1156
1157xnnpack_benchmark(
1158 name = "deconvolution_bench",
1159 srcs = ["bench/deconvolution.cc"],
1160 copts = xnnpack_optional_tflite_copts(),
1161 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1162)
1163
1164xnnpack_benchmark(
1165 name = "global_average_pooling_bench",
1166 srcs = ["bench/global-average-pooling.cc"],
1167 deps = OPERATOR_BENCHMARK_DEPS,
1168)
1169
1170xnnpack_benchmark(
1171 name = "max_pooling_bench",
1172 srcs = ["bench/max-pooling.cc"],
1173 deps = OPERATOR_BENCHMARK_DEPS,
1174)
1175
1176xnnpack_benchmark(
1177 name = "sigmoid_bench",
1178 srcs = ["bench/sigmoid.cc"],
1179 deps = OPERATOR_BENCHMARK_DEPS,
1180)
1181
1182xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07001183 name = "prelu_bench",
1184 srcs = ["bench/prelu.cc"],
1185 copts = xnnpack_optional_tflite_copts(),
1186 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1187)
1188
1189xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001190 name = "softargmax_bench",
1191 srcs = ["bench/softargmax.cc"],
1192 deps = OPERATOR_BENCHMARK_DEPS,
1193)
1194
Marat Dukhanc068bb62019-10-04 13:24:39 -07001195############################# End-to-end benchmarks ############################
1196
1197cc_library(
1198 name = "mobilenet_v1",
1199 srcs = ["models/mobilenet-v1.cc"],
1200 hdrs = ["models/models.h"],
1201 linkstatic = True,
1202 deps = [
1203 ":XNNPACK",
1204 "@pthreadpool",
1205 ],
1206)
1207
1208cc_library(
1209 name = "mobilenet_v2",
1210 srcs = ["models/mobilenet-v2.cc"],
1211 hdrs = ["models/models.h"],
1212 linkstatic = True,
1213 deps = [
1214 ":XNNPACK",
1215 "@pthreadpool",
1216 ],
1217)
1218
1219xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07001220 name = "f32_dwconv_e2e_bench",
1221 srcs = ["bench/f32-dwconv-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1222 copts = ["-Wno-unused-function"],
1223 deps = MICROKERNEL_BENCHMARK_DEPS + [
1224 ":XNNPACK",
1225 ":mobilenet_v1",
1226 ":mobilenet_v2",
1227 ],
1228)
1229
1230xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07001231 name = "f32_gemm_e2e_bench",
1232 srcs = ["bench/f32-gemm-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1233 copts = ["-Wno-unused-function"],
1234 deps = MICROKERNEL_BENCHMARK_DEPS + [
1235 ":XNNPACK",
1236 ":mobilenet_v1",
1237 ":mobilenet_v2",
1238 ],
1239)
1240
1241xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07001242 name = "end2end_bench",
1243 srcs = ["bench/end2end.cc"],
1244 deps = [
1245 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07001246 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07001247 ":mobilenet_v1",
1248 ":mobilenet_v2",
1249 "@pthreadpool",
1250 ],
1251)
1252
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001253#################### Accuracy evaluation for math functions ####################
1254
1255xnnpack_benchmark(
1256 name = "f32_exp_eval",
1257 srcs = [
1258 "eval/f32-exp.cc",
1259 "src/xnnpack/AlignedAllocator.h",
1260 ] + ACCURACY_EVAL_HDRS,
1261 deps = ACCURACY_EVAL_DEPS,
1262)
1263
Marat Dukhan515c9772019-10-17 18:07:57 -07001264xnnpack_benchmark(
1265 name = "f32_expminus_eval",
1266 srcs = [
1267 "eval/f32-expminus.cc",
