blob: 1853960626f6fd30306d9b688296df134fbdffeb [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
9licenses(["notice"])
10
11exports_files(["LICENSE"])
12
13load(":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")
14
15OPERATOR_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",
68 "src/prelu.c",
69 "src/sigmoid.c",
70 "src/softargmax.c",
71 "src/unpooling.c",
72]
73
74SCALAR_UKERNELS = [
75 "src/f32-argmaxpool/mp9p8q-scalar.c",
76 "src/f32-argmaxpool/up4-scalar.c",
77 "src/f32-argmaxpool/up9-scalar.c",
78 "src/f32-avgpool/mp9p8q-scalar.c",
79 "src/f32-avgpool/up9-scalar.c",
80 "src/f32-clamp/scalar.c",
Erich Elsen563df5f2019-10-23 08:02:21 -070081 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070082 "src/f32-igemm/1x4-scalar.c",
83 "src/f32-igemm/2x4-scalar.c",
84 "src/f32-igemm/4x2-scalar.c",
85 "src/f32-igemm/4x4-scalar.c",
86 "src/f32-dwconv/up1x25-scalar.c",
87 "src/f32-dwconv/up1x4-scalar.c",
88 "src/f32-dwconv/up1x9-scalar.c",
Erich Elsen0cc2c532019-10-15 04:44:18 -070089 "src/f32-dwconv-spchw/3x3p1-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -070090 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -070091 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070092 "src/f32-gavgpool/mp7p7q-scalar.c",
93 "src/f32-gavgpool/up7-scalar.c",
94 "src/f32-gemm/1x4-scalar.c",
95 "src/f32-gemm/2x4-scalar.c",
96 "src/f32-gemm/4x2-scalar.c",
97 "src/f32-gemm/4x4-scalar.c",
98 "src/f32-gemminc/1x4-scalar.c",
99 "src/f32-gemminc/2x4-scalar.c",
100 "src/f32-gemminc/4x4-scalar.c",
101 "src/f32-hswish/scalar.c",
102 "src/f32-maxpool/9p8q-scalar.c",
103 "src/f32-pavgpool/mp9p8q-scalar.c",
104 "src/f32-pavgpool/up9-scalar.c",
105 "src/f32-ppmm/2x4-scalar.c",
106 "src/f32-ppmm/3x3-scalar.c",
107 "src/f32-ppmm/4x2-scalar.c",
108 "src/f32-ppmm/4x4-scalar.c",
109 "src/f32-prelu/x4-scalar.c",
110 "src/f32-rmax/scalar.c",
111 "src/f32-spmm/1x1-scalar-pipelined.c",
112 "src/f32-spmm/1x1-scalar-unroll2.c",
113 "src/f32-spmm/1x1-scalar.c",
114 "src/f32-spmm/2x1-scalar-pipelined.c",
115 "src/f32-spmm/2x1-scalar-unroll2.c",
116 "src/f32-spmm/2x1-scalar.c",
117 "src/f32-spmm/4x1-scalar-pipelined.c",
118 "src/f32-spmm/4x1-scalar-unroll2.c",
119 "src/f32-spmm/4x1-scalar.c",
120 "src/f32-spmm/8x1-scalar-pipelined.c",
121 "src/f32-spmm/8x1-scalar-unroll2.c",
122 "src/f32-spmm/8x1-scalar.c",
123 "src/f32-vadd/scalar.c",
124 "src/f32-vmul/scalar.c",
125 "src/f32-vmulcaddc/c1-scalar-x2.c",
126 "src/f32-vsub/scalar.c",
127 "src/q8-avgpool/mp9p8q-scalar.c",
128 "src/q8-avgpool/up9-scalar.c",
129 "src/q8-igemm/2x2-scalar.c",
130 "src/q8-dwconv/up1x9-scalar.c",
131 "src/q8-gavgpool/mp7p7q-scalar.c",
132 "src/q8-gavgpool/up7-scalar.c",
133 "src/q8-gemm/2x2-scalar.c",
134 "src/q8-vadd/scalar.c",
135 "src/u8-clamp/scalar.c",
136 "src/u8-lut32norm/scalar.c",
137 "src/u8-maxpool/9p8q-scalar.c",
138 "src/u8-rmax/scalar.c",
139 "src/x32-packx/x2-scalar.c",
140 "src/x32-packx/x3-scalar.c",
141 "src/x32-packx/x4-scalar.c",
142 "src/x32-pad/x2-scalar.c",
143 "src/x32-unpool/scalar.c",
144 "src/x32-zip/x2-scalar.c",
145 "src/x32-zip/x3-scalar.c",
146 "src/x32-zip/x4-scalar.c",
147 "src/x32-zip/xm-scalar.c",
148 "src/x8-lut/scalar.c",
149 "src/x8-zip/x2-scalar.c",
150 "src/x8-zip/x3-scalar.c",
151 "src/x8-zip/x4-scalar.c",
152 "src/x8-zip/xm-scalar.c",
153]
154
155PSIMD_UKERNELS = [
156 "src/f32-argmaxpool/mp9p8q-psimd.c",
157 "src/f32-argmaxpool/up4-psimd.c",
158 "src/f32-argmaxpool/up9-psimd.c",
159 "src/f32-avgpool/mp9p8q-psimd.c",
160 "src/f32-avgpool/up9-psimd.c",
161 "src/f32-clamp/psimd.c",
162 "src/f32-igemm/1x8-psimd-loadsplat.c",
163 "src/f32-igemm/1x8-psimd-splat.c",
164 "src/f32-igemm/1x8s4-psimd.c",
165 "src/f32-igemm/4x2c4-psimd.c",
166 "src/f32-igemm/4x8-psimd-loadsplat.c",
167 "src/f32-igemm/4x8-psimd-splat.c",
168 "src/f32-igemm/4x8s4-psimd.c",
169 "src/f32-igemm/6x8-psimd-loadsplat.c",
170 "src/f32-igemm/6x8-psimd-splat.c",
171 "src/f32-igemm/6x8s4-psimd.c",
172 "src/f32-dwconv/up4x25-psimd.c",
173 "src/f32-dwconv/up4x4-psimd.c",
174 "src/f32-dwconv/up4x9-psimd.c",
175 "src/f32-gavgpool/mp7p7q-psimd.c",
176 "src/f32-gavgpool/up7-psimd.c",
177 "src/f32-gemm/1x8-psimd-loadsplat.c",
178 "src/f32-gemm/1x8-psimd-splat.c",
179 "src/f32-gemm/1x8s4-psimd.c",
180 "src/f32-gemm/4x8-psimd-loadsplat.c",
181 "src/f32-gemm/4x8-psimd-splat.c",
182 "src/f32-gemm/4x8s4-psimd.c",
183 "src/f32-gemm/6x8-psimd-loadsplat.c",
184 "src/f32-gemm/6x8-psimd-splat.c",
185 "src/f32-gemm/6x8s4-psimd.c",
186 "src/f32-gemminc/1x8-psimd-loadsplat.c",
187 "src/f32-gemminc/1x8-psimd-splat.c",
188 "src/f32-gemminc/1x8s4-psimd.c",
189 "src/f32-gemminc/4x8-psimd-loadsplat.c",
190 "src/f32-gemminc/4x8-psimd-splat.c",
191 "src/f32-gemminc/4x8s4-psimd.c",
192 "src/f32-gemminc/6x8-psimd-loadsplat.c",
193 "src/f32-gemminc/6x8-psimd-splat.c",
194 "src/f32-gemminc/6x8s4-psimd.c",
195 "src/f32-hswish/psimd.c",
196 "src/f32-maxpool/9p8q-psimd.c",
197 "src/f32-pavgpool/mp9p8q-psimd.c",
198 "src/f32-pavgpool/up9-psimd.c",
199 "src/f32-ppmm/4x8-psimd.c",
200 "src/f32-prelu/x4-psimd.c",
201 "src/f32-vadd/psimd.c",
202 "src/f32-vmul/psimd.c",
203 "src/f32-vmulcaddc/c4-psimd-x2.c",
204 "src/f32-vsub/psimd.c",
205 "src/x32-packx/x4-psimd.c",
206 "src/x32-pad/x2-psimd.c",
207 "src/x32-unpool/psimd.c",
208 "src/x32-zip/x2-psimd.c",
209 "src/x32-zip/x3-psimd.c",
210 "src/x32-zip/x4-psimd.c",
211 "src/x32-zip/xm-psimd.c",
212]
213
214# ISA-specific micro-kernels
215NEON_UKERNELS = [
216 "src/f32-avgpool/mp9p8q-neon.c",
217 "src/f32-avgpool/up9-neon.c",
218 "src/f32-clamp/neon.c",
219 "src/f32-igemm/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700220 "src/f32-igemm/4x2-neon-ld64.c",
221 "src/f32-igemm/4x4-neon-ld64.c",
222 "src/f32-igemm/4x8-neon-ld128.c",
223 "src/f32-igemm/4x8-neon-ld64.c",
224 "src/f32-igemm/6x8-neon-ld64.c",
225 "src/f32-dwconv/up4x9-neon.c",
226 "src/f32-gavgpool-spchw/neon-x4.c",
227 "src/f32-gavgpool/mp7p7q-neon.c",
228 "src/f32-gavgpool/up7-neon.c",
229 "src/f32-gemm/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700230 "src/f32-gemm/4x2-neon-ld64.c",
231 "src/f32-gemm/4x8-neon-ld128.c",
232 "src/f32-gemm/4x8-neon-ld64.c",
233 "src/f32-gemm/5x8-neon-ld64.c",
234 "src/f32-gemm/6x8-neon-ld64.c",
235 "src/f32-gemminc/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700236 "src/f32-gemminc/4x8-neon-ld128.c",
237 "src/f32-gemminc/4x8-neon-ld64.c",
238 "src/f32-gemminc/5x8-neon-ld64.c",
239 "src/f32-gemminc/6x8-neon-ld64.c",
240 "src/f32-hswish/neon.c",
