blob: 0876f4c6b532d160bd63f58637a047c565113053 [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 = [
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",
Marat Dukhan80bafd22019-11-18 10:16:01 -0800386 "src/math/sigmoid-neonfma-p5-nr2fma.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 Dukhan80bafd22019-11-18 10:16:01 -0800521 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700522]
523
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800524SSE41_UKERNELS = [
525 "src/f32-prelu/sse41-2x4.c",
526 "src/f32-prelu/sse41-2x8.c",
527]
528
Marat Dukhan08c4a432019-10-03 09:29:21 -0700529AVX_UKERNELS = [
530 "src/f32-rmax/avx.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700531 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700532]
533
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700534AVX2_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700535 "src/f32-raddexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700536 "src/f32-raddextexp/avx2-p5-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700537 "src/f32-raddstoreexpminusmax/avx2-p5-unroll64.c",
538 "src/f32-vscaleexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700539 "src/f32-vscaleextexp/avx2-p5-unroll64.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700540 "src/math/exp-avx2-p5.c",
541 "src/math/exp-avx2-perm-p3.c",
542 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -0700543 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700544 "src/math/extexp-avx2-p5.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700545]
546
Marat Dukhan08c4a432019-10-03 09:29:21 -0700547AVX512F_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700548 "src/f32-raddexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700549 "src/f32-raddextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700550 "src/f32-raddstoreexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700551 "src/f32-rmax/avx512f.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700552 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700553 "src/f32-vscaleexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700554 "src/f32-vscaleextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700555 "src/math/exp-avx512f-p5-scalef.c",
556 "src/math/exp-avx512f-p5.c",
557 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -0700558 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700559 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700560]
561
562AARCH32_ASM_UKERNELS = [
563 "src/q8-dwconv/up8x9-aarch32-neon.S",
564]
565
566AARCH64_ASM_UKERNELS = [
567 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
568 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
569 "src/f32-gemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700570 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700571 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a57.S",
572 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a75.S",
573 "src/f32-gemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700574 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700575 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a57.S",
576 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a75.S",
577 "src/f32-gemm/4x8-aarch64-neonfma-ld128.S",
578 "src/f32-gemm/4x8-aarch64-neonfma-ld64.S",
579 "src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700580 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700581 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a57.S",
582 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S",
583 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S",
584 "src/f32-gemm/6x8-aarch64-neonfma-ld128.S",
585 "src/f32-gemm/6x8-aarch64-neonfma-ld64.S",
586 "src/f32-gemminc/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700587 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700588 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a57.S",
589 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a75.S",
590 "src/f32-gemminc/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700591 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700592 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a57.S",
593 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a75.S",
594 "src/f32-gemminc/4x8-aarch64-neonfma-ld128.S",
595 "src/f32-gemminc/4x8-aarch64-neonfma-ld64.S",
596 "src/f32-gemminc/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700597 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700598 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a57.S",
599 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a73.S",
600 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a75.S",
601 "src/f32-gemminc/6x8-aarch64-neonfma-ld128.S",
602 "src/f32-gemminc/6x8-aarch64-neonfma-ld64.S",
603 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700604 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700605 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S",
606 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S",
607 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700608 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700609 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
610 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700611 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700612 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
613 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
614 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
615]
616
617INTERNAL_MICROKERNEL_HDRS = [
618 "src/xnnpack/argmaxpool.h",
619 "src/xnnpack/avgpool.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800620 "src/xnnpack/bilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700621 "src/xnnpack/clamp.h",
622 "src/xnnpack/common.h",
623 "src/xnnpack/conv.h",
624 "src/xnnpack/dwconv.h",
625 "src/xnnpack/gavgpool.h",
626 "src/xnnpack/gemm.h",
627 "src/xnnpack/hswish.h",
628 "src/xnnpack/igemm.h",
629 "src/xnnpack/lut.h",
630 "src/xnnpack/math.h",
631 "src/xnnpack/maxpool.h",
632 "src/xnnpack/packx.h",
633 "src/xnnpack/pad.h",
634 "src/xnnpack/params.h",
635 "src/xnnpack/pavgpool.h",
636 "src/xnnpack/ppmm.h",