1268 "src/xnnpack/AlignedAllocator.h",
1269 ] + ACCURACY_EVAL_HDRS,
1270 deps = ACCURACY_EVAL_DEPS,
1271)
1272
Marat Dukhan98ba4412019-10-23 02:14:28 -07001273xnnpack_benchmark(
1274 name = "f32_extexp_eval",
1275 srcs = [
1276 "eval/f32-extexp.cc",
1277 "src/xnnpack/AlignedAllocator.h",
1278 ] + ACCURACY_EVAL_HDRS,
1279 deps = ACCURACY_EVAL_DEPS,
1280)
1281
Marat Dukhan346a9e52019-11-15 09:06:30 -08001282xnnpack_benchmark(
1283 name = "f32_sigmoid_eval",
1284 srcs = [
1285 "eval/f32-sigmoid.cc",
1286 "src/xnnpack/AlignedAllocator.h",
1287 ] + ACCURACY_EVAL_HDRS,
1288 deps = ACCURACY_EVAL_DEPS,
1289)
1290
Marat Dukhan08c4a432019-10-03 09:29:21 -07001291######################### Unit tests for micro-kernels #########################
1292
1293xnnpack_unit_test(
1294 name = "f16_gemm_test",
1295 srcs = [
1296 "test/f16-gemm.cc",
1297 "test/gemm-microkernel-tester.h",
1298 "src/xnnpack/AlignedAllocator.h",
1299 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1300 deps = MICROKERNEL_TEST_DEPS,
1301)
1302
1303xnnpack_unit_test(
1304 name = "f32_argmaxpool_test",
1305 srcs = [
1306 "test/f32-argmaxpool.cc",
1307 "test/argmaxpool-microkernel-tester.h",
1308 "src/xnnpack/AlignedAllocator.h",
1309 ] + MICROKERNEL_TEST_HDRS,
1310 deps = MICROKERNEL_TEST_DEPS,
1311)
1312
1313xnnpack_unit_test(
1314 name = "f32_avgpool_test",
1315 srcs = [
1316 "test/f32-avgpool.cc",
1317 "test/avgpool-microkernel-tester.h",
1318 "src/xnnpack/AlignedAllocator.h",
1319 ] + MICROKERNEL_TEST_HDRS,
1320 deps = MICROKERNEL_TEST_DEPS,
1321)
1322
1323xnnpack_unit_test(
Marat Dukhan35dacfb2019-11-07 19:18:16 -08001324 name = "f32_bilinear_test",
1325 srcs = [
1326 "test/f32-bilinear.cc",
1327 "test/bilinear-microkernel-tester.h",
1328 "src/xnnpack/AlignedAllocator.h",
1329 ] + MICROKERNEL_TEST_HDRS,
1330 deps = MICROKERNEL_TEST_DEPS,
1331)
1332
1333xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001334 name = "f32_clamp_test",
1335 srcs = [
1336 "test/f32-clamp.cc",
1337 "test/clamp-microkernel-tester.h",
1338 ] + MICROKERNEL_TEST_HDRS,
1339 deps = MICROKERNEL_TEST_DEPS,
1340)
1341
1342xnnpack_unit_test(
1343 name = "f32_igemm_test",
1344 srcs = [
1345 "test/f32-igemm.cc",
1346 "test/gemm-microkernel-tester.h",
1347 "src/xnnpack/AlignedAllocator.h",
1348 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1349 deps = MICROKERNEL_TEST_DEPS,
1350)
1351
1352xnnpack_unit_test(
1353 name = "f32_conv_hwc_test",
1354 srcs = [
1355 "test/f32-conv-hwc.cc",
1356 "test/conv-hwc-microkernel-tester.h",
1357 "src/xnnpack/AlignedAllocator.h",
1358 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1359 deps = MICROKERNEL_TEST_DEPS,
1360)
1361
1362xnnpack_unit_test(
1363 name = "f32_conv_hwc2spchw_test",
1364 srcs = [
1365 "test/f32-conv-hwc2spchw.cc",
1366 "test/conv-hwc2spchw-microkernel-tester.h",
1367 "src/xnnpack/AlignedAllocator.h",
1368 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1369 deps = MICROKERNEL_TEST_DEPS,
1370)
1371
1372xnnpack_unit_test(
1373 name = "f32_dwconv_test",
1374 srcs = [
1375 "test/f32-dwconv.cc",
1376 "test/dwconv-microkernel-tester.h",
1377 "src/xnnpack/AlignedAllocator.h",