241 "src/f32-pavgpool/mp9p8q-neon.c",
242 "src/f32-pavgpool/up9-neon.c",
243 "src/f32-ppmm/4x8-neon.c",
244 "src/f32-ppmm/8x8-neon.c",
245 "src/f32-rmax/neon.c",
246 "src/f32-vmulcaddc/c4-neon-x2.c",
247 "src/q8-avgpool/mp9p8q-neon.c",
248 "src/q8-avgpool/up9-neon.c",
249 "src/q8-igemm/4x8-neon.c",
250 "src/q8-igemm/8x8-neon.c",
251 "src/q8-dwconv/up8x9-neon.c",
252 "src/q8-gavgpool/mp7p7q-neon.c",
253 "src/q8-gavgpool/up7-neon.c",
254 "src/q8-gemm/4x8-neon.c",
255 "src/q8-gemm/8x8-neon.c",
256 "src/q8-vadd/neon.c",
257 "src/u8-clamp/neon.c",
258 "src/u8-maxpool/9p8q-neon.c",
259 "src/u8-rmax/neon.c",
260 "src/x32-packx/x4-neon-st4.c",
261 "src/x32-pad/x2-neon.c",
262 "src/x32-zip/x2-neon.c",
263 "src/x32-zip/x3-neon.c",
264 "src/x32-zip/x4-neon.c",
265 "src/x32-zip/xm-neon.c",
266 "src/x8-zip/x2-neon.c",
267 "src/x8-zip/x3-neon.c",
268 "src/x8-zip/x4-neon.c",
269 "src/x8-zip/xm-neon.c",
270]
271
272NEONFMA_UKERNELS = [
Frank Barcharddb45b6a2019-10-09 16:42:45 -0700273 "src/f32-igemm/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700274 "src/f32-igemm/4x2-neonfma-ld64.c",
275 "src/f32-igemm/4x4-neonfma-ld64.c",
276 "src/f32-igemm/4x8-neonfma-ld128.c",
277 "src/f32-igemm/4x8-neonfma-ld64.c",
278 "src/f32-igemm/6x8-neonfma-ld64.c",
279 "src/f32-dwconv/up4x9-neonfma.c",
280 "src/f32-dwconv/up8x9-neonfma.c",
281 "src/f32-gemm/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700282 "src/f32-gemm/4x2-neonfma-ld64.c",
283 "src/f32-gemm/4x8-neonfma-ld128.c",
284 "src/f32-gemm/4x8-neonfma-ld64.c",
285 "src/f32-gemm/5x8-neonfma-ld64.c",
286 "src/f32-gemm/6x8-neonfma-ld64.c",
287 "src/f32-gemminc/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700288 "src/f32-gemminc/4x8-neonfma-ld128.c",
289 "src/f32-gemminc/4x8-neonfma-ld64.c",
290 "src/f32-gemminc/5x8-neonfma-ld64.c",
291 "src/f32-gemminc/6x8-neonfma-ld64.c",
292 "src/f32-hswish/neonfma.c",
293 "src/f32-ppmm/4x8-neonfma.c",
294 "src/f32-ppmm/8x8-neonfma.c",
295 "src/f32-vmulcaddc/c4-neonfma-x2.c",
296]
297
298AARCH64_NEONFMA_UKERNELS = [
299 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
300 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
301 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
302 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
303 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
304 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
305 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
306 "src/f32-spmm/12x1-neonfma.c",
307 "src/f32-spmm/12x2-neonfma.c",
308 "src/f32-spmm/12x4-neonfma.c",
309 "src/f32-spmm/16x1-neonfma-pipelined.c",
310 "src/f32-spmm/16x1-neonfma-unroll2.c",
311 "src/f32-spmm/16x1-neonfma.c",
312 "src/f32-spmm/16x2-neonfma.c",
313 "src/f32-spmm/16x4-neonfma.c",
314 "src/f32-spmm/4x1-neonfma-pipelined.c",
315 "src/f32-spmm/4x1-neonfma-unroll2.c",
316 "src/f32-spmm/4x1-neonfma.c",
317 "src/f32-spmm/4x2-neonfma.c",
318 "src/f32-spmm/4x4-neonfma.c",
319 "src/f32-spmm/8x1-neonfma-pipelined.c",
320 "src/f32-spmm/8x1-neonfma-unroll2.c",
321 "src/f32-spmm/8x1-neonfma.c",
322 "src/f32-spmm/8x2-neonfma.c",
323 "src/f32-spmm/8x4-neonfma.c",
324]
325
326AARCH64_NEONFP16ARITH_UKERNELS = [
327 "src/f16-gemm/4x8-neonfp16arith-ld64.c",
328 "src/f16-gemm/6x8-neonfp16arith-ld64.c",
329 "src/f16-gemm/8x8-neonfp16arith-ld64.c",
330]
331
332SSE_UKERNELS = [
333 "src/f32-avgpool/mp9p8q-sse.c",
334 "src/f32-avgpool/up9-sse.c",
335 "src/f32-clamp/sse.c",
336 "src/f32-igemm/1x8-sse-dup.c",
337 "src/f32-igemm/1x8-sse-load1.c",
338 "src/f32-igemm/1x8s4-sse.c",
339 "src/f32-igemm/4x2c4-sse.c",
340 "src/f32-igemm/4x8-sse-dup.c",
341 "src/f32-igemm/4x8-sse-load1.c",
342 "src/f32-igemm/4x8s4-sse.c",
343 "src/f32-dwconv/up4x25-sse.c",
344 "src/f32-dwconv/up4x4-sse.c",
345 "src/f32-dwconv/up4x9-sse.c",
346 "src/f32-gavgpool-spchw/sse-x4.c",
347 "src/f32-gavgpool/mp7p7q-sse.c",
348 "src/f32-gavgpool/up7-sse.c",
349 "src/f32-gemm/1x8-sse-dup.c",
350 "src/f32-gemm/1x8-sse-load1.c",
351 "src/f32-gemm/1x8s4-sse.c",
352 "src/f32-gemm/4x8-sse-dup.c",
353 "src/f32-gemm/4x8-sse-load1.c",
354 "src/f32-gemm/4x8s4-sse.c",
355 "src/f32-gemminc/1x8-sse-dup.c",
356 "src/f32-gemminc/1x8-sse-load1.c",
357 "src/f32-gemminc/1x8s4-sse.c",
358 "src/f32-gemminc/4x8-sse-dup.c",
359 "src/f32-gemminc/4x8-sse-load1.c",
360 "src/f32-gemminc/4x8s4-sse.c",
361 "src/f32-hswish/sse.c",
362 "src/f32-maxpool/9p8q-sse.c",
363 "src/f32-pavgpool/mp9p8q-sse.c",
364 "src/f32-pavgpool/up9-sse.c",
365 "src/f32-dwconv-spchw/3x3p1-sse.c",
366 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
367 "src/f32-ppmm/4x8-sse.c",
368 "src/f32-prelu/x4-sse.c",
369 "src/f32-rmax/sse.c",
370 "src/f32-spmm/4x1-sse.c",
371 "src/f32-spmm/8x1-sse.c",
372 "src/f32-vadd/sse.c",
373 "src/f32-vmul/sse.c",
374 "src/f32-vmulcaddc/c4-sse-x2.c",
375 "src/f32-vsub/sse.c",
376 "src/x32-packx/x4-sse.c",
377]
378
379SSE2_UKERNELS = [
380 "src/f32-argmaxpool/mp9p8q-sse2.c",
381 "src/f32-argmaxpool/up4-sse2.c",
382 "src/f32-argmaxpool/up9-sse2.c",
383 "src/q8-avgpool/mp9p8q-sse2.c",
384 "src/q8-avgpool/up9-sse2.c",
385 "src/q8-igemm/4x4c2-sse2.c",
386 "src/q8-dwconv/up8x9-sse2.c",
387 "src/q8-gavgpool/mp7p7q-sse2.c",
388 "src/q8-gavgpool/up7-sse2.c",
389 "src/q8-gemm/2x4c8-sse2.c",
390 "src/q8-gemm/4x4c2-sse2.c",
391 "src/q8-vadd/sse2.c",
392 "src/u8-clamp/sse2.c",
393 "src/u8-maxpool/9p8q-sse2.c",
394 "src/u8-rmax/sse2.c",
395 "src/x32-pad/x2-sse2.c",
396 "src/x32-zip/x2-sse2.c",
397 "src/x32-zip/x3-sse2.c",
398 "src/x32-zip/x4-sse2.c",
399 "src/x32-zip/xm-sse2.c",
400 "src/x8-zip/x2-sse2.c",
401 "src/x8-zip/x3-sse2.c",
402 "src/x8-zip/x4-sse2.c",
403 "src/x8-zip/xm-sse2.c",
404]
405
406AVX_UKERNELS = [
407 "src/f32-rmax/avx.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700408 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700409]
410
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700411AVX2_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700412 "src/f32-raddexpminusmax/avx2-p5-unroll64.c",
413 "src/f32-raddstoreexpminusmax/avx2-p5-unroll64.c",
414 "src/f32-vscaleexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700415 "src/math/exp-avx2-p5.c",
416 "src/math/exp-avx2-perm-p3.c",
417 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -0700418 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700419 "src/math/extexp-avx2-p5.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700420]
421