637 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700638 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700639 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700640 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700641 "src/xnnpack/rmax.h",
642 "src/xnnpack/scalar-utils.h",
643 "src/xnnpack/spmm.h",
644 "src/xnnpack/unpool.h",
645 "src/xnnpack/vadd.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -0800646 "src/xnnpack/vbinop.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700647 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700648 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700649 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700650 "src/xnnpack/vscaleextexp.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800651 "src/xnnpack/vunop.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700652 "src/xnnpack/zip.h",
653]
654
655INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
656 "include/xnnpack.h",
657 "src/xnnpack/allocator.h",
658 "src/xnnpack/compute.h",
659 "src/xnnpack/im2col.h",
660 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700661 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700662 "src/xnnpack/operator.h",
663 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700664 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700665 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700666 "src/xnnpack/requantization.h",
667]
668
669ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
670 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700671]
672
673MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700674 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700675 "include/xnnpack.h",
676]
677
678MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
679 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700680 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700681 "src/xnnpack/requantization.h",
682 "include/xnnpack.h",
683]
684
685OPERATOR_TEST_PARAMS_HDRS = [
686 "src/xnnpack/params.h",
687 "src/xnnpack/common.h",
688]
689
690WEIGHTS_PACK_HDRS = [
691 "src/xnnpack/pack.h",
692 "src/xnnpack/operator.h",
693 "src/xnnpack/compute.h",
694]
695
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700696LOGGING_COPTS = select({
697 # No logging in optimized mode
698 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
699 # Full logging in debug mode
700 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
701 # Error-only logging in default (fastbuild) mode
702 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
703})
704
705LOGGING_HDRS = [
706 "src/xnnpack/log.h",
707]
708
Marat Dukhan08c4a432019-10-03 09:29:21 -0700709xnnpack_cc_library(
710 name = "scalar_ukernels",
711 srcs = SCALAR_UKERNELS,
712 hdrs = INTERNAL_HDRS,
713 aarch32_copts = ["-marm"],
714 copts = xnnpack_std_copts(),
715 deps = [
716 "@FP16",
717 "@FXdiv",
718 ],
719)
720
721xnnpack_cc_library(
722 name = "psimd_ukernels",
723 srcs = PSIMD_UKERNELS,
724 hdrs = INTERNAL_HDRS,
725 aarch32_copts = [
726 "-marm",
727 "-mfpu=neon",
728 ],
729 copts = xnnpack_std_copts(),
730 optimized_copts = [
731 "-O3",
732 "-ffast-math",
733 ],
734 deps = [
735 "@FP16",
736 "@psimd",
737 ],
738)
739
740xnnpack_cc_library(
741 name = "neon_ukernels",
742 hdrs = INTERNAL_HDRS,
743 aarch32_copts = [
744 "-marm",
745 "-mfpu=neon",
746 ],
747 aarch32_srcs = NEON_UKERNELS,
748 aarch64_srcs = NEON_UKERNELS,
749 copts = xnnpack_std_copts(),
750 deps = ["@FP16"],
751)
752
753xnnpack_cc_library(
754 name = "neonfma_ukernels",
755 hdrs = INTERNAL_HDRS,
756 aarch32_copts = [
757 "-marm",
758 "-mfpu=neon-vfpv4",
759 ],
760 aarch32_srcs = NEONFMA_UKERNELS,
761 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
762 copts = xnnpack_std_copts(),
763 deps = ["@FP16"],
764)
765
766xnnpack_cc_library(
767 name = "neonfp16arith_ukernels",
768 hdrs = INTERNAL_HDRS,
769 aarch64_copts = ["-march=armv8.2-a+fp16"],
770 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
771 copts = xnnpack_std_copts(),
772 deps = ["@FP16"],
773)
774
775xnnpack_cc_library(
776 name = "sse2_ukernels",
777 hdrs = INTERNAL_HDRS,
778 copts = xnnpack_std_copts(),
779 x86_copts = ["-msse2"],
780 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
781 deps = ["@FP16"],
782)
783
784xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800785 name = "sse41_ukernels",
786 hdrs = INTERNAL_HDRS,
787 copts = xnnpack_std_copts(),
788 x86_copts = ["-msse4.1"],
789 x86_srcs = SSE41_UKERNELS,
790 deps = ["@FP16"],
791)
792
793xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700794 name = "avx_ukernels",
795 hdrs = INTERNAL_HDRS,
796 copts = xnnpack_std_copts(),
797 x86_copts = ["-mavx"],
798 x86_srcs = AVX_UKERNELS,
799 deps = ["@FP16"],
800)
801
802xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700803 name = "avx2_ukernels",
804 hdrs = INTERNAL_HDRS,
805 copts = xnnpack_std_copts(),
806 x86_copts = [
807 "-mfma",
808 "-mavx2",
809 ],
810 x86_srcs = AVX2_UKERNELS,
811 deps = ["@FP16"],
812)
813
814xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700815 name = "avx512f_ukernels",
816 hdrs = INTERNAL_HDRS,
817 copts = xnnpack_std_copts(),
818 x86_copts = ["-mavx512f"],
819 x86_srcs = AVX512F_UKERNELS,
820 deps = ["@FP16"],
821)
822
823xnnpack_cc_library(
824 name = "asm_ukernels",
825 hdrs = ["src/xnnpack/assembly.h"],
826 aarch32_srcs = AARCH32_ASM_UKERNELS,
827 aarch64_srcs = AARCH64_ASM_UKERNELS,
828)
829
830xnnpack_aggregate_library(
831 name = "ukernels",
832 aarch32_deps = [
833 ":psimd_ukernels",
834 ":neon_ukernels",
835 ":neonfma_ukernels",
836 ":asm_ukernels",
837 ],
838 aarch64_deps = [
839 ":psimd_ukernels",
840 ":neon_ukernels",
841 ":neonfma_ukernels",
842 ":neonfp16arith_ukernels",
843 ":asm_ukernels",
844 ],
845 generic_deps = [":scalar_ukernels"],
846 wasmsimd_deps = [
847 ":psimd_ukernels",
848 ],
849 x86_deps = [
850 ":psimd_ukernels",
851 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800852 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700853 ":avx_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700854 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700855 ":avx512f_ukernels",
856 ],
857)
858