1378 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1379 deps = MICROKERNEL_TEST_DEPS,
1380)
1381
1382xnnpack_unit_test(
1383 name = "f32_dwconv_spchw_test",
1384 srcs = [
1385 "test/f32-dwconv-spchw.cc",
1386 "test/dwconv-spchw-microkernel-tester.h",
1387 "src/xnnpack/AlignedAllocator.h",
1388 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1389 deps = MICROKERNEL_TEST_DEPS,
1390)
1391
1392xnnpack_unit_test(
1393 name = "f32_gavgpool_test",
1394 srcs = [
1395 "test/f32-gavgpool.cc",
1396 "test/gavgpool-microkernel-tester.h",
1397 "src/xnnpack/AlignedAllocator.h",
1398 ] + MICROKERNEL_TEST_HDRS,
1399 deps = MICROKERNEL_TEST_DEPS,
1400)
1401
1402xnnpack_unit_test(
1403 name = "f32_gavgpool_spchw_test",
1404 srcs = [
1405 "test/f32-gavgpool-spchw.cc",
1406 "test/gavgpool-spchw-microkernel-tester.h",
1407 "src/xnnpack/AlignedAllocator.h",
1408 ] + MICROKERNEL_TEST_HDRS,
1409 deps = MICROKERNEL_TEST_DEPS,
1410)
1411
1412xnnpack_unit_test(
1413 name = "f32_gemm_test",
1414 srcs = [
1415 "test/f32-gemm.cc",
1416 "test/gemm-microkernel-tester.h",
1417 "src/xnnpack/AlignedAllocator.h",
1418 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1419 deps = MICROKERNEL_TEST_DEPS,
1420)
1421
1422xnnpack_unit_test(
1423 name = "f32_gemminc_test",
1424 srcs = [
1425 "test/f32-gemminc.cc",
1426 "test/gemm-microkernel-tester.h",
1427 "src/xnnpack/AlignedAllocator.h",
1428 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1429 deps = MICROKERNEL_TEST_DEPS,
1430)
1431
1432xnnpack_unit_test(
1433 name = "f32_hswish_test",
1434 srcs = [
1435 "test/f32-hswish.cc",
1436 "test/hswish-microkernel-tester.h",
1437 ] + MICROKERNEL_TEST_HDRS,
1438 deps = MICROKERNEL_TEST_DEPS,
1439)
1440
1441xnnpack_unit_test(
1442 name = "f32_maxpool_test",
1443 srcs = [
1444 "test/f32-maxpool.cc",
1445 "test/maxpool-microkernel-tester.h",
1446 ] + MICROKERNEL_TEST_HDRS,
1447 deps = MICROKERNEL_TEST_DEPS,
1448)
1449
1450xnnpack_unit_test(
1451 name = "f32_pavgpool_test",
1452 srcs = [
1453 "test/f32-pavgpool.cc",
1454 "test/avgpool-microkernel-tester.h",
1455 "src/xnnpack/AlignedAllocator.h",
1456 ] + MICROKERNEL_TEST_HDRS,
1457 deps = MICROKERNEL_TEST_DEPS,
1458)
1459
1460xnnpack_unit_test(
1461 name = "f32_ppmm_test",
1462 srcs = [
1463 "test/f32-ppmm.cc",
1464 "test/gemm-microkernel-tester.h",
1465 "src/xnnpack/AlignedAllocator.h",
1466 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1467 deps = MICROKERNEL_TEST_DEPS,
1468)
1469
1470xnnpack_unit_test(
1471 name = "f32_prelu_test",
1472 srcs = [
1473 "test/f32-prelu.cc",
1474 "test/prelu-microkernel-tester.h",
1475 "src/xnnpack/AlignedAllocator.h",
1476 ] + MICROKERNEL_TEST_HDRS,
1477 deps = MICROKERNEL_TEST_DEPS,
1478)
1479
1480xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001481 name = "f32_raddexpminusmax_test",
1482 srcs = [
1483 "test/f32-raddexpminusmax.cc",
1484 "test/raddexpminusmax-microkernel-tester.h",
1485 ] + MICROKERNEL_TEST_HDRS,
1486 deps = MICROKERNEL_TEST_DEPS,
1487)
1488
1489xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001490 name = "f32_raddextexp_test",
1491 srcs = [
1492 "test/f32-raddextexp.cc",
1493 "test/raddextexp-microkernel-tester.h",
1494 ] + MICROKERNEL_TEST_HDRS,