Marat Dukhan08c4a432019-10-03 09:29:21 -0700422AVX512F_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700423 "src/f32-raddexpminusmax/avx512f-p5-scalef-unroll128.c",
424 "src/f32-raddstoreexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700425 "src/f32-rmax/avx512f.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700426 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700427 "src/f32-vscaleexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700428 "src/math/exp-avx512f-p5-scalef.c",
429 "src/math/exp-avx512f-p5.c",
430 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700431 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700432]
433
434AARCH32_ASM_UKERNELS = [
435 "src/q8-dwconv/up8x9-aarch32-neon.S",
436]
437
438AARCH64_ASM_UKERNELS = [
439 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
440 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
441 "src/f32-gemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700442 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700443 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a57.S",
444 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a75.S",
445 "src/f32-gemm/4x12-aarch64-neonfma-cortex-a53.S",
446 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a57.S",
447 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a75.S",
448 "src/f32-gemm/4x8-aarch64-neonfma-ld128.S",
449 "src/f32-gemm/4x8-aarch64-neonfma-ld64.S",
450 "src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700451 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700452 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a57.S",
453 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S",
454 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S",
455 "src/f32-gemm/6x8-aarch64-neonfma-ld128.S",
456 "src/f32-gemm/6x8-aarch64-neonfma-ld64.S",
457 "src/f32-gemminc/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700458 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700459 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a57.S",
460 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a75.S",
461 "src/f32-gemminc/4x12-aarch64-neonfma-cortex-a53.S",
462 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a57.S",
463 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a75.S",
464 "src/f32-gemminc/4x8-aarch64-neonfma-ld128.S",
465 "src/f32-gemminc/4x8-aarch64-neonfma-ld64.S",
466 "src/f32-gemminc/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700467 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700468 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a57.S",
469 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a73.S",
470 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a75.S",
471 "src/f32-gemminc/6x8-aarch64-neonfma-ld128.S",
472 "src/f32-gemminc/6x8-aarch64-neonfma-ld64.S",
473 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700474 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700475 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S",
476 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S",
477 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
478 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
479 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700480 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700481 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
482 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
483 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
484]
485
486INTERNAL_MICROKERNEL_HDRS = [
487 "src/xnnpack/argmaxpool.h",
488 "src/xnnpack/avgpool.h",
489 "src/xnnpack/clamp.h",
490 "src/xnnpack/common.h",
491 "src/xnnpack/conv.h",
492 "src/xnnpack/dwconv.h",
493 "src/xnnpack/gavgpool.h",
494 "src/xnnpack/gemm.h",
495 "src/xnnpack/hswish.h",
496 "src/xnnpack/igemm.h",
497 "src/xnnpack/lut.h",
498 "src/xnnpack/math.h",
499 "src/xnnpack/maxpool.h",
500 "src/xnnpack/packx.h",
501 "src/xnnpack/pad.h",
502 "src/xnnpack/params.h",
503 "src/xnnpack/pavgpool.h",
504 "src/xnnpack/ppmm.h",
505 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700506 "src/xnnpack/raddexpminusmax.h",
507 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700508 "src/xnnpack/rmax.h",
509 "src/xnnpack/scalar-utils.h",
510 "src/xnnpack/spmm.h",
511 "src/xnnpack/unpool.h",
512 "src/xnnpack/vadd.h",
513 "src/xnnpack/vmul.h",
514 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700515 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700516 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700517 "src/xnnpack/vsub.h",
518 "src/xnnpack/zip.h",
519]
520
521INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
522 "include/xnnpack.h",
523 "src/xnnpack/allocator.h",
524 "src/xnnpack/compute.h",
525 "src/xnnpack/im2col.h",
526 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700527 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700528 "src/xnnpack/operator.h",
529 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700530 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700531 "src/xnnpack/requantization.h",
532]
533
534ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
535 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700536]
537
538MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
539 "src/xnnpack/requantization.h",
540 "include/xnnpack.h",
541]
542
543MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
544 "src/xnnpack/isa-checks.h",
545 "src/xnnpack/requantization.h",
546 "include/xnnpack.h",
547]
548
549OPERATOR_TEST_PARAMS_HDRS = [
550 "src/xnnpack/params.h",
551 "src/xnnpack/common.h",
552]
553
554WEIGHTS_PACK_HDRS = [
555 "src/xnnpack/pack.h",
556 "src/xnnpack/operator.h",
557 "src/xnnpack/compute.h",
558]
559
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700560LOGGING_COPTS = select({
561 # No logging in optimized mode
562 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
563 # Full logging in debug mode
564 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
565 # Error-only logging in default (fastbuild) mode
566 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
567})
568
569LOGGING_HDRS = [
570 "src/xnnpack/log.h",
571]
572
Marat Dukhan08c4a432019-10-03 09:29:21 -0700573xnnpack_cc_library(
574 name = "scalar_ukernels",
575 srcs = SCALAR_UKERNELS,
576 hdrs = INTERNAL_HDRS,
577 aarch32_copts = ["-marm"],
578 copts = xnnpack_std_copts(),
579 deps = [
580 "@FP16",
581 "@FXdiv",
582 ],
583)
584
585xnnpack_cc_library(
586 name = "psimd_ukernels",
587 srcs = PSIMD_UKERNELS,
588 hdrs = INTERNAL_HDRS,