859xnnpack_cc_library(
860 name = "im2col",
861 srcs = ["src/im2col.c"],
862 hdrs = [
863 "src/xnnpack/common.h",
864 "src/xnnpack/im2col.h",
865 ],
866 copts = xnnpack_std_copts(),
867)
868
869xnnpack_cc_library(
870 name = "indirection",
871 srcs = ["src/indirection.c"],
872 hdrs = INTERNAL_HDRS,
873 copts = xnnpack_std_copts(),
874 deps = [
875 "@FP16",
876 "@FXdiv",
877 "@pthreadpool",
878 ],
879)
880
881xnnpack_cc_library(
882 name = "operator_run",
883 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700884 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
885 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700886 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
887 "-Wno-vla",
888 ],
889 deps = [
890 "@FP16",
891 "@FXdiv",
892 "@clog",
893 "@pthreadpool",
894 ],
895)
896
897cc_library(
898 name = "enable_assembly",
899 defines = select({
900 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
901 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -0700902 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700903 }),
904)
905
Marat Dukhancf056b22019-10-07 10:26:29 -0700906xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700907 name = "operators",
908 srcs = OPERATOR_SRCS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700909 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -0700910 ],
911 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700912 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700913 "-Isrc",
914 "-Iinclude",
915 ] + select({
916 ":debug_build": [],
917 "//conditions:default": xnnpack_min_size_copts(),
918 }),
Marat Dukhancf056b22019-10-07 10:26:29 -0700919 wasm_srcs = ["src/wasm-stubs.c"],
920 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700921 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700922 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700923 "@FP16",
924 "@FXdiv",
925 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700926 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700927 ],
928)
929
930cc_library(
931 name = "XNNPACK",
932 srcs = [
933 "src/init.c",
934 ],
935 copts = xnnpack_std_copts() + LOGGING_COPTS + [
936 "-Isrc",
937 "-Iinclude",
938 ] + select({
939 ":debug_build": [],
940 "//conditions:default": xnnpack_min_size_copts(),
941 }),
942 includes = ["include"],
943 linkstatic = True,
944 textual_hdrs = ["include/xnnpack.h"],
945 visibility = xnnpack_visibility(),
946 deps = [
947 ":enable_assembly",
948 ":ukernels",
949 ":operator_run",
950 ":operators",
951 "@clog",
952 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -0700953 ] + select({
954 ":emscripten": [],
955 "//conditions:default": ["@cpuinfo"],
956 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700957)
958
959cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700960 name = "xnnpack_operators_nhwc_f32",
961 srcs = [
962 "src/init.c",
963 ],
964 copts = xnnpack_std_copts() + LOGGING_COPTS + [
965 "-Isrc",
966 "-Iinclude",
967 ] + select({
968 ":debug_build": [],
969 "//conditions:default": xnnpack_min_size_copts(),
970 }),
971 defines = [
972 "XNN_NO_Q8_OPERATORS",
973 "XNN_NO_U8_OPERATORS",
974 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -0800975 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700976 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700977 includes = ["include"],
978 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700979 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700980 visibility = xnnpack_visibility(),
981 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700982 ":enable_assembly",
983 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700984 ":operator_run",
985 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700986 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700987 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700988 ] + select({
989 ":emscripten": [],
990 "//conditions:default": ["@cpuinfo"],
991 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700992)
993
Marat Dukhancf056b22019-10-07 10:26:29 -0700994xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700995 name = "bench_utils",
996 srcs = ["bench/utils.cc"],
997 hdrs = ["bench/utils.h"],
998 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -0800999 deps = [
1000 "@com_google_benchmark//:benchmark",
1001 "@cpuinfo",
1002 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001003)
1004
Frank Barchard7e955972019-10-11 10:34:25 -07001005######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07001006
1007xnnpack_benchmark(
1008 name = "q8_gemm_bench",
1009 srcs = [
1010 "bench/gemm.h",
1011 "bench/q8-gemm.cc",
1012 "src/xnnpack/AlignedAllocator.h",
1013 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1014 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
1015 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
1016)
1017
1018xnnpack_benchmark(
1019 name = "f16_gemm_bench",
1020 srcs = [
1021 "bench/f16-gemm.cc",
1022 "bench/gemm.h",
1023 "src/xnnpack/AlignedAllocator.h",
1024 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1025 copts = ["-Wno-unused-function"],
1026 deps = MICROKERNEL_BENCHMARK_DEPS,
1027)
1028
1029xnnpack_benchmark(
1030 name = "f32_igemm_bench",
1031 srcs = [
1032 "bench/f32-igemm.cc",
1033 "bench/conv.h",
1034 "src/xnnpack/AlignedAllocator.h",
1035 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07001036 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001037)
1038
1039xnnpack_benchmark(
1040 name = "f32_conv_hwc_bench",
1041 srcs = [
1042 "bench/f32-conv-hwc.cc",
1043 "bench/dconv.h",
1044 "src/xnnpack/AlignedAllocator.h",
1045 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1046 copts = ["-Wno-unused-function"],
1047 deps = MICROKERNEL_BENCHMARK_DEPS,
1048)
1049
1050xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07001051 name = "f32_conv_hwc2spchw_bench",
1052 srcs = [
1053 "bench/f32-conv-hwc2spchw.cc",
1054 "bench/dconv.h",
1055 "src/xnnpack/AlignedAllocator.h",
1056 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1057 copts = ["-Wno-unused-function"],
1058 deps = MICROKERNEL_BENCHMARK_DEPS,
1059)
1060
1061xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001062 name = "f32_dwconv_bench",
