blob: 590f0dac08b24809852df4bc366cb8aa52eadbe4 [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",
451 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a57.S",
452 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S",
453 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S",
454 "src/f32-gemm/6x8-aarch64-neonfma-ld128.S",
455 "src/f32-gemm/6x8-aarch64-neonfma-ld64.S",
456 "src/f32-gemminc/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700457 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700458 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a57.S",
459 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a75.S",
460 "src/f32-gemminc/4x12-aarch64-neonfma-cortex-a53.S",
461 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a57.S",
462 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a75.S",
463 "src/f32-gemminc/4x8-aarch64-neonfma-ld128.S",
464 "src/f32-gemminc/4x8-aarch64-neonfma-ld64.S",
465 "src/f32-gemminc/5x8-aarch64-neonfma-cortex-a75.S",
466 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a57.S",
467 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a73.S",
468 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a75.S",
469 "src/f32-gemminc/6x8-aarch64-neonfma-ld128.S",
470 "src/f32-gemminc/6x8-aarch64-neonfma-ld64.S",
471 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700472 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700473 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S",
474 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S",
475 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
476 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
477 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
478 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
479 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
480 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
481]
482
483INTERNAL_MICROKERNEL_HDRS = [
484 "src/xnnpack/argmaxpool.h",
485 "src/xnnpack/avgpool.h",
486 "src/xnnpack/clamp.h",
487 "src/xnnpack/common.h",
488 "src/xnnpack/conv.h",
489 "src/xnnpack/dwconv.h",
490 "src/xnnpack/gavgpool.h",
491 "src/xnnpack/gemm.h",
492 "src/xnnpack/hswish.h",
493 "src/xnnpack/igemm.h",
494 "src/xnnpack/lut.h",
495 "src/xnnpack/math.h",
496 "src/xnnpack/maxpool.h",
497 "src/xnnpack/packx.h",
498 "src/xnnpack/pad.h",
499 "src/xnnpack/params.h",
500 "src/xnnpack/pavgpool.h",
501 "src/xnnpack/ppmm.h",
502 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700503 "src/xnnpack/raddexpminusmax.h",
504 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700505 "src/xnnpack/rmax.h",
506 "src/xnnpack/scalar-utils.h",
507 "src/xnnpack/spmm.h",
508 "src/xnnpack/unpool.h",
509 "src/xnnpack/vadd.h",
510 "src/xnnpack/vmul.h",
511 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700512 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700513 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700514 "src/xnnpack/vsub.h",
515 "src/xnnpack/zip.h",
516]
517
518INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
519 "include/xnnpack.h",
520 "src/xnnpack/allocator.h",
521 "src/xnnpack/compute.h",
522 "src/xnnpack/im2col.h",
523 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700524 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700525 "src/xnnpack/operator.h",
526 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700527 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700528 "src/xnnpack/requantization.h",
529]
530
531ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
532 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700533]
534
535MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
536 "src/xnnpack/requantization.h",
537 "include/xnnpack.h",
538]
539
540MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
541 "src/xnnpack/isa-checks.h",
542 "src/xnnpack/requantization.h",
543 "include/xnnpack.h",
544]
545
546OPERATOR_TEST_PARAMS_HDRS = [
547 "src/xnnpack/params.h",
548 "src/xnnpack/common.h",
549]
550
551WEIGHTS_PACK_HDRS = [
552 "src/xnnpack/pack.h",
553 "src/xnnpack/operator.h",
554 "src/xnnpack/compute.h",
555]
556
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700557LOGGING_COPTS = select({
558 # No logging in optimized mode