589 aarch32_copts = [
590 "-marm",
591 "-mfpu=neon",
592 ],
593 copts = xnnpack_std_copts(),
594 optimized_copts = [
595 "-O3",
596 "-ffast-math",
597 ],
598 deps = [
599 "@FP16",
600 "@psimd",
601 ],
602)
603
604xnnpack_cc_library(
605 name = "neon_ukernels",
606 hdrs = INTERNAL_HDRS,
607 aarch32_copts = [
608 "-marm",
609 "-mfpu=neon",
610 ],
611 aarch32_srcs = NEON_UKERNELS,
612 aarch64_srcs = NEON_UKERNELS,
613 copts = xnnpack_std_copts(),
614 deps = ["@FP16"],
615)
616
617xnnpack_cc_library(
618 name = "neonfma_ukernels",
619 hdrs = INTERNAL_HDRS,
620 aarch32_copts = [
621 "-marm",
622 "-mfpu=neon-vfpv4",
623 ],
624 aarch32_srcs = NEONFMA_UKERNELS,
625 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
626 copts = xnnpack_std_copts(),
627 deps = ["@FP16"],
628)
629
630xnnpack_cc_library(
631 name = "neonfp16arith_ukernels",
632 hdrs = INTERNAL_HDRS,
633 aarch64_copts = ["-march=armv8.2-a+fp16"],
634 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
635 copts = xnnpack_std_copts(),
636 deps = ["@FP16"],
637)
638
639xnnpack_cc_library(
640 name = "sse2_ukernels",
641 hdrs = INTERNAL_HDRS,
642 copts = xnnpack_std_copts(),
643 x86_copts = ["-msse2"],
644 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
645 deps = ["@FP16"],
646)
647
648xnnpack_cc_library(
649 name = "avx_ukernels",
650 hdrs = INTERNAL_HDRS,
651 copts = xnnpack_std_copts(),
652 x86_copts = ["-mavx"],
653 x86_srcs = AVX_UKERNELS,
654 deps = ["@FP16"],
655)
656
657xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700658 name = "avx2_ukernels",
659 hdrs = INTERNAL_HDRS,
660 copts = xnnpack_std_copts(),
661 x86_copts = [
662 "-mfma",
663 "-mavx2",
664 ],
665 x86_srcs = AVX2_UKERNELS,
666 deps = ["@FP16"],
667)
668
669xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700670 name = "avx512f_ukernels",
671 hdrs = INTERNAL_HDRS,
672 copts = xnnpack_std_copts(),
673 x86_copts = ["-mavx512f"],
674 x86_srcs = AVX512F_UKERNELS,
675 deps = ["@FP16"],
676)
677
678xnnpack_cc_library(
679 name = "asm_ukernels",
680 hdrs = ["src/xnnpack/assembly.h"],
681 aarch32_srcs = AARCH32_ASM_UKERNELS,
682 aarch64_srcs = AARCH64_ASM_UKERNELS,
683)
684
685xnnpack_aggregate_library(
686 name = "ukernels",
687 aarch32_deps = [
688 ":psimd_ukernels",
689 ":neon_ukernels",
690 ":neonfma_ukernels",
691 ":asm_ukernels",
692 ],
693 aarch64_deps = [
694 ":psimd_ukernels",
695 ":neon_ukernels",
696 ":neonfma_ukernels",
697 ":neonfp16arith_ukernels",
698 ":asm_ukernels",
699 ],
700 generic_deps = [":scalar_ukernels"],
701 wasmsimd_deps = [
702 ":psimd_ukernels",
703 ],
704 x86_deps = [
705 ":psimd_ukernels",
706 ":sse2_ukernels",
707 ":avx_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700708 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700709 ":avx512f_ukernels",
710 ],
711)
712
713xnnpack_cc_library(
714 name = "im2col",
715 srcs = ["src/im2col.c"],
716 hdrs = [
717 "src/xnnpack/common.h",
718 "src/xnnpack/im2col.h",
719 ],
720 copts = xnnpack_std_copts(),
721)
722
723xnnpack_cc_library(
724 name = "indirection",
725 srcs = ["src/indirection.c"],
726 hdrs = INTERNAL_HDRS,
727 copts = xnnpack_std_copts(),
728 deps = [
729 "@FP16",
730 "@FXdiv",
731 "@pthreadpool",
732 ],
733)
734
735xnnpack_cc_library(
736 name = "operator_run",
737 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700738 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
739 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700740 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
741 "-Wno-vla",
742 ],
743 deps = [
744 "@FP16",
745 "@FXdiv",
746 "@clog",
747 "@pthreadpool",
748 ],
749)
750
751cc_library(
752 name = "enable_assembly",
753 defines = select({
754 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
755 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -0700756 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700757 }),
758)
759
Marat Dukhancf056b22019-10-07 10:26:29 -0700760xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700761 name = "operators",
762 srcs = OPERATOR_SRCS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700763 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -0700764 ],
765 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700766 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700767 "-Isrc",
768 "-Iinclude",
769 ] + select({
770 ":debug_build": [],
771 "//conditions:default": xnnpack_min_size_copts(),
772 }),
Marat Dukhancf056b22019-10-07 10:26:29 -0700773 wasm_srcs = ["src/wasm-stubs.c"],
774 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700775 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700776 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700777 "@FP16",
778 "@FXdiv",
779 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700780 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700781 ],
782)
783
784cc_library(
785 name = "XNNPACK",
786 srcs = [
787 "src/init.c",
788 ],
789 copts = xnnpack_std_copts() + LOGGING_COPTS + [
790 "-Isrc",
791 "-Iinclude",
792 ] + select({
793 ":debug_build": [],
794 "//conditions:default": xnnpack_min_size_copts(),
795 }),
796 includes = ["include"],
797 linkstatic = True,
798 textual_hdrs = ["include/xnnpack.h"],
799 visibility = xnnpack_visibility(),
800 deps = [
801 ":enable_assembly",
802 ":ukernels",
803 ":operator_run",
804 ":operators",
805 "@clog",
806 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -0700807 ] + select({
808 ":emscripten": [],
809 "//conditions:default": ["@cpuinfo"],
810 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700811)
812
813cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700814 name = "xnnpack_operators_nhwc_f32",
815 srcs = [
816 "src/init.c",
817 ],
818 copts = xnnpack_std_copts() + LOGGING_COPTS + [
819 "-Isrc",
820 "-Iinclude",
821 ] + select({
822 ":debug_build": [],
823 "//conditions:default": xnnpack_min_size_copts(),
824 }),
825 defines = [
826 "XNN_NO_Q8_OPERATORS",
827 "XNN_NO_U8_OPERATORS",
828 "XNN_NO_X8_OPERATORS",
829 "XNN_NO_SPNCHW_OPERATORS",
830 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700831 includes = ["include"],
832 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700833 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700834 visibility = xnnpack_visibility(),
835 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700836 ":enable_assembly",
837 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700838 ":operator_run",