1063 srcs = [
1064 "bench/f32-dwconv.cc",
1065 "bench/dwconv.h",
1066 "src/xnnpack/AlignedAllocator.h",
1067 ] + MICROKERNEL_BENCHMARK_HDRS,
1068 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1069)
1070
1071xnnpack_benchmark(
1072 name = "f32_dwconv_spchw_bench",
1073 srcs = [
1074 "bench/f32-dwconv-spchw.cc",
1075 "bench/dwconv.h",
1076 "src/xnnpack/AlignedAllocator.h",
1077 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1078 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1079)
1080
1081xnnpack_benchmark(
1082 name = "f32_gemm_bench",
1083 srcs = [
1084 "bench/f32-gemm.cc",
1085 "bench/gemm.h",
1086 "src/xnnpack/AlignedAllocator.h",
1087 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1088 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07001089 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001090)
1091
1092xnnpack_benchmark(
1093 name = "f32_rmax_bench",
1094 srcs = [
1095 "bench/f32-rmax.cc",
1096 "src/xnnpack/AlignedAllocator.h",
1097 ] + MICROKERNEL_BENCHMARK_HDRS,
1098 deps = MICROKERNEL_BENCHMARK_DEPS,
1099)
1100
1101xnnpack_benchmark(
1102 name = "f32_spmm_bench",
1103 srcs = [
1104 "bench/f32-spmm.cc",
1105 "bench/gemm.h",
1106 "src/xnnpack/AlignedAllocator.h",
1107 ] + MICROKERNEL_BENCHMARK_HDRS,
1108 copts = ["-Wno-unused-function"],
1109 deps = MICROKERNEL_BENCHMARK_DEPS,
1110)
1111
1112xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07001113 name = "f32_softargmax_bench",
1114 srcs = [
1115 "bench/f32-softargmax.cc",
1116 ] + MICROKERNEL_BENCHMARK_HDRS,
1117 copts = ["-Wno-unused-function"],
1118 deps = MICROKERNEL_BENCHMARK_DEPS,
1119)
1120
1121xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001122 name = "f32_im2col_gemm_bench",
1123 srcs = [
1124 "bench/f32-im2col-gemm.cc",
1125 "bench/conv.h",
1126 "src/xnnpack/AlignedAllocator.h",
1127 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1128 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
1129)
1130
1131########################### Benchmarks for operators ###########################
1132
1133xnnpack_benchmark(
1134 name = "add_bench",
1135 srcs = ["bench/add.cc"],
1136 deps = OPERATOR_BENCHMARK_DEPS,
1137)
1138
1139xnnpack_benchmark(
1140 name = "average_pooling_bench",
1141 srcs = ["bench/average-pooling.cc"],
1142 deps = OPERATOR_BENCHMARK_DEPS,
1143)
1144
1145xnnpack_benchmark(
1146 name = "channel_shuffle_bench",
1147 srcs = ["bench/channel-shuffle.cc"],
1148 deps = OPERATOR_BENCHMARK_DEPS,
1149)
1150
1151xnnpack_benchmark(
1152 name = "convolution_bench",
1153 srcs = ["bench/convolution.cc"],
1154 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1155 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1156)
1157
1158xnnpack_benchmark(
1159 name = "deconvolution_bench",
1160 srcs = ["bench/deconvolution.cc"],
1161 copts = xnnpack_optional_tflite_copts(),
1162 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1163)
1164
1165xnnpack_benchmark(
1166 name = "global_average_pooling_bench",
1167 srcs = ["bench/global-average-pooling.cc"],
1168 deps = OPERATOR_BENCHMARK_DEPS,
1169)
1170
1171xnnpack_benchmark(
1172 name = "max_pooling_bench",
1173 srcs = ["bench/max-pooling.cc"],
1174 deps = OPERATOR_BENCHMARK_DEPS,
1175)
1176
1177xnnpack_benchmark(
1178 name = "sigmoid_bench",
1179 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08001180 copts = xnnpack_optional_tflite_copts(),
1181 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001182)
1183
1184xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07001185 name = "prelu_bench",
1186 srcs = ["bench/prelu.cc"],
1187 copts = xnnpack_optional_tflite_copts(),
1188 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1189)
1190
1191xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001192 name = "softargmax_bench",
1193 srcs = ["bench/softargmax.cc"],
1194 deps = OPERATOR_BENCHMARK_DEPS,
1195)
1196
Marat Dukhanc068bb62019-10-04 13:24:39 -07001197############################# End-to-end benchmarks ############################
1198
1199cc_library(
1200 name = "mobilenet_v1",
1201 srcs = ["models/mobilenet-v1.cc"],
1202 hdrs = ["models/models.h"],
1203 linkstatic = True,
1204 deps = [
1205 ":XNNPACK",
1206 "@pthreadpool",
1207 ],
1208)
1209
1210cc_library(
1211 name = "mobilenet_v2",
1212 srcs = ["models/mobilenet-v2.cc"],
1213 hdrs = ["models/models.h"],
1214 linkstatic = True,
1215 deps = [
1216 ":XNNPACK",
1217 "@pthreadpool",
1218 ],
1219)
1220
1221xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07001222 name = "f32_dwconv_e2e_bench",
1223 srcs = ["bench/f32-dwconv-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1224 copts = ["-Wno-unused-function"],
1225 deps = MICROKERNEL_BENCHMARK_DEPS + [
1226 ":XNNPACK",
1227 ":mobilenet_v1",
1228 ":mobilenet_v2",
1229 ],
1230)
1231
1232xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07001233 name = "f32_gemm_e2e_bench",
1234 srcs = ["bench/f32-gemm-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1235 copts = ["-Wno-unused-function"],
1236 deps = MICROKERNEL_BENCHMARK_DEPS + [
1237 ":XNNPACK",
1238 ":mobilenet_v1",
1239 ":mobilenet_v2",
1240 ],
1241)
1242
1243xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07001244 name = "end2end_bench",
1245 srcs = ["bench/end2end.cc"],
1246 deps = [
1247 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07001248 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07001249 ":mobilenet_v1",
1250 ":mobilenet_v2",
1251 "@pthreadpool",
1252 ],
1253)
1254
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001255#################### Accuracy evaluation for math functions ####################
1256
1257xnnpack_benchmark(
1258 name = "f32_exp_eval",
1259 srcs = [
1260 "eval/f32-exp.cc",
1261 "src/xnnpack/AlignedAllocator.h",
1262 ] + ACCURACY_EVAL_HDRS,
1263 deps = ACCURACY_EVAL_DEPS,
1264)
1265
Marat Dukhan515c9772019-10-17 18:07:57 -07001266xnnpack_benchmark(
1267 name = "f32_expminus_eval",
1268 srcs = [
1269 "eval/f32-expminus.cc",
1270 "src/xnnpack/AlignedAllocator.h",
1271 ] + ACCURACY_EVAL_HDRS,
1272 deps = ACCURACY_EVAL_DEPS,
1273)
1274
Marat Dukhan98ba4412019-10-23 02:14:28 -07001275xnnpack_benchmark(
1276 name = "f32_extexp_eval",