559 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
560 # Full logging in debug mode
561 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
562 # Error-only logging in default (fastbuild) mode
563 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
564})
565
566LOGGING_HDRS = [
567 "src/xnnpack/log.h",
568]
569
Marat Dukhan08c4a432019-10-03 09:29:21 -0700570xnnpack_cc_library(
571 name = "scalar_ukernels",
572 srcs = SCALAR_UKERNELS,
573 hdrs = INTERNAL_HDRS,
574 aarch32_copts = ["-marm"],
575 copts = xnnpack_std_copts(),
576 deps = [
577 "@FP16",
578 "@FXdiv",
579 ],
580)
581
582xnnpack_cc_library(
583 name = "psimd_ukernels",
584 srcs = PSIMD_UKERNELS,
585 hdrs = INTERNAL_HDRS,
586 aarch32_copts = [
587 "-marm",
588 "-mfpu=neon",
589 ],
590 copts = xnnpack_std_copts(),
591 optimized_copts = [
592 "-O3",
593 "-ffast-math",
594 ],
595 deps = [
596 "@FP16",
597 "@psimd",
598 ],
599)
600
601xnnpack_cc_library(
602 name = "neon_ukernels",
603 hdrs = INTERNAL_HDRS,
604 aarch32_copts = [
605 "-marm",
606 "-mfpu=neon",
607 ],
608 aarch32_srcs = NEON_UKERNELS,
609 aarch64_srcs = NEON_UKERNELS,
610 copts = xnnpack_std_copts(),
611 deps = ["@FP16"],
612)
613
614xnnpack_cc_library(
615 name = "neonfma_ukernels",
616 hdrs = INTERNAL_HDRS,
617 aarch32_copts = [
618 "-marm",
619 "-mfpu=neon-vfpv4",
620 ],
621 aarch32_srcs = NEONFMA_UKERNELS,
622 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
623 copts = xnnpack_std_copts(),
624 deps = ["@FP16"],
625)
626
627xnnpack_cc_library(
628 name = "neonfp16arith_ukernels",
629 hdrs = INTERNAL_HDRS,
630 aarch64_copts = ["-march=armv8.2-a+fp16"],
631 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
632 copts = xnnpack_std_copts(),
633 deps = ["@FP16"],
634)
635
636xnnpack_cc_library(
637 name = "sse2_ukernels",
638 hdrs = INTERNAL_HDRS,
639 copts = xnnpack_std_copts(),
640 x86_copts = ["-msse2"],
641 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
642 deps = ["@FP16"],
643)
644
645xnnpack_cc_library(
646 name = "avx_ukernels",
647 hdrs = INTERNAL_HDRS,
648 copts = xnnpack_std_copts(),
649 x86_copts = ["-mavx"],
650 x86_srcs = AVX_UKERNELS,
651 deps = ["@FP16"],
652)
653
654xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700655 name = "avx2_ukernels",
656 hdrs = INTERNAL_HDRS,
657 copts = xnnpack_std_copts(),
658 x86_copts = [
659 "-mfma",
660 "-mavx2",
661 ],
662 x86_srcs = AVX2_UKERNELS,
663 deps = ["@FP16"],
664)
665
666xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700667 name = "avx512f_ukernels",
668 hdrs = INTERNAL_HDRS,
669 copts = xnnpack_std_copts(),
670 x86_copts = ["-mavx512f"],
671 x86_srcs = AVX512F_UKERNELS,
672 deps = ["@FP16"],
673)
674
675xnnpack_cc_library(
676 name = "asm_ukernels",
677 hdrs = ["src/xnnpack/assembly.h"],
678 aarch32_srcs = AARCH32_ASM_UKERNELS,
679 aarch64_srcs = AARCH64_ASM_UKERNELS,
680)
681
682xnnpack_aggregate_library(
683 name = "ukernels",
684 aarch32_deps = [
685 ":psimd_ukernels",
686 ":neon_ukernels",
687 ":neonfma_ukernels",
688 ":asm_ukernels",
689 ],
690 aarch64_deps = [
691 ":psimd_ukernels",
692 ":neon_ukernels",
693 ":neonfma_ukernels",
694 ":neonfp16arith_ukernels",
695 ":asm_ukernels",
696 ],
697 generic_deps = [":scalar_ukernels"],
698 wasmsimd_deps = [
699 ":psimd_ukernels",
700 ],
701 x86_deps = [
702 ":psimd_ukernels",
703 ":sse2_ukernels",
704 ":avx_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700705 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700706 ":avx512f_ukernels",
707 ],
708)
709
710xnnpack_cc_library(
711 name = "im2col",
712 srcs = ["src/im2col.c"],
713 hdrs = [
714 "src/xnnpack/common.h",
715 "src/xnnpack/im2col.h",
716 ],
717 copts = xnnpack_std_copts(),
718)
719
720xnnpack_cc_library(
721 name = "indirection",
722 srcs = ["src/indirection.c"],
723 hdrs = INTERNAL_HDRS,
724 copts = xnnpack_std_copts(),
725 deps = [
726 "@FP16",
727 "@FXdiv",
728 "@pthreadpool",
729 ],
730)
731
732xnnpack_cc_library(
733 name = "operator_run",
734 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700735 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
736 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700737 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