839 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700840 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700841 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700842 ] + select({
843 ":emscripten": [],
844 "//conditions:default": ["@cpuinfo"],
845 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700846)
847
Marat Dukhancf056b22019-10-07 10:26:29 -0700848xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700849 name = "bench_utils",
850 srcs = ["bench/utils.cc"],
851 hdrs = ["bench/utils.h"],
852 copts = ["-Wno-unused-result"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700853 deps = ["@cpuinfo"],
854)
855
Frank Barchard7e955972019-10-11 10:34:25 -0700856######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -0700857
858xnnpack_benchmark(
859 name = "q8_gemm_bench",
860 srcs = [
861 "bench/gemm.h",
862 "bench/q8-gemm.cc",
863 "src/xnnpack/AlignedAllocator.h",
864 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
865 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
866 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
867)
868
869xnnpack_benchmark(
870 name = "f16_gemm_bench",
871 srcs = [
872 "bench/f16-gemm.cc",
873 "bench/gemm.h",
874 "src/xnnpack/AlignedAllocator.h",
875 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
876 copts = ["-Wno-unused-function"],
877 deps = MICROKERNEL_BENCHMARK_DEPS,
878)
879
880xnnpack_benchmark(
881 name = "f32_igemm_bench",
882 srcs = [
883 "bench/f32-igemm.cc",
884 "bench/conv.h",
885 "src/xnnpack/AlignedAllocator.h",
886 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -0700887 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700888)
889
890xnnpack_benchmark(
891 name = "f32_conv_hwc_bench",
892 srcs = [
893 "bench/f32-conv-hwc.cc",
894 "bench/dconv.h",
895 "src/xnnpack/AlignedAllocator.h",
896 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
897 copts = ["-Wno-unused-function"],
898 deps = MICROKERNEL_BENCHMARK_DEPS,
899)
900
901xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -0700902 name = "f32_conv_hwc2spchw_bench",
903 srcs = [
904 "bench/f32-conv-hwc2spchw.cc",
905 "bench/dconv.h",
906 "src/xnnpack/AlignedAllocator.h",
907 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
908 copts = ["-Wno-unused-function"],
909 deps = MICROKERNEL_BENCHMARK_DEPS,
910)
911
912xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700913 name = "f32_dwconv_bench",
914 srcs = [
915 "bench/f32-dwconv.cc",
916 "bench/dwconv.h",
917 "src/xnnpack/AlignedAllocator.h",
918 ] + MICROKERNEL_BENCHMARK_HDRS,
919 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
920)
921
922xnnpack_benchmark(
923 name = "f32_dwconv_spchw_bench",
924 srcs = [
925 "bench/f32-dwconv-spchw.cc",
926 "bench/dwconv.h",
927 "src/xnnpack/AlignedAllocator.h",
928 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
929 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
930)
931
932xnnpack_benchmark(
933 name = "f32_gemm_bench",
934 srcs = [
935 "bench/f32-gemm.cc",
936 "bench/gemm.h",
937 "src/xnnpack/AlignedAllocator.h",
938 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
939 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -0700940 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700941)
942
943xnnpack_benchmark(
944 name = "f32_rmax_bench",
945 srcs = [
946 "bench/f32-rmax.cc",
947 "src/xnnpack/AlignedAllocator.h",
948 ] + MICROKERNEL_BENCHMARK_HDRS,
949 deps = MICROKERNEL_BENCHMARK_DEPS,
950)
951
952xnnpack_benchmark(
953 name = "f32_spmm_bench",
954 srcs = [
955 "bench/f32-spmm.cc",
956 "bench/gemm.h",
957 "src/xnnpack/AlignedAllocator.h",
958 ] + MICROKERNEL_BENCHMARK_HDRS,
959 copts = ["-Wno-unused-function"],
960 deps = MICROKERNEL_BENCHMARK_DEPS,
961)
962
963xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -0700964 name = "f32_softargmax_bench",
965 srcs = [
966 "bench/f32-softargmax.cc",
967 ] + MICROKERNEL_BENCHMARK_HDRS,
968 copts = ["-Wno-unused-function"],
969 deps = MICROKERNEL_BENCHMARK_DEPS,
970)
971
972xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700973 name = "f32_im2col_gemm_bench",
974 srcs = [
975 "bench/f32-im2col-gemm.cc",
976 "bench/conv.h",
977 "src/xnnpack/AlignedAllocator.h",
978 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
979 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
980)
981
982########################### Benchmarks for operators ###########################
983
984xnnpack_benchmark(
985 name = "add_bench",
986 srcs = ["bench/add.cc"],
987 deps = OPERATOR_BENCHMARK_DEPS,
988)
989
990xnnpack_benchmark(
991 name = "average_pooling_bench",
992 srcs = ["bench/average-pooling.cc"],
993 deps = OPERATOR_BENCHMARK_DEPS,
994)
995
996xnnpack_benchmark(
997 name = "channel_shuffle_bench",
998 srcs = ["bench/channel-shuffle.cc"],
999 deps = OPERATOR_BENCHMARK_DEPS,
1000)
1001
1002xnnpack_benchmark(
1003 name = "convolution_bench",
1004 srcs = ["bench/convolution.cc"],
1005 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1006 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1007)
1008
1009xnnpack_benchmark(
1010 name = "deconvolution_bench",
1011 srcs = ["bench/deconvolution.cc"],
1012 copts = xnnpack_optional_tflite_copts(),
1013 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1014)
1015
1016xnnpack_benchmark(
1017 name = "global_average_pooling_bench",
1018 srcs = ["bench/global-average-pooling.cc"],
1019 deps = OPERATOR_BENCHMARK_DEPS,
1020)
1021
1022xnnpack_benchmark(
1023 name = "max_pooling_bench",
1024 srcs = ["bench/max-pooling.cc"],
1025 deps = OPERATOR_BENCHMARK_DEPS,
1026)
1027
1028xnnpack_benchmark(
1029 name = "sigmoid_bench",
1030 srcs = ["bench/sigmoid.cc"],
1031 deps = OPERATOR_BENCHMARK_DEPS,
1032)
1033
1034xnnpack_benchmark(
1035 name = "softargmax_bench",
1036 srcs = ["bench/softargmax.cc"],
1037 deps = OPERATOR_BENCHMARK_DEPS,
1038)
1039
Marat Dukhanc068bb62019-10-04 13:24:39 -07001040############################# End-to-end benchmarks ############################
1041
1042cc_library(
1043 name = "mobilenet_v1",
1044 srcs = ["models/mobilenet-v1.cc"],
1045 hdrs = ["models/models.h"],
1046 linkstatic = True,
1047 deps = [
1048 ":XNNPACK",
1049 "@pthreadpool",
1050 ],
1051)
1052
1053cc_library(
1054 name = "mobilenet_v2",
1055 srcs = ["models/mobilenet-v2.cc"],
1056 hdrs = ["models/models.h"],
1057 linkstatic = True,
1058 deps = [
1059 ":XNNPACK",
1060 "@pthreadpool",
1061 ],
1062)
1063
1064xnnpack_benchmark(
1065 name = "end2end_bench",
1066 srcs = ["bench/end2end.cc"],
1067 deps = [
1068 ":XNNPACK",
1069 ":mobilenet_v1",