1495 deps = MICROKERNEL_TEST_DEPS,
1496)
1497
1498xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001499 name = "f32_raddstoreexpminusmax_test",
1500 srcs = [
1501 "test/f32-raddstoreexpminusmax.cc",
1502 "test/raddstoreexpminusmax-microkernel-tester.h",
1503 ] + MICROKERNEL_TEST_HDRS,
1504 deps = MICROKERNEL_TEST_DEPS,
1505)
1506
1507xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001508 name = "f32_rmax_test",
1509 srcs = [
1510 "test/f32-rmax.cc",
1511 "test/rmax-microkernel-tester.h",
1512 ] + MICROKERNEL_TEST_HDRS,
1513 deps = MICROKERNEL_TEST_DEPS,
1514)
1515
1516xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08001517 name = "f32_sigmoid_test",
1518 srcs = [
1519 "test/f32-sigmoid.cc",
1520 "test/vunop-microkernel-tester.h",
1521 ] + MICROKERNEL_TEST_HDRS,
1522 deps = MICROKERNEL_TEST_DEPS,
1523)
1524
1525xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001526 name = "f32_spmm_test",
1527 srcs = [
1528 "test/f32-spmm.cc",
1529 "test/spmm-microkernel-tester.h",
1530 "src/xnnpack/AlignedAllocator.h",
1531 ] + MICROKERNEL_TEST_HDRS,
1532 deps = MICROKERNEL_TEST_DEPS,
1533)
1534
1535xnnpack_unit_test(
1536 name = "f32_vadd_test",
1537 srcs = [
1538 "test/f32-vadd.cc",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001539 "test/vbinop-microkernel-tester.h",
1540 ] + MICROKERNEL_TEST_HDRS,
1541 deps = MICROKERNEL_TEST_DEPS,
1542)
1543
1544xnnpack_unit_test(
1545 name = "f32_vaddc_test",
1546 srcs = [
1547 "test/f32-vaddc.cc",
1548 "test/vbinopc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001549 ] + MICROKERNEL_TEST_HDRS,
1550 deps = MICROKERNEL_TEST_DEPS,
1551)
1552
1553xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001554 name = "f32_vmul_test",
1555 srcs = [
1556 "test/f32-vmul.cc",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001557 "test/vbinop-microkernel-tester.h",
1558 ] + MICROKERNEL_TEST_HDRS,
1559 deps = MICROKERNEL_TEST_DEPS,
1560)
1561
1562xnnpack_unit_test(
1563 name = "f32_vmulc_test",
1564 srcs = [
1565 "test/f32-vmulc.cc",
1566 "test/vbinopc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001567 ] + MICROKERNEL_TEST_HDRS,
1568 deps = MICROKERNEL_TEST_DEPS,
1569)
1570
1571xnnpack_unit_test(
1572 name = "f32_vmulcaddc_test",
1573 srcs = [
1574 "test/f32-vmulcaddc.cc",
1575 "test/vmulcaddc-microkernel-tester.h",
1576 "src/xnnpack/AlignedAllocator.h",
1577 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1578 deps = MICROKERNEL_TEST_DEPS,
1579)
1580
1581xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001582 name = "f32_vscale_test",
1583 srcs = [
1584 "test/f32-vscale.cc",
1585 "test/vscale-microkernel-tester.h",
1586 ] + MICROKERNEL_TEST_HDRS,
1587 deps = MICROKERNEL_TEST_DEPS,
1588)
1589
1590xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001591 name = "f32_vscaleexpminusmax_test",
1592 srcs = [
1593 "test/f32-vscaleexpminusmax.cc",
1594 "test/vscaleexpminusmax-microkernel-tester.h",
1595 ] + MICROKERNEL_TEST_HDRS,
1596 deps = MICROKERNEL_TEST_DEPS,
1597)
1598
1599xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001600 name = "f32_vscaleextexp_test",
1601 srcs = [
1602 "test/f32-vscaleextexp.cc",