1277 srcs = [
1278 "eval/f32-extexp.cc",
1279 "src/xnnpack/AlignedAllocator.h",
1280 ] + ACCURACY_EVAL_HDRS,
1281 deps = ACCURACY_EVAL_DEPS,
1282)
1283
Marat Dukhan346a9e52019-11-15 09:06:30 -08001284xnnpack_benchmark(
1285 name = "f32_sigmoid_eval",
1286 srcs = [
1287 "eval/f32-sigmoid.cc",
1288 "src/xnnpack/AlignedAllocator.h",
1289 ] + ACCURACY_EVAL_HDRS,
1290 deps = ACCURACY_EVAL_DEPS,
1291)
1292
Marat Dukhan08c4a432019-10-03 09:29:21 -07001293######################### Unit tests for micro-kernels #########################
1294
1295xnnpack_unit_test(
1296 name = "f16_gemm_test",
1297 srcs = [
1298 "test/f16-gemm.cc",
1299 "test/gemm-microkernel-tester.h",
1300 "src/xnnpack/AlignedAllocator.h",
1301 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1302 deps = MICROKERNEL_TEST_DEPS,
1303)
1304
1305xnnpack_unit_test(
1306 name = "f32_argmaxpool_test",
1307 srcs = [
1308 "test/f32-argmaxpool.cc",
1309 "test/argmaxpool-microkernel-tester.h",
1310 "src/xnnpack/AlignedAllocator.h",
1311 ] + MICROKERNEL_TEST_HDRS,
1312 deps = MICROKERNEL_TEST_DEPS,
1313)
1314
1315xnnpack_unit_test(
1316 name = "f32_avgpool_test",
1317 srcs = [
1318 "test/f32-avgpool.cc",
1319 "test/avgpool-microkernel-tester.h",
1320 "src/xnnpack/AlignedAllocator.h",
1321 ] + MICROKERNEL_TEST_HDRS,
1322 deps = MICROKERNEL_TEST_DEPS,
1323)
1324
1325xnnpack_unit_test(
Marat Dukhan35dacfb2019-11-07 19:18:16 -08001326 name = "f32_bilinear_test",
1327 srcs = [
1328 "test/f32-bilinear.cc",
1329 "test/bilinear-microkernel-tester.h",
1330 "src/xnnpack/AlignedAllocator.h",
1331 ] + MICROKERNEL_TEST_HDRS,
1332 deps = MICROKERNEL_TEST_DEPS,
1333)
1334
1335xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001336 name = "f32_clamp_test",
1337 srcs = [
1338 "test/f32-clamp.cc",
1339 "test/clamp-microkernel-tester.h",
1340 ] + MICROKERNEL_TEST_HDRS,
1341 deps = MICROKERNEL_TEST_DEPS,
1342)
1343
1344xnnpack_unit_test(
1345 name = "f32_igemm_test",
1346 srcs = [
1347 "test/f32-igemm.cc",
1348 "test/gemm-microkernel-tester.h",
1349 "src/xnnpack/AlignedAllocator.h",
1350 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1351 deps = MICROKERNEL_TEST_DEPS,
1352)
1353
1354xnnpack_unit_test(
1355 name = "f32_conv_hwc_test",
1356 srcs = [
1357 "test/f32-conv-hwc.cc",
1358 "test/conv-hwc-microkernel-tester.h",
1359 "src/xnnpack/AlignedAllocator.h",
1360 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1361 deps = MICROKERNEL_TEST_DEPS,
1362)
1363
1364xnnpack_unit_test(
1365 name = "f32_conv_hwc2spchw_test",
1366 srcs = [
1367 "test/f32-conv-hwc2spchw.cc",
1368 "test/conv-hwc2spchw-microkernel-tester.h",
1369 "src/xnnpack/AlignedAllocator.h",
1370 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1371 deps = MICROKERNEL_TEST_DEPS,
1372)
1373
1374xnnpack_unit_test(
1375 name = "f32_dwconv_test",
1376 srcs = [
1377 "test/f32-dwconv.cc",
1378 "test/dwconv-microkernel-tester.h",
1379 "src/xnnpack/AlignedAllocator.h",
1380 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1381 deps = MICROKERNEL_TEST_DEPS,
1382)
1383
1384xnnpack_unit_test(
1385 name = "f32_dwconv_spchw_test",
1386 srcs = [
1387 "test/f32-dwconv-spchw.cc",
1388 "test/dwconv-spchw-microkernel-tester.h",
1389 "src/xnnpack/AlignedAllocator.h",
1390 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1391 deps = MICROKERNEL_TEST_DEPS,
1392)
1393
1394xnnpack_unit_test(
1395 name = "f32_gavgpool_test",
1396 srcs = [
1397 "test/f32-gavgpool.cc",
1398 "test/gavgpool-microkernel-tester.h",
1399 "src/xnnpack/AlignedAllocator.h",
1400 ] + MICROKERNEL_TEST_HDRS,
1401 deps = MICROKERNEL_TEST_DEPS,
1402)
1403
1404xnnpack_unit_test(
1405 name = "f32_gavgpool_spchw_test",
1406 srcs = [
1407 "test/f32-gavgpool-spchw.cc",
1408 "test/gavgpool-spchw-microkernel-tester.h",
1409 "src/xnnpack/AlignedAllocator.h",
1410 ] + MICROKERNEL_TEST_HDRS,
1411 deps = MICROKERNEL_TEST_DEPS,
1412)
1413
1414xnnpack_unit_test(
1415 name = "f32_gemm_test",
1416 srcs = [
1417 "test/f32-gemm.cc",
1418 "test/gemm-microkernel-tester.h",
1419 "src/xnnpack/AlignedAllocator.h",
1420 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1421 deps = MICROKERNEL_TEST_DEPS,
1422)
1423
1424xnnpack_unit_test(
1425 name = "f32_gemminc_test",
1426 srcs = [
1427 "test/f32-gemminc.cc",
1428 "test/gemm-microkernel-tester.h",
1429 "src/xnnpack/AlignedAllocator.h",
1430 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1431 deps = MICROKERNEL_TEST_DEPS,
1432)
1433
1434xnnpack_unit_test(
1435 name = "f32_hswish_test",
1436 srcs = [
1437 "test/f32-hswish.cc",
1438 "test/hswish-microkernel-tester.h",
1439 ] + MICROKERNEL_TEST_HDRS,
1440 deps = MICROKERNEL_TEST_DEPS,
1441)
1442
1443xnnpack_unit_test(
1444 name = "f32_maxpool_test",
1445 srcs = [
1446 "test/f32-maxpool.cc",
1447 "test/maxpool-microkernel-tester.h",
1448 ] + MICROKERNEL_TEST_HDRS,
1449 deps = MICROKERNEL_TEST_DEPS,
1450)
1451
1452xnnpack_unit_test(
1453 name = "f32_pavgpool_test",
1454 srcs = [
1455 "test/f32-pavgpool.cc",
1456 "test/avgpool-microkernel-tester.h",
1457 "src/xnnpack/AlignedAllocator.h",
1458 ] + MICROKERNEL_TEST_HDRS,
1459 deps = MICROKERNEL_TEST_DEPS,
1460)
1461
1462xnnpack_unit_test(
1463 name = "f32_ppmm_test",
1464 srcs = [
1465 "test/f32-ppmm.cc",
1466 "test/gemm-microkernel-tester.h",
1467 "src/xnnpack/AlignedAllocator.h",
1468 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1469 deps = MICROKERNEL_TEST_DEPS,
1470)
1471
1472xnnpack_unit_test(
1473 name = "f32_prelu_test",
1474 srcs = [
1475 "test/f32-prelu.cc",
1476 "test/prelu-microkernel-tester.h",
1477 "src/xnnpack/AlignedAllocator.h",
1478 ] + MICROKERNEL_TEST_HDRS,
1479 deps = MICROKERNEL_TEST_DEPS,
1480)
1481
1482xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001483 name = "f32_raddexpminusmax_test",
1484 srcs = [
1485 "test/f32-raddexpminusmax.cc",
1486 "test/raddexpminusmax-microkernel-tester.h",
1487 ] + MICROKERNEL_TEST_HDRS,
1488 deps = MICROKERNEL_TEST_DEPS,
1489)
1490
1491xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001492 name = "f32_raddextexp_test",
1493 srcs = [
1494 "test/f32-raddextexp.cc",
1495 "test/raddextexp-microkernel-tester.h",
1496 ] + MICROKERNEL_TEST_HDRS,
1497 deps = MICROKERNEL_TEST_DEPS,
1498)
1499