738 "-Wno-vla",
739 ],
740 deps = [
741 "@FP16",
742 "@FXdiv",
743 "@clog",
744 "@pthreadpool",
745 ],
746)
747
748cc_library(
749 name = "enable_assembly",
750 defines = select({
751 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
752 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -0700753 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700754 }),
755)
756
Marat Dukhancf056b22019-10-07 10:26:29 -0700757xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700758 name = "operators",
759 srcs = OPERATOR_SRCS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700760 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -0700761 ],
762 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700763 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700764 "-Isrc",
765 "-Iinclude",
766 ] + select({
767 ":debug_build": [],
768 "//conditions:default": xnnpack_min_size_copts(),
769 }),
Marat Dukhancf056b22019-10-07 10:26:29 -0700770 wasm_srcs = ["src/wasm-stubs.c"],
771 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700772 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700773 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700774 "@FP16",
775 "@FXdiv",
776 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700777 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700778 ],
779)
780
781cc_library(
782 name = "XNNPACK",
783 srcs = [
784 "src/init.c",
785 ],
786 copts = xnnpack_std_copts() + LOGGING_COPTS + [
787 "-Isrc",
788 "-Iinclude",
789 ] + select({
790 ":debug_build": [],
791 "//conditions:default": xnnpack_min_size_copts(),
792 }),
793 includes = ["include"],
794 linkstatic = True,
795 textual_hdrs = ["include/xnnpack.h"],
796 visibility = xnnpack_visibility(),
797 deps = [
798 ":enable_assembly",
799 ":ukernels",
800 ":operator_run",
801 ":operators",
802 "@clog",
803 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -0700804 ] + select({
805 ":emscripten": [],
806 "//conditions:default": ["@cpuinfo"],
807 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700808)
809
810cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700811 name = "xnnpack_operators_nhwc_f32",
812 srcs = [
813 "src/init.c",
814 ],
815 copts = xnnpack_std_copts() + LOGGING_COPTS + [
816 "-Isrc",
817 "-Iinclude",
818 ] + select({
819 ":debug_build": [],
820 "//conditions:default": xnnpack_min_size_copts(),
821 }),
822 defines = [
823 "XNN_NO_Q8_OPERATORS",
824 "XNN_NO_U8_OPERATORS",
825 "XNN_NO_X8_OPERATORS",
826 "XNN_NO_SPNCHW_OPERATORS",
827 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700828 includes = ["include"],
829 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700830 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700831 visibility = xnnpack_visibility(),
832 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700833 ":enable_assembly",
834 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700835 ":operator_run",
836 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700837 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700838 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700839 ] + select({
840 ":emscripten": [],
841 "//conditions:default": ["@cpuinfo"],
842 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700843)
844
Marat Dukhancf056b22019-10-07 10:26:29 -0700845xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700846 name = "bench_utils",
847 srcs = ["bench/utils.cc"],
848 hdrs = ["bench/utils.h"],
849 copts = ["-Wno-unused-result"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700850 deps = ["@cpuinfo"],
851)
852
Frank Barchard7e955972019-10-11 10:34:25 -0700853######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -0700854
855xnnpack_benchmark(
856 name = "q8_gemm_bench",
857 srcs = [
858 "bench/gemm.h",
859 "bench/q8-gemm.cc",
860 "src/xnnpack/AlignedAllocator.h",
861 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
862 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
863 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
864)
865
866xnnpack_benchmark(
867 name = "f16_gemm_bench",
868 srcs = [
869 "bench/f16-gemm.cc",
870 "bench/gemm.h",
871 "src/xnnpack/AlignedAllocator.h",
872 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