1603 "test/vscaleextexp-microkernel-tester.h",
1604 ] + MICROKERNEL_TEST_HDRS,
1605 deps = MICROKERNEL_TEST_DEPS,
1606)
1607
1608xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001609 name = "f32_vsub_test",
1610 srcs = [
1611 "test/f32-vsub.cc",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001612 "test/vbinop-microkernel-tester.h",
1613 ] + MICROKERNEL_TEST_HDRS,
1614 deps = MICROKERNEL_TEST_DEPS,
1615)
1616
1617xnnpack_unit_test(
1618 name = "f32_vsubc_test",
1619 srcs = [
1620 "test/f32-vsubc.cc",
1621 "test/vbinopc-microkernel-tester.h",
1622 ] + MICROKERNEL_TEST_HDRS,
1623 deps = MICROKERNEL_TEST_DEPS,
1624)
1625
1626xnnpack_unit_test(
1627 name = "f32_vrsubc_test",
1628 srcs = [
1629 "test/f32-vrsubc.cc",
1630 "test/vbinopc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001631 ] + MICROKERNEL_TEST_HDRS,
1632 deps = MICROKERNEL_TEST_DEPS,
1633)
1634
1635xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001636 name = "q8_avgpool_test",
1637 srcs = [
1638 "test/q8-avgpool.cc",
1639 "test/avgpool-microkernel-tester.h",
1640 "src/xnnpack/AlignedAllocator.h",
1641 ] + MICROKERNEL_TEST_HDRS,
1642 deps = MICROKERNEL_TEST_DEPS,
1643)
1644
1645xnnpack_unit_test(
1646 name = "q8_igemm_test",
1647 srcs = [
1648 "test/q8-igemm.cc",
1649 "test/gemm-microkernel-tester.h",
1650 "src/xnnpack/AlignedAllocator.h",
1651 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1652 deps = MICROKERNEL_TEST_DEPS,
1653)
1654
1655xnnpack_unit_test(
1656 name = "q8_dwconv_test",
1657 srcs = [
1658 "test/q8-dwconv.cc",
1659 "test/dwconv-microkernel-tester.h",
1660 "src/xnnpack/AlignedAllocator.h",
1661 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1662 deps = MICROKERNEL_TEST_DEPS,
1663)
1664
1665xnnpack_unit_test(
1666 name = "q8_gavgpool_test",
1667 srcs = [
1668 "test/q8-gavgpool.cc",
1669 "test/gavgpool-microkernel-tester.h",
1670 "src/xnnpack/AlignedAllocator.h",
1671 ] + MICROKERNEL_TEST_HDRS,
1672 deps = MICROKERNEL_TEST_DEPS,
1673)
1674
1675xnnpack_unit_test(
1676 name = "q8_gemm_test",
1677 srcs = [
1678 "test/q8-gemm.cc",
1679 "test/gemm-microkernel-tester.h",
1680 "src/xnnpack/AlignedAllocator.h",
1681 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1682 deps = MICROKERNEL_TEST_DEPS,
1683)
1684
1685xnnpack_unit_test(
1686 name = "q8_vadd_test",
1687 srcs = [
1688 "test/q8-vadd.cc",
1689 "test/vadd-microkernel-tester.h",
1690 ] + MICROKERNEL_TEST_HDRS,
1691 deps = MICROKERNEL_TEST_DEPS,
1692)
1693
1694xnnpack_unit_test(
1695 name = "u8_clamp_test",
1696 srcs = [
1697 "test/u8-clamp.cc",
1698 "test/clamp-microkernel-tester.h",
1699 ] + MICROKERNEL_TEST_HDRS,
1700 deps = MICROKERNEL_TEST_DEPS,
1701)
1702
1703xnnpack_unit_test(
1704 name = "u8_lut32norm_test",
1705 srcs = [
1706 "test/u8-lut32norm.cc",
1707 "test/lut-norm-microkernel-tester.h",
1708 ] + MICROKERNEL_TEST_HDRS,
1709 deps = MICROKERNEL_TEST_DEPS,
1710)
1711
1712xnnpack_unit_test(
1713 name = "u8_maxpool_test",
1714 srcs = [
1715 "test/u8-maxpool.cc",
1716 "test/maxpool-microkernel-tester.h",
1717 ] + MICROKERNEL_TEST_HDRS,
1718 deps = MICROKERNEL_TEST_DEPS,
1719)
1720
1721xnnpack_unit_test(
1722 name = "u8_rmax_test",