1500xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001501 name = "f32_raddstoreexpminusmax_test",
1502 srcs = [
1503 "test/f32-raddstoreexpminusmax.cc",
1504 "test/raddstoreexpminusmax-microkernel-tester.h",
1505 ] + MICROKERNEL_TEST_HDRS,
1506 deps = MICROKERNEL_TEST_DEPS,
1507)
1508
1509xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001510 name = "f32_rmax_test",
1511 srcs = [
1512 "test/f32-rmax.cc",
1513 "test/rmax-microkernel-tester.h",
1514 ] + MICROKERNEL_TEST_HDRS,
1515 deps = MICROKERNEL_TEST_DEPS,
1516)
1517
1518xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08001519 name = "f32_sigmoid_test",
1520 srcs = [
1521 "test/f32-sigmoid.cc",
1522 "test/vunop-microkernel-tester.h",
1523 ] + MICROKERNEL_TEST_HDRS,
1524 deps = MICROKERNEL_TEST_DEPS,
1525)
1526
1527xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001528 name = "f32_spmm_test",
1529 srcs = [
1530 "test/f32-spmm.cc",
1531 "test/spmm-microkernel-tester.h",
1532 "src/xnnpack/AlignedAllocator.h",
1533 ] + MICROKERNEL_TEST_HDRS,
1534 deps = MICROKERNEL_TEST_DEPS,
1535)
1536
1537xnnpack_unit_test(
1538 name = "f32_vadd_test",
1539 srcs = [
1540 "test/f32-vadd.cc",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001541 "test/vbinop-microkernel-tester.h",
1542 ] + MICROKERNEL_TEST_HDRS,
1543 deps = MICROKERNEL_TEST_DEPS,
1544)
1545
1546xnnpack_unit_test(
1547 name = "f32_vaddc_test",
1548 srcs = [
1549 "test/f32-vaddc.cc",
1550 "test/vbinopc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001551 ] + MICROKERNEL_TEST_HDRS,
1552 deps = MICROKERNEL_TEST_DEPS,
1553)
1554
1555xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001556 name = "f32_vmul_test",
1557 srcs = [
1558 "test/f32-vmul.cc",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001559 "test/vbinop-microkernel-tester.h",
1560 ] + MICROKERNEL_TEST_HDRS,
1561 deps = MICROKERNEL_TEST_DEPS,
1562)
1563
1564xnnpack_unit_test(
1565 name = "f32_vmulc_test",
1566 srcs = [
1567 "test/f32-vmulc.cc",
1568 "test/vbinopc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001569 ] + MICROKERNEL_TEST_HDRS,
1570 deps = MICROKERNEL_TEST_DEPS,
1571)
1572
1573xnnpack_unit_test(
1574 name = "f32_vmulcaddc_test",
1575 srcs = [
1576 "test/f32-vmulcaddc.cc",
1577 "test/vmulcaddc-microkernel-tester.h",
1578 "src/xnnpack/AlignedAllocator.h",
1579 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1580 deps = MICROKERNEL_TEST_DEPS,
1581)
1582
1583xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001584 name = "f32_vscale_test",
1585 srcs = [
1586 "test/f32-vscale.cc",
1587 "test/vscale-microkernel-tester.h",
1588 ] + MICROKERNEL_TEST_HDRS,
1589 deps = MICROKERNEL_TEST_DEPS,
1590)
1591
1592xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001593 name = "f32_vscaleexpminusmax_test",
1594 srcs = [
1595 "test/f32-vscaleexpminusmax.cc",
1596 "test/vscaleexpminusmax-microkernel-tester.h",
1597 ] + MICROKERNEL_TEST_HDRS,
1598 deps = MICROKERNEL_TEST_DEPS,
1599)
1600
1601xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001602 name = "f32_vscaleextexp_test",
1603 srcs = [
1604 "test/f32-vscaleextexp.cc",
1605 "test/vscaleextexp-microkernel-tester.h",
1606 ] + MICROKERNEL_TEST_HDRS,
1607 deps = MICROKERNEL_TEST_DEPS,
1608)
1609
1610xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001611 name = "f32_vsub_test",
1612 srcs = [
1613 "test/f32-vsub.cc",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001614 "test/vbinop-microkernel-tester.h",
1615 ] + MICROKERNEL_TEST_HDRS,
1616 deps = MICROKERNEL_TEST_DEPS,
1617)
1618
1619xnnpack_unit_test(
1620 name = "f32_vsubc_test",
1621 srcs = [
1622 "test/f32-vsubc.cc",
1623 "test/vbinopc-microkernel-tester.h",
1624 ] + MICROKERNEL_TEST_HDRS,
1625 deps = MICROKERNEL_TEST_DEPS,
1626)
1627
1628xnnpack_unit_test(
1629 name = "f32_vrsubc_test",
1630 srcs = [
1631 "test/f32-vrsubc.cc",
1632 "test/vbinopc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001633 ] + MICROKERNEL_TEST_HDRS,
1634 deps = MICROKERNEL_TEST_DEPS,
1635)
1636
1637xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001638 name = "q8_avgpool_test",
1639 srcs = [
1640 "test/q8-avgpool.cc",
1641 "test/avgpool-microkernel-tester.h",
1642 "src/xnnpack/AlignedAllocator.h",
1643 ] + MICROKERNEL_TEST_HDRS,
1644 deps = MICROKERNEL_TEST_DEPS,
1645)
1646
1647xnnpack_unit_test(
1648 name = "q8_igemm_test",
1649 srcs = [
1650 "test/q8-igemm.cc",
1651 "test/gemm-microkernel-tester.h",
1652 "src/xnnpack/AlignedAllocator.h",
1653 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1654 deps = MICROKERNEL_TEST_DEPS,
1655)
1656
1657xnnpack_unit_test(
1658 name = "q8_dwconv_test",
1659 srcs = [
1660 "test/q8-dwconv.cc",
1661 "test/dwconv-microkernel-tester.h",
1662 "src/xnnpack/AlignedAllocator.h",
1663 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1664 deps = MICROKERNEL_TEST_DEPS,
1665)
1666
1667xnnpack_unit_test(
1668 name = "q8_gavgpool_test",
1669 srcs = [
1670 "test/q8-gavgpool.cc",
1671 "test/gavgpool-microkernel-tester.h",
1672 "src/xnnpack/AlignedAllocator.h",
1673 ] + MICROKERNEL_TEST_HDRS,
1674 deps = MICROKERNEL_TEST_DEPS,
1675)
1676
1677xnnpack_unit_test(
1678 name = "q8_gemm_test",
1679 srcs = [
1680 "test/q8-gemm.cc",
1681 "test/gemm-microkernel-tester.h",
1682 "src/xnnpack/AlignedAllocator.h",
1683 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1684 deps = MICROKERNEL_TEST_DEPS,
1685)
1686
1687xnnpack_unit_test(
1688 name = "q8_vadd_test",
1689 srcs = [
1690 "test/q8-vadd.cc",
1691 "test/vadd-microkernel-tester.h",
1692 ] + MICROKERNEL_TEST_HDRS,
1693 deps = MICROKERNEL_TEST_DEPS,
1694)
1695
1696xnnpack_unit_test(
1697 name = "u8_clamp_test",
1698 srcs = [
1699 "test/u8-clamp.cc",
1700 "test/clamp-microkernel-tester.h",
1701 ] + MICROKERNEL_TEST_HDRS,
1702 deps = MICROKERNEL_TEST_DEPS,
1703)
1704
1705xnnpack_unit_test(
1706 name = "u8_lut32norm_test",
1707 srcs = [
1708 "test/u8-lut32norm.cc",
1709 "test/lut-norm-microkernel-tester.h",
1710 ] + MICROKERNEL_TEST_HDRS,
1711 deps = MICROKERNEL_TEST_DEPS,
1712)
1713
1714xnnpack_unit_test(
1715 name = "u8_maxpool_test",
1716 srcs = [
1717 "test/u8-maxpool.cc",
1718 "test/maxpool-microkernel-tester.h",
1719 ] + MICROKERNEL_TEST_HDRS,
1720 deps = MICROKERNEL_TEST_DEPS,