1070 ":mobilenet_v2",
1071 "@pthreadpool",
1072 ],
1073)
1074
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001075#################### Accuracy evaluation for math functions ####################
1076
1077xnnpack_benchmark(
1078 name = "f32_exp_eval",
1079 srcs = [
1080 "eval/f32-exp.cc",
1081 "src/xnnpack/AlignedAllocator.h",
1082 ] + ACCURACY_EVAL_HDRS,
1083 deps = ACCURACY_EVAL_DEPS,
1084)
1085
Marat Dukhan515c9772019-10-17 18:07:57 -07001086xnnpack_benchmark(
1087 name = "f32_expminus_eval",
1088 srcs = [
1089 "eval/f32-expminus.cc",
1090 "src/xnnpack/AlignedAllocator.h",
1091 ] + ACCURACY_EVAL_HDRS,
1092 deps = ACCURACY_EVAL_DEPS,
1093)
1094
Marat Dukhan98ba4412019-10-23 02:14:28 -07001095xnnpack_benchmark(
1096 name = "f32_extexp_eval",
1097 srcs = [
1098 "eval/f32-extexp.cc",
1099 "src/xnnpack/AlignedAllocator.h",
1100 ] + ACCURACY_EVAL_HDRS,
1101 deps = ACCURACY_EVAL_DEPS,
1102)
1103
Marat Dukhan08c4a432019-10-03 09:29:21 -07001104######################### Unit tests for micro-kernels #########################
1105
1106xnnpack_unit_test(
1107 name = "f16_gemm_test",
1108 srcs = [
1109 "test/f16-gemm.cc",
1110 "test/gemm-microkernel-tester.h",
1111 "src/xnnpack/AlignedAllocator.h",
1112 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1113 deps = MICROKERNEL_TEST_DEPS,
1114)
1115
1116xnnpack_unit_test(
1117 name = "f32_argmaxpool_test",
1118 srcs = [
1119 "test/f32-argmaxpool.cc",
1120 "test/argmaxpool-microkernel-tester.h",
1121 "src/xnnpack/AlignedAllocator.h",
1122 ] + MICROKERNEL_TEST_HDRS,
1123 deps = MICROKERNEL_TEST_DEPS,
1124)
1125
1126xnnpack_unit_test(
1127 name = "f32_avgpool_test",
1128 srcs = [
1129 "test/f32-avgpool.cc",
1130 "test/avgpool-microkernel-tester.h",
1131 "src/xnnpack/AlignedAllocator.h",
1132 ] + MICROKERNEL_TEST_HDRS,
1133 deps = MICROKERNEL_TEST_DEPS,
1134)
1135
1136xnnpack_unit_test(
1137 name = "f32_clamp_test",
1138 srcs = [
1139 "test/f32-clamp.cc",
1140 "test/clamp-microkernel-tester.h",
1141 ] + MICROKERNEL_TEST_HDRS,
1142 deps = MICROKERNEL_TEST_DEPS,
1143)
1144
1145xnnpack_unit_test(
1146 name = "f32_igemm_test",
1147 srcs = [
1148 "test/f32-igemm.cc",
1149 "test/gemm-microkernel-tester.h",
1150 "src/xnnpack/AlignedAllocator.h",
1151 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1152 deps = MICROKERNEL_TEST_DEPS,
1153)
1154
1155xnnpack_unit_test(
1156 name = "f32_conv_hwc_test",
1157 srcs = [
1158 "test/f32-conv-hwc.cc",
1159 "test/conv-hwc-microkernel-tester.h",
1160 "src/xnnpack/AlignedAllocator.h",
1161 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1162 deps = MICROKERNEL_TEST_DEPS,
1163)
1164
1165xnnpack_unit_test(
1166 name = "f32_conv_hwc2spchw_test",
1167 srcs = [
1168 "test/f32-conv-hwc2spchw.cc",
1169 "test/conv-hwc2spchw-microkernel-tester.h",
1170 "src/xnnpack/AlignedAllocator.h",
1171 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1172 deps = MICROKERNEL_TEST_DEPS,
1173)
1174
1175xnnpack_unit_test(
1176 name = "f32_dwconv_test",
1177 srcs = [
1178 "test/f32-dwconv.cc",
1179 "test/dwconv-microkernel-tester.h",
1180 "src/xnnpack/AlignedAllocator.h",
1181 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1182 deps = MICROKERNEL_TEST_DEPS,
1183)
1184
1185xnnpack_unit_test(
1186 name = "f32_dwconv_spchw_test",
1187 srcs = [
1188 "test/f32-dwconv-spchw.cc",
1189 "test/dwconv-spchw-microkernel-tester.h",
1190 "src/xnnpack/AlignedAllocator.h",
1191 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1192 deps = MICROKERNEL_TEST_DEPS,
1193)
1194
1195xnnpack_unit_test(
1196 name = "f32_gavgpool_test",
1197 srcs = [
1198 "test/f32-gavgpool.cc",
1199 "test/gavgpool-microkernel-tester.h",
1200 "src/xnnpack/AlignedAllocator.h",
1201 ] + MICROKERNEL_TEST_HDRS,
1202 deps = MICROKERNEL_TEST_DEPS,
1203)
1204
1205xnnpack_unit_test(
1206 name = "f32_gavgpool_spchw_test",
1207 srcs = [
1208 "test/f32-gavgpool-spchw.cc",
1209 "test/gavgpool-spchw-microkernel-tester.h",
1210 "src/xnnpack/AlignedAllocator.h",
1211 ] + MICROKERNEL_TEST_HDRS,
1212 deps = MICROKERNEL_TEST_DEPS,
1213)
1214
1215xnnpack_unit_test(
1216 name = "f32_gemm_test",
1217 srcs = [
1218 "test/f32-gemm.cc",
1219 "test/gemm-microkernel-tester.h",
1220 "src/xnnpack/AlignedAllocator.h",
1221 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1222 deps = MICROKERNEL_TEST_DEPS,
1223)
1224
1225xnnpack_unit_test(
1226 name = "f32_gemminc_test",
1227 srcs = [
1228 "test/f32-gemminc.cc",
1229 "test/gemm-microkernel-tester.h",
1230 "src/xnnpack/AlignedAllocator.h",
1231 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1232 deps = MICROKERNEL_TEST_DEPS,
1233)
1234
1235xnnpack_unit_test(
1236 name = "f32_hswish_test",
1237 srcs = [
1238 "test/f32-hswish.cc",
1239 "test/hswish-microkernel-tester.h",
1240 ] + MICROKERNEL_TEST_HDRS,
1241 deps = MICROKERNEL_TEST_DEPS,
1242)
1243
1244xnnpack_unit_test(
1245 name = "f32_maxpool_test",
1246 srcs = [
1247 "test/f32-maxpool.cc",
1248 "test/maxpool-microkernel-tester.h",
1249 ] + MICROKERNEL_TEST_HDRS,
1250 deps = MICROKERNEL_TEST_DEPS,
1251)
1252
1253xnnpack_unit_test(
1254 name = "f32_pavgpool_test",
1255 srcs = [
1256 "test/f32-pavgpool.cc",
1257 "test/avgpool-microkernel-tester.h",
1258 "src/xnnpack/AlignedAllocator.h",
1259 ] + MICROKERNEL_TEST_HDRS,
1260 deps = MICROKERNEL_TEST_DEPS,
1261)
1262
1263xnnpack_unit_test(
1264 name = "f32_ppmm_test",
1265 srcs = [
1266 "test/f32-ppmm.cc",
1267 "test/gemm-microkernel-tester.h",
1268 "src/xnnpack/AlignedAllocator.h",
1269 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1270 deps = MICROKERNEL_TEST_DEPS,
1271)
1272
1273xnnpack_unit_test(
1274 name = "f32_prelu_test",
1275 srcs = [
1276 "test/f32-prelu.cc",
1277 "test/prelu-microkernel-tester.h",
1278 "src/xnnpack/AlignedAllocator.h",
1279 ] + MICROKERNEL_TEST_HDRS,
1280 deps = MICROKERNEL_TEST_DEPS,
1281)
1282
1283xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001284 name = "f32_raddexpminusmax_test",
1285 srcs = [
1286 "test/f32-raddexpminusmax.cc",
1287 "test/raddexpminusmax-microkernel-tester.h",
1288 ] + MICROKERNEL_TEST_HDRS,
1289 deps = MICROKERNEL_TEST_DEPS,
1290)
1291
1292xnnpack_unit_test(
1293 name = "f32_raddstoreexpminusmax_test",
1294 srcs = [
1295 "test/f32-raddstoreexpminusmax.cc",
1296 "test/raddstoreexpminusmax-microkernel-tester.h",
1297 ] + MICROKERNEL_TEST_HDRS,
1298 deps = MICROKERNEL_TEST_DEPS,
1299)
1300
1301xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001302 name = "f32_rmax_test",
1303 srcs = [
1304 "test/f32-rmax.cc",