1723 srcs = [
1724 "test/u8-rmax.cc",
1725 "test/rmax-microkernel-tester.h",
1726 ] + MICROKERNEL_TEST_HDRS,
1727 deps = MICROKERNEL_TEST_DEPS,
1728)
1729
1730xnnpack_unit_test(
1731 name = "x32_packx_test",
1732 srcs = [
1733 "test/x32-packx.cc",
1734 "test/pack-microkernel-tester.h",
1735 "src/xnnpack/AlignedAllocator.h",
1736 ] + MICROKERNEL_TEST_HDRS,
1737 deps = MICROKERNEL_TEST_DEPS,
1738)
1739
1740xnnpack_unit_test(
1741 name = "x32_pad_test",
1742 srcs = [
1743 "test/x32-pad.cc",
1744 "test/pad-microkernel-tester.h",
1745 ] + MICROKERNEL_TEST_HDRS,
1746 deps = MICROKERNEL_TEST_DEPS,
1747)
1748
1749xnnpack_unit_test(
1750 name = "x32_unpool_test",
1751 srcs = [
1752 "test/x32-unpool.cc",
1753 "test/unpool-microkernel-tester.h",
1754 ] + MICROKERNEL_TEST_HDRS,
1755 deps = MICROKERNEL_TEST_DEPS,
1756)
1757
1758xnnpack_unit_test(
1759 name = "x32_zip_test",
1760 srcs = [
1761 "test/x32-zip.cc",
1762 "test/zip-microkernel-tester.h",
1763 ] + MICROKERNEL_TEST_HDRS,
1764 deps = MICROKERNEL_TEST_DEPS,
1765)
1766
1767xnnpack_unit_test(
1768 name = "x8_lut_test",
1769 srcs = [
1770 "test/x8-lut.cc",
1771 "test/lut-microkernel-tester.h",
1772 ] + MICROKERNEL_TEST_HDRS,
1773 deps = MICROKERNEL_TEST_DEPS,
1774)
1775
1776xnnpack_unit_test(
1777 name = "x8_zip_test",
1778 srcs = [
1779 "test/x8-zip.cc",
1780 "test/zip-microkernel-tester.h",
1781 ] + MICROKERNEL_TEST_HDRS,
1782 deps = MICROKERNEL_TEST_DEPS,
1783)
1784
1785########################### Size test for the library ##########################
1786
1787xnnpack_binary(
1788 name = "size_test",
1789 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001790 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001791)
1792
1793########################### Unit tests for operators ###########################
1794
1795xnnpack_unit_test(
1796 name = "add_test",
1797 srcs = [
1798 "test/add.cc",
1799 "test/add-operator-tester.h",
1800 ],
1801 deps = OPERATOR_TEST_DEPS,
1802)
1803
1804xnnpack_unit_test(
1805 name = "argmax_pooling_test",
1806 srcs = [
1807 "test/argmax-pooling.cc",
1808 "test/argmax-pooling-operator-tester.h",
1809 ] + OPERATOR_TEST_PARAMS_HDRS,
1810 deps = OPERATOR_TEST_DEPS,
1811)
1812
1813xnnpack_unit_test(
1814 name = "average_pooling_test",
1815 srcs = [
1816 "test/average-pooling.cc",
1817 "test/average-pooling-operator-tester.h",
1818 ] + OPERATOR_TEST_PARAMS_HDRS,
1819 deps = OPERATOR_TEST_DEPS,
1820)
1821
1822xnnpack_unit_test(
1823 name = "channel_pad_test",
1824 srcs = [
1825 "test/channel-pad.cc",
1826 "test/channel-pad-operator-tester.h",
1827 ] + OPERATOR_TEST_PARAMS_HDRS,
1828 deps = OPERATOR_TEST_DEPS,
1829)
1830
1831xnnpack_unit_test(
1832 name = "channel_shuffle_test",
1833 srcs = [
1834 "test/channel-shuffle.cc",
1835 "test/channel-shuffle-operator-tester.h",
1836 ],
1837 deps = OPERATOR_TEST_DEPS,
1838)
1839
1840xnnpack_unit_test(
1841 name = "clamp_test",
1842 srcs = [
1843 "test/clamp.cc",
1844 "test/clamp-operator-tester.h",
1845 ],
1846 deps = OPERATOR_TEST_DEPS,
1847)
1848
1849xnnpack_unit_test(
1850 name = "convolution_test",
1851 srcs = [
1852 "test/convolution.cc",
1853 "test/convolution-operator-tester.h",