873 copts = ["-Wno-unused-function"],
874 deps = MICROKERNEL_BENCHMARK_DEPS,
875)
876
877xnnpack_benchmark(
878 name = "f32_igemm_bench",
879 srcs = [
880 "bench/f32-igemm.cc",
881 "bench/conv.h",
882 "src/xnnpack/AlignedAllocator.h",
883 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -0700884 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700885)
886
887xnnpack_benchmark(
888 name = "f32_conv_hwc_bench",
889 srcs = [
890 "bench/f32-conv-hwc.cc",
891 "bench/dconv.h",
892 "src/xnnpack/AlignedAllocator.h",
893 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
894 copts = ["-Wno-unused-function"],
895 deps = MICROKERNEL_BENCHMARK_DEPS,
896)
897
898xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -0700899 name = "f32_conv_hwc2spchw_bench",
900 srcs = [
901 "bench/f32-conv-hwc2spchw.cc",
902 "bench/dconv.h",
903 "src/xnnpack/AlignedAllocator.h",
904 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
905 copts = ["-Wno-unused-function"],
906 deps = MICROKERNEL_BENCHMARK_DEPS,
907)
908
909xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700910 name = "f32_dwconv_bench",
911 srcs = [
912 "bench/f32-dwconv.cc",
913 "bench/dwconv.h",
914 "src/xnnpack/AlignedAllocator.h",
915 ] + MICROKERNEL_BENCHMARK_HDRS,
916 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
917)
918
919xnnpack_benchmark(
920 name = "f32_dwconv_spchw_bench",
921 srcs = [
922 "bench/f32-dwconv-spchw.cc",
923 "bench/dwconv.h",
924 "src/xnnpack/AlignedAllocator.h",
925 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
926 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
927)
928
929xnnpack_benchmark(
930 name = "f32_gemm_bench",
931 srcs = [
932 "bench/f32-gemm.cc",
933 "bench/gemm.h",
934 "src/xnnpack/AlignedAllocator.h",
935 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
936 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -0700937 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700938)
939
940xnnpack_benchmark(
941 name = "f32_rmax_bench",
942 srcs = [
943 "bench/f32-rmax.cc",
944 "src/xnnpack/AlignedAllocator.h",
945 ] + MICROKERNEL_BENCHMARK_HDRS,
946 deps = MICROKERNEL_BENCHMARK_DEPS,
947)
948
949xnnpack_benchmark(
950 name = "f32_spmm_bench",
951 srcs = [
952 "bench/f32-spmm.cc",
953 "bench/gemm.h",
954 "src/xnnpack/AlignedAllocator.h",
955 ] + MICROKERNEL_BENCHMARK_HDRS,
956 copts = ["-Wno-unused-function"],
957 deps = MICROKERNEL_BENCHMARK_DEPS,
958)
959
960xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -0700961 name = "f32_softargmax_bench",
962 srcs = [
963 "bench/f32-softargmax.cc",
964 ] + MICROKERNEL_BENCHMARK_HDRS,
965 copts = ["-Wno-unused-function"],
966 deps = MICROKERNEL_BENCHMARK_DEPS,
967)
968
969xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700970 name = "f32_im2col_gemm_bench",
971 srcs = [
972 "bench/f32-im2col-gemm.cc",
973 "bench/conv.h",
974 "src/xnnpack/AlignedAllocator.h",
975 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
976 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
977)
978
979########################### Benchmarks for operators ###########################
980
981xnnpack_benchmark(
982 name = "add_bench",
983 srcs = ["bench/add.cc"],
984 deps = OPERATOR_BENCHMARK_DEPS,
985)
986
987xnnpack_benchmark(
988 name = "average_pooling_bench",
989 srcs = ["bench/average-pooling.cc"],
990 deps = OPERATOR_BENCHMARK_DEPS,
991)
992
993xnnpack_benchmark(
994 name = "channel_shuffle_bench",
995 srcs = ["bench/channel-shuffle.cc"],
996 deps = OPERATOR_BENCHMARK_DEPS,
997)
998
999xnnpack_benchmark(
1000 name = "convolution_bench",
1001 srcs = ["bench/convolution.cc"],
1002 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1003 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1004)
1005
1006xnnpack_benchmark(
1007 name = "deconvolution_bench",
1008 srcs = ["bench/deconvolution.cc"],
1009 copts = xnnpack_optional_tflite_copts(),
1010 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1011)
1012
1013xnnpack_benchmark(
1014 name = "global_average_pooling_bench",
1015 srcs = ["bench/global-average-pooling.cc"],
1016 deps = OPERATOR_BENCHMARK_DEPS,
1017)
1018
1019xnnpack_benchmark(
1020 name = "max_pooling_bench",
1021 srcs = ["bench/max-pooling.cc"],
1022 deps = OPERATOR