1721)
1722
1723xnnpack_unit_test(
1724 name = "u8_rmax_test",
1725 srcs = [
1726 "test/u8-rmax.cc",
1727 "test/rmax-microkernel-tester.h",
1728 ] + MICROKERNEL_TEST_HDRS,
1729 deps = MICROKERNEL_TEST_DEPS,
1730)
1731
1732xnnpack_unit_test(
1733 name = "x32_packx_test",
1734 srcs = [
1735 "test/x32-packx.cc",
1736 "test/pack-microkernel-tester.h",
1737 "src/xnnpack/AlignedAllocator.h",
1738 ] + MICROKERNEL_TEST_HDRS,
1739 deps = MICROKERNEL_TEST_DEPS,
1740)
1741
1742xnnpack_unit_test(
1743 name = "x32_pad_test",
1744 srcs = [
1745 "test/x32-pad.cc",
1746 "test/pad-microkernel-tester.h",
1747 ] + MICROKERNEL_TEST_HDRS,
1748 deps = MICROKERNEL_TEST_DEPS,
1749)
1750
1751xnnpack_unit_test(
1752 name = "x32_unpool_test",
1753 srcs = [
1754 "test/x32-unpool.cc",
1755 "test/unpool-microkernel-tester.h",
1756 ] + MICROKERNEL_TEST_HDRS,
1757 deps = MICROKERNEL_TEST_DEPS,
1758)
1759
1760xnnpack_unit_test(
1761 name = "x32_zip_test",
1762 srcs = [
1763 "test/x32-zip.cc",
1764 "test/zip-microkernel-tester.h",
1765 ] + MICROKERNEL_TEST_HDRS,
1766 deps = MICROKERNEL_TEST_DEPS,
1767)
1768
1769xnnpack_unit_test(
1770 name = "x8_lut_test",
1771 srcs = [
1772 "test/x8-lut.cc",
1773 "test/lut-microkernel-tester.h",
1774 ] + MICROKERNEL_TEST_HDRS,
1775 deps = MICROKERNEL_TEST_DEPS,
1776)
1777
1778xnnpack_unit_test(
1779 name = "x8_zip_test",
1780 srcs = [
1781 "test/x8-zip.cc",
1782 "test/zip-microkernel-tester.h",
1783 ] + MICROKERNEL_TEST_HDRS,
1784 deps = MICROKERNEL_TEST_DEPS,
1785)
1786
1787########################### Size test for the library ##########################
1788
1789xnnpack_binary(
1790 name = "size_test",
1791 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001792 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001793)
1794
1795########################### Unit tests for operators ###########################
1796
1797xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001798 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001799 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001800 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001801 "test/add-operator-tester.h",
1802 ],
1803 deps = OPERATOR_TEST_DEPS,
1804)
1805
1806xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001807 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001808 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001809 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001810 "test/argmax-pooling-operator-tester.h",
1811 ] + OPERATOR_TEST_PARAMS_HDRS,
1812 deps = OPERATOR_TEST_DEPS,
1813)
1814
1815xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001816 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001817 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001818 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001819 "test/average-pooling-operator-tester.h",
1820 ] + OPERATOR_TEST_PARAMS_HDRS,
1821 deps = OPERATOR_TEST_DEPS,
1822)
1823
1824xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001825 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001826 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001827 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001828 "test/channel-pad-operator-tester.h",
1829 ] + OPERATOR_TEST_PARAMS_HDRS,
1830 deps = OPERATOR_TEST_DEPS,
1831)
1832
1833xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001834 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001835 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001836 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001837 "test/channel-shuffle-operator-tester.h",
1838 ],
1839 deps = OPERATOR_TEST_DEPS,
1840)
1841
1842xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001843 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001844 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001845 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001846 "test/clamp-operator-tester.h",
1847 ],
1848 deps = OPERATOR_TEST_DEPS,
1849)
1850
1851xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001852 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001853 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001854 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001855 "test/convolution-operator-tester.h",
1856 ],
1857 deps = OPERATOR_TEST_DEPS,
1858)
1859
1860xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001861 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001862 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001863 "test/convolution-nchw.cc",
1864 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001865 ],
1866 deps = OPERATOR_TEST_DEPS,
1867)
1868
1869xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001870 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001871 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001872 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001873 "test/deconvolution-operator-tester.h",
1874 ] + OPERATOR_TEST_PARAMS_HDRS,
1875 deps = OPERATOR_TEST_DEPS,
1876)
1877
1878xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001879 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001880 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001881 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001882 "test/fully-connected-operator-tester.h",
1883 ],
1884 deps = OPERATOR_TEST_DEPS,
1885)
1886
1887xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001888 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001889 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001890 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001891 "test/global-average-pooling-operator-tester.h",
1892 ] + OPERATOR_TEST_PARAMS_HDRS,
1893 deps = OPERATOR_TEST_DEPS,
1894)
1895
1896xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001897 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001898 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001899 "test/global-average-pooling-ncw.cc",
1900 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001901 ],
1902 deps = OPERATOR_TEST_DEPS,
1903)
1904
1905xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001906 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001907 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001908 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001909 "test/hardswish-operator-tester.h",