1854 ],
1855 deps = OPERATOR_TEST_DEPS,
1856)
1857
1858xnnpack_unit_test(
1859 name = "convolution_spnchw_test",
1860 srcs = [
1861 "test/convolution-spnchw.cc",
1862 "test/convolution-spnchw-operator-tester.h",
1863 ],
1864 deps = OPERATOR_TEST_DEPS,
1865)
1866
1867xnnpack_unit_test(
1868 name = "deconvolution_test",
1869 srcs = [
1870 "test/deconvolution.cc",
1871 "test/deconvolution-operator-tester.h",
1872 ] + OPERATOR_TEST_PARAMS_HDRS,
1873 deps = OPERATOR_TEST_DEPS,
1874)
1875
1876xnnpack_unit_test(
1877 name = "fully_connected_test",
1878 srcs = [
1879 "test/fully-connected.cc",
1880 "test/fully-connected-operator-tester.h",
1881 ],
1882 deps = OPERATOR_TEST_DEPS,
1883)
1884
1885xnnpack_unit_test(
1886 name = "global_average_pooling_test",
1887 srcs = [
1888 "test/global-average-pooling.cc",
1889 "test/global-average-pooling-operator-tester.h",
1890 ] + OPERATOR_TEST_PARAMS_HDRS,
1891 deps = OPERATOR_TEST_DEPS,
1892)
1893
1894xnnpack_unit_test(
1895 name = "global_average_pooling_spnchw_test",
1896 srcs = [
1897 "test/global-average-pooling-spnchw.cc",
1898 "test/global-average-pooling-spnchw-operator-tester.h",
1899 ],
1900 deps = OPERATOR_TEST_DEPS,
1901)
1902
1903xnnpack_unit_test(
1904 name = "hardswish_test",
1905 srcs = [
1906 "test/hardswish.cc",
1907 "test/hardswish-operator-tester.h",
1908 ],
1909 deps = OPERATOR_TEST_DEPS,
1910)
1911
1912xnnpack_unit_test(
1913 name = "leaky_relu_test",
1914 srcs = [
1915 "test/leaky-relu.cc",
1916 "test/leaky-relu-operator-tester.h",
1917 ],
1918 deps = OPERATOR_TEST_DEPS,
1919)
1920
1921xnnpack_unit_test(
1922 name = "max_pooling_test",
1923 srcs = [
1924 "test/max-pooling.cc",
1925 "test/max-pooling-operator-tester.h",
1926 ] + OPERATOR_TEST_PARAMS_HDRS,
1927 deps = OPERATOR_TEST_DEPS,
1928)
1929
1930xnnpack_unit_test(
Marat Dukhanca2733c2019-11-15 23:21:17 -08001931 name = "multiply_test",
1932 srcs = [
1933 "test/multiply.cc",
1934 "test/multiply-operator-tester.h",
1935 ],
1936 deps = OPERATOR_TEST_DEPS,
1937)
1938
1939xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001940 name = "prelu_test",
1941 srcs = [
1942 "test/prelu.cc",
1943 "test/prelu-operator-tester.h",
1944 ] + OPERATOR_TEST_PARAMS_HDRS,
1945 deps = OPERATOR_TEST_DEPS,
1946)
1947
1948xnnpack_unit_test(
Marat Dukhan69722492019-11-11 19:55:50 -08001949 name = "resize_bilinear_test",
1950 srcs = [
1951 "test/resize-bilinear.cc",
1952 "test/resize-bilinear-operator-tester.h",
1953 ] + OPERATOR_TEST_PARAMS_HDRS,
1954 deps = OPERATOR_TEST_DEPS,
1955)
1956
1957xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001958 name = "sigmoid_test",
1959 srcs = [
1960 "test/sigmoid.cc",
1961 "test/sigmoid-operator-tester.h",
1962 ],
1963 deps = OPERATOR_TEST_DEPS,
1964)
1965
1966xnnpack_unit_test(
1967 name = "softargmax_test",
1968 srcs = [
1969 "test/softargmax.cc",
1970 "test/softargmax-operator-tester.h",
1971 ],
1972 deps = OPERATOR_TEST_DEPS,
1973)
1974
1975xnnpack_unit_test(
1976 name = "unpooling_test",
1977 srcs = [
1978 "test/unpooling.cc",
1979 "test/unpooling-operator-tester.h",
1980 ],
1981 deps = OPERATOR_TEST_DEPS,
1982)
1983
1984############################# Build configurations #############################