1305 "test/rmax-microkernel-tester.h",
1306 ] + MICROKERNEL_TEST_HDRS,
1307 deps = MICROKERNEL_TEST_DEPS,
1308)
1309
1310xnnpack_unit_test(
1311 name = "f32_spmm_test",
1312 srcs = [
1313 "test/f32-spmm.cc",
1314 "test/spmm-microkernel-tester.h",
1315 "src/xnnpack/AlignedAllocator.h",
1316 ] + MICROKERNEL_TEST_HDRS,
1317 deps = MICROKERNEL_TEST_DEPS,
1318)
1319
1320xnnpack_unit_test(
1321 name = "f32_vadd_test",
1322 srcs = [
1323 "test/f32-vadd.cc",
1324 "test/vadd-microkernel-tester.h",
1325 ] + MICROKERNEL_TEST_HDRS,
1326 deps = MICROKERNEL_TEST_DEPS,
1327)
1328
1329xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001330 name = "f32_vmul_test",
1331 srcs = [
1332 "test/f32-vmul.cc",
1333 "test/vmul-microkernel-tester.h",
1334 ] + MICROKERNEL_TEST_HDRS,
1335 deps = MICROKERNEL_TEST_DEPS,
1336)
1337
1338xnnpack_unit_test(
1339 name = "f32_vmulcaddc_test",
1340 srcs = [
1341 "test/f32-vmulcaddc.cc",
1342 "test/vmulcaddc-microkernel-tester.h",
1343 "src/xnnpack/AlignedAllocator.h",
1344 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1345 deps = MICROKERNEL_TEST_DEPS,
1346)
1347
1348xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001349 name = "f32_vscale_test",
1350 srcs = [
1351 "test/f32-vscale.cc",
1352 "test/vscale-microkernel-tester.h",
1353 ] + MICROKERNEL_TEST_HDRS,
1354 deps = MICROKERNEL_TEST_DEPS,
1355)
1356
1357xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001358 name = "f32_vscaleexpminusmax_test",
1359 srcs = [
1360 "test/f32-vscaleexpminusmax.cc",
1361 "test/vscaleexpminusmax-microkernel-tester.h",
1362 ] + MICROKERNEL_TEST_HDRS,
1363 deps = MICROKERNEL_TEST_DEPS,
1364)
1365
1366xnnpack_unit_test(
1367 name = "f32_vsub_test",
1368 srcs = [
1369 "test/f32-vsub.cc",
1370 "test/vsub-microkernel-tester.h",
1371 ] + MICROKERNEL_TEST_HDRS,
1372 deps = MICROKERNEL_TEST_DEPS,
1373)
1374
1375xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001376 name = "q8_avgpool_test",
1377 srcs = [
1378 "test/q8-avgpool.cc",
1379 "test/avgpool-microkernel-tester.h",
1380 "src/xnnpack/AlignedAllocator.h",
1381 ] + MICROKERNEL_TEST_HDRS,
1382 deps = MICROKERNEL_TEST_DEPS,
1383)
1384
1385xnnpack_unit_test(
1386 name = "q8_igemm_test",
1387 srcs = [
1388 "test/q8-igemm.cc",
1389 "test/gemm-microkernel-tester.h",
1390 "src/xnnpack/AlignedAllocator.h",
1391 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1392 deps = MICROKERNEL_TEST_DEPS,
1393)
1394
1395xnnpack_unit_test(
1396 name = "q8_dwconv_test",
1397 srcs = [
1398 "test/q8-dwconv.cc",
1399 "test/dwconv-microkernel-tester.h",
1400 "src/xnnpack/AlignedAllocator.h",
1401 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1402 deps = MICROKERNEL_TEST_DEPS,
1403)
1404
1405xnnpack_unit_test(
1406 name = "q8_gavgpool_test",
1407 srcs = [
1408 "test/q8-gavgpool.cc",
1409 "test/gavgpool-microkernel-tester.h",
1410 "src/xnnpack/AlignedAllocator.h",
1411 ] + MICROKERNEL_TEST_HDRS,
1412 deps = MICROKERNEL_TEST_DEPS,
1413)
1414
1415xnnpack_unit_test(
1416 name = "q8_gemm_test",
1417 srcs = [
1418 "test/q8-gemm.cc",
1419 "test/gemm-microkernel-tester.h",
1420 "src/xnnpack/AlignedAllocator.h",
1421 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1422 deps = MICROKERNEL_TEST_DEPS,
1423)
1424
1425xnnpack_unit_test(
1426 name = "q8_vadd_test",
1427 srcs = [
1428 "test/q8-vadd.cc",
1429 "test/vadd-microkernel-tester.h",
1430 ] + MICROKERNEL_TEST_HDRS,
1431 deps = MICROKERNEL_TEST_DEPS,
1432)
1433
1434xnnpack_unit_test(
1435 name = "u8_clamp_test",
1436 srcs = [
1437 "test/u8-clamp.cc",
1438 "test/clamp-microkernel-tester.h",
1439 ] + MICROKERNEL_TEST_HDRS,
1440 deps = MICROKERNEL_TEST_DEPS,
1441)
1442
1443xnnpack_unit_test(
1444 name = "u8_lut32norm_test",
1445 srcs = [
1446 "test/u8-lut32norm.cc",
1447 "test/lut-norm-microkernel-tester.h",
1448 ] + MICROKERNEL_TEST_HDRS,
1449 deps = MICROKERNEL_TEST_DEPS,
1450)
1451
1452xnnpack_unit_test(
1453 name = "u8_maxpool_test",
1454 srcs = [
1455 "test/u8-maxpool.cc",
1456 "test/maxpool-microkernel-tester.h",
1457 ] + MICROKERNEL_TEST_HDRS,
1458 deps = MICROKERNEL_TEST_DEPS,
1459)
1460
1461xnnpack_unit_test(
1462 name = "u8_rmax_test",
1463 srcs = [
1464 "test/u8-rmax.cc",
1465 "test/rmax-microkernel-tester.h",
1466 ] + MICROKERNEL_TEST_HDRS,
1467 deps = MICROKERNEL_TEST_DEPS,
1468)
1469
1470xnnpack_unit_test(
1471 name = "x32_packx_test",
1472 srcs = [
1473 "test/x32-packx.cc",
1474 "test/pack-microkernel-tester.h",
1475 "src/xnnpack/AlignedAllocator.h",
1476 ] + MICROKERNEL_TEST_HDRS,
1477 deps = MICROKERNEL_TEST_DEPS,
1478)
1479
1480xnnpack_unit_test(
1481 name = "x32_pad_test",
1482 srcs = [
1483 "test/x32-pad.cc",
1484 "test/pad-microkernel-tester.h",
1485 ] + MICROKERNEL_TEST_HDRS,
1486 deps = MICROKERNEL_TEST_DEPS,
1487)
1488
1489xnnpack_unit_test(
1490 name = "x32_unpool_test",
1491 srcs = [
1492 "test/x32-unpool.cc",
1493 "test/unpool-microkernel-tester.h",
1494 ] + MICROKERNEL_TEST_HDRS,
1495 deps = MICROKERNEL_TEST_DEPS,
1496)
1497
1498xnnpack_unit_test(
1499 name = "x32_zip_test",
1500 srcs = [
1501 "test/x32-zip.cc",
1502 "test/zip-microkernel-tester.h",
1503 ] + MICROKERNEL_TEST_HDRS,
1504 deps = MICROKERNEL_TEST_DEPS,
1505)
1506
1507xnnpack_unit_test(
1508 name = "x8_lut_test",
1509 srcs = [
1510 "test/x8-lut.cc",
1511 "test/lut-microkernel-tester.h",
1512 ] + MICROKERNEL_TEST_HDRS,
1513 deps = MICROKERNEL_TEST_DEPS,
1514)
1515
1516xnnpack_unit_test(
1517 name = "x8_zip_test",
1518 srcs = [
1519 "test/x8-zip.cc",
1520 "test/zip-microkernel-tester.h",
1521 ] + MICROKERNEL_TEST_HDRS,
1522 deps = MICROKERNEL_TEST_DEPS,
1523)
1524
1525########################### Size test for the library ##########################
1526
1527xnnpack_binary(
1528 name = "size_test",
1529 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001530 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001531)
1532
1533########################### Unit tests for operators ###########################
1534
1535xnnpack_unit_test(
1536 name = "add_test",
1537 srcs = [
1538 "test/add.cc",
1539 "test/add-operator-tester.h",
1540 ],
1541 deps = OPERATOR_TEST_DEPS,
1542)
1543
1544xnnpack_unit_test(
1545 name = "argmax_pooling_test",
1546 srcs = [
1547 "test/argmax-pooling.cc",
1548 "test/argmax-pooling-operator-tester.h",
1549 ] + OPERATOR_TEST_PARAMS_HDRS,
1550 deps = OPERATOR_TEST_DEPS,
1551)
1552
1553xnnpack_unit_test(
1554 name = "average_pooling_test",
1555 srcs = [
1556 "test/average-pooling.cc",
1557 "test/average-pooling-operator-tester.h",