1910 ],
1911 deps = OPERATOR_TEST_DEPS,
1912)
1913
1914xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001915 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001916 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001917 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001918 "test/leaky-relu-operator-tester.h",
1919 ],
1920 deps = OPERATOR_TEST_DEPS,
1921)
1922
1923xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001924 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001925 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001926 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001927 "test/max-pooling-operator-tester.h",
1928 ] + OPERATOR_TEST_PARAMS_HDRS,
1929 deps = OPERATOR_TEST_DEPS,
1930)
1931
1932xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001933 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08001934 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001935 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08001936 "test/multiply-operator-tester.h",
1937 ],
1938 deps = OPERATOR_TEST_DEPS,
1939)
1940
1941xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001942 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001943 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001944 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001945 "test/prelu-operator-tester.h",
1946 ] + OPERATOR_TEST_PARAMS_HDRS,
1947 deps = OPERATOR_TEST_DEPS,
1948)
1949
1950xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001951 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08001952 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001953 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08001954 "test/resize-bilinear-operator-tester.h",
1955 ] + OPERATOR_TEST_PARAMS_HDRS,
1956 deps = OPERATOR_TEST_DEPS,
1957)
1958
1959xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001960 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001961 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001962 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001963 "test/sigmoid-operator-tester.h",
1964 ],
1965 deps = OPERATOR_TEST_DEPS,
1966)
1967
1968xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001969 name = "softargmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001970 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001971 "test/softargmax-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001972 "test/softargmax-operator-tester.h",
1973 ],
1974 deps = OPERATOR_TEST_DEPS,
1975)
1976
1977xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08001978 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001979 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08001980 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001981 "test/unpooling-operator-tester.h",
1982 ],
1983 deps = OPERATOR_TEST_DEPS,
1984)
1985
1986############################# Build configurations #############################
1987
Marat Dukhanb8642352019-10-30 15:43:02 -07001988# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07001989config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07001990 name = "xnn_enable_assembly_explicit_true",
1991 define_values = {"xnn_enable_assembly": "true"},
1992)
1993
1994# Disables usage of assembly kernels.
1995config_setting(
1996 name = "xnn_enable_assembly_explicit_false",
1997 define_values = {"xnn_enable_assembly": "false"},
1998)
1999
2000# Builds with -c dbg
2001config_setting(
2002 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002003 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07002004 "compilation_mode": "dbg",
2005 },
2006)
2007
2008# Builds with -c opt
2009config_setting(
2010 name = "optimized_build",
2011 values = {
2012 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002013 },
2014)
2015
2016config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07002017 name = "linux_k8",
2018 values = {"cpu": "k8"},
2019)
2020
2021config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07002022 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07002023 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07002024)
2025
2026config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002027 name = "android",
2028 values = {"crosstool_top": "//external:android/crosstool"},
2029)
2030
2031config_setting(
2032 name = "android_armv7",
2033 values = {
2034 "crosstool_top": "//external:android/crosstool",
2035 "cpu": "armeabi-v7a",
2036 },
2037)
2038
2039config_setting(
2040 name = "android_arm64",
2041 values = {
2042 "crosstool_top": "//external:android/crosstool",
2043 "cpu": "arm64-v8a",
2044 },
2045)
2046
2047config_setting(
2048 name = "android_x86",
2049 values = {
2050 "crosstool_top": "//external:android/crosstool",
2051 "cpu": "x86",
2052 },
2053)
2054
2055config_setting(
2056 name = "android_x86_64",
2057 values = {
2058 "crosstool_top": "//external:android/crosstool",
2059 "cpu": "x86_64",
2060 },
2061)
2062
2063config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07002064 name = "macos_x86_64",
2065 values = {
2066 "apple_platform_type": "macos",
2067 "cpu": "darwin",
2068 },
2069)
2070
2071config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002072 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002073 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07002074)
2075
2076config_setting(
2077 name = "emscripten_wasm",
2078 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002079 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002080 "cpu": "wasm",
2081 },
2082)
2083
2084config_setting(
2085 name = "emscripten_wasmsimd",
2086 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002087 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002088 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07002089 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002090 },
2091)
2092
2093config_setting(
2094 name = "emscripten_asmjs",
2095 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002096 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002097 "cpu": "asmjs",
2098 },
2099)