1985
Marat Dukhanb8642352019-10-30 15:43:02 -07001986# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07001987config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07001988 name = "xnn_enable_assembly_explicit_true",
1989 define_values = {"xnn_enable_assembly": "true"},
1990)
1991
1992# Disables usage of assembly kernels.
1993config_setting(
1994 name = "xnn_enable_assembly_explicit_false",
1995 define_values = {"xnn_enable_assembly": "false"},
1996)
1997
1998# Builds with -c dbg
1999config_setting(
2000 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002001 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07002002 "compilation_mode": "dbg",
2003 },
2004)
2005
2006# Builds with -c opt
2007config_setting(
2008 name = "optimized_build",
2009 values = {
2010 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002011 },
2012)
2013
2014config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07002015 name = "linux_k8",
2016 values = {"cpu": "k8"},
2017)
2018
2019config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07002020 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07002021 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07002022)
2023
2024config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002025 name = "android",
2026 values = {"crosstool_top": "//external:android/crosstool"},
2027)
2028
2029config_setting(
2030 name = "android_armv7",
2031 values = {
2032 "crosstool_top": "//external:android/crosstool",
2033 "cpu": "armeabi-v7a",
2034 },
2035)
2036
2037config_setting(
2038 name = "android_arm64",
2039 values = {
2040 "crosstool_top": "//external:android/crosstool",
2041 "cpu": "arm64-v8a",
2042 },
2043)
2044
2045config_setting(
2046 name = "android_x86",
2047 values = {
2048 "crosstool_top": "//external:android/crosstool",
2049 "cpu": "x86",
2050 },
2051)
2052
2053config_setting(
2054 name = "android_x86_64",
2055 values = {
2056 "crosstool_top": "//external:android/crosstool",
2057 "cpu": "x86_64",
2058 },
2059)
2060
2061config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07002062 name = "macos_x86_64",
2063 values = {
2064 "apple_platform_type": "macos",
2065 "cpu": "darwin",
2066 },
2067)
2068
2069config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002070 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002071 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07002072)
2073
2074config_setting(
2075 name = "emscripten_wasm",
2076 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002077 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002078 "cpu": "wasm",
2079 },
2080)
2081
2082config_setting(
2083 name = "emscripten_wasmsimd",
2084 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002085 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002086 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07002087 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002088 },
2089)
2090
2091config_setting(
2092 name = "emscripten_asmjs",
2093 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002094 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002095 "cpu": "asmjs",
2096 },
2097)