1558 ] + OPERATOR_TEST_PARAMS_HDRS,
1559 deps = OPERATOR_TEST_DEPS,
1560)
1561
1562xnnpack_unit_test(
1563 name = "channel_pad_test",
1564 srcs = [
1565 "test/channel-pad.cc",
1566 "test/channel-pad-operator-tester.h",
1567 ] + OPERATOR_TEST_PARAMS_HDRS,
1568 deps = OPERATOR_TEST_DEPS,
1569)
1570
1571xnnpack_unit_test(
1572 name = "channel_shuffle_test",
1573 srcs = [
1574 "test/channel-shuffle.cc",
1575 "test/channel-shuffle-operator-tester.h",
1576 ],
1577 deps = OPERATOR_TEST_DEPS,
1578)
1579
1580xnnpack_unit_test(
1581 name = "clamp_test",
1582 srcs = [
1583 "test/clamp.cc",
1584 "test/clamp-operator-tester.h",
1585 ],
1586 deps = OPERATOR_TEST_DEPS,
1587)
1588
1589xnnpack_unit_test(
1590 name = "convolution_test",
1591 srcs = [
1592 "test/convolution.cc",
1593 "test/convolution-operator-tester.h",
1594 ],
1595 deps = OPERATOR_TEST_DEPS,
1596)
1597
1598xnnpack_unit_test(
1599 name = "convolution_spnchw_test",
1600 srcs = [
1601 "test/convolution-spnchw.cc",
1602 "test/convolution-spnchw-operator-tester.h",
1603 ],
1604 deps = OPERATOR_TEST_DEPS,
1605)
1606
1607xnnpack_unit_test(
1608 name = "deconvolution_test",
1609 srcs = [
1610 "test/deconvolution.cc",
1611 "test/deconvolution-operator-tester.h",
1612 ] + OPERATOR_TEST_PARAMS_HDRS,
1613 deps = OPERATOR_TEST_DEPS,
1614)
1615
1616xnnpack_unit_test(
1617 name = "fully_connected_test",
1618 srcs = [
1619 "test/fully-connected.cc",
1620 "test/fully-connected-operator-tester.h",
1621 ],
1622 deps = OPERATOR_TEST_DEPS,
1623)
1624
1625xnnpack_unit_test(
1626 name = "global_average_pooling_test",
1627 srcs = [
1628 "test/global-average-pooling.cc",
1629 "test/global-average-pooling-operator-tester.h",
1630 ] + OPERATOR_TEST_PARAMS_HDRS,
1631 deps = OPERATOR_TEST_DEPS,
1632)
1633
1634xnnpack_unit_test(
1635 name = "global_average_pooling_spnchw_test",
1636 srcs = [
1637 "test/global-average-pooling-spnchw.cc",
1638 "test/global-average-pooling-spnchw-operator-tester.h",
1639 ],
1640 deps = OPERATOR_TEST_DEPS,
1641)
1642
1643xnnpack_unit_test(
1644 name = "hardswish_test",
1645 srcs = [
1646 "test/hardswish.cc",
1647 "test/hardswish-operator-tester.h",
1648 ],
1649 deps = OPERATOR_TEST_DEPS,
1650)
1651
1652xnnpack_unit_test(
1653 name = "leaky_relu_test",
1654 srcs = [
1655 "test/leaky-relu.cc",
1656 "test/leaky-relu-operator-tester.h",
1657 ],
1658 deps = OPERATOR_TEST_DEPS,
1659)
1660
1661xnnpack_unit_test(
1662 name = "max_pooling_test",
1663 srcs = [
1664 "test/max-pooling.cc",
1665 "test/max-pooling-operator-tester.h",
1666 ] + OPERATOR_TEST_PARAMS_HDRS,
1667 deps = OPERATOR_TEST_DEPS,
1668)
1669
1670xnnpack_unit_test(
1671 name = "prelu_test",
1672 srcs = [
1673 "test/prelu.cc",
1674 "test/prelu-operator-tester.h",
1675 ] + OPERATOR_TEST_PARAMS_HDRS,
1676 deps = OPERATOR_TEST_DEPS,
1677)
1678
1679xnnpack_unit_test(
1680 name = "sigmoid_test",
1681 srcs = [
1682 "test/sigmoid.cc",
1683 "test/sigmoid-operator-tester.h",
1684 ],
1685 deps = OPERATOR_TEST_DEPS,
1686)
1687
1688xnnpack_unit_test(
1689 name = "softargmax_test",
1690 srcs = [
1691 "test/softargmax.cc",
1692 "test/softargmax-operator-tester.h",
1693 ],
1694 deps = OPERATOR_TEST_DEPS,
1695)
1696
1697xnnpack_unit_test(
1698 name = "unpooling_test",
1699 srcs = [
1700 "test/unpooling.cc",
1701 "test/unpooling-operator-tester.h",
1702 ],
1703 deps = OPERATOR_TEST_DEPS,
1704)
1705
1706############################# Build configurations #############################
1707
1708config_setting(
1709 name = "linux_k8",
1710 values = {
1711 "cpu": "k8",
1712 },
1713)
1714
1715config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07001716 name = "linux_aarch64",
1717 values = {
1718 "cpu": "aarch64",
1719 },
1720)
1721
1722config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001723 name = "android",
1724 values = {"crosstool_top": "//external:android/crosstool"},
1725)
1726
1727config_setting(
1728 name = "android_armv7",
1729 values = {
1730 "crosstool_top": "//external:android/crosstool",
1731 "cpu": "armeabi-v7a",
1732 },
1733)
1734
1735config_setting(
1736 name = "android_arm64",
1737 values = {
1738 "crosstool_top": "//external:android/crosstool",
1739 "cpu": "arm64-v8a",
1740 },
1741)
1742
1743config_setting(
1744 name = "android_x86",
1745 values = {
1746 "crosstool_top": "//external:android/crosstool",
1747 "cpu": "x86",
1748 },
1749)
1750
1751config_setting(
1752 name = "android_x86_64",
1753 values = {
1754 "crosstool_top": "//external:android/crosstool",
1755 "cpu": "x86_64",
1756 },
1757)
1758
1759config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07001760 name = "macos_x86_64",
1761 values = {
1762 "apple_platform_type": "macos",
1763 "cpu": "darwin",
1764 },
1765)
1766
1767config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001768 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001769 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07001770)
1771
1772config_setting(
1773 name = "emscripten_wasm",
1774 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001775 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001776 "cpu": "wasm",
1777 },
1778)
1779
1780config_setting(
1781 name = "emscripten_wasmsimd",
1782 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001783 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001784 "cpu": "wasm",
1785 "features": "wasmsimd",
1786 },
1787)
1788
1789config_setting(
1790 name = "emscripten_asmjs",
1791 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001792 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001793 "cpu": "asmjs",
1794 },
1795)
1796
Marat Dukhan174706e2019-10-09 16:33:53 -07001797# Builds with -c dbg
Marat Dukhan08c4a432019-10-03 09:29:21 -07001798config_setting(
1799 name = "debug_build",
1800 values = {
1801 "compilation_mode": "dbg",
1802 },
Marat Dukhan08c4a432019-10-03 09:29:21 -07001803)
1804
Marat Dukhan174706e2019-10-09 16:33:53 -07001805# Builds with -c opt
Marat Dukhan08c4a432019-10-03 09:29:21 -07001806config_setting(
1807 name = "optimized_build",
1808 values = {
1809 "compilation_mode": "opt",
1810 },
Marat Dukhan08c4a432019-10-03 09:29:21 -07001811)
1812
1813# Enables usage of assembly kernels.
1814config_setting(
1815 name = "xnn_enable_assembly_explicit_true",
1816 define_values = {"xnn_enable_assembly": "true"},
1817)
1818
1819# Disables usage of assembly kernels.
1820config_setting(
1821 name = "xnn_enable_assembly_explicit_false",
1822 define_values = {"xnn_enable_assembly": "false"},
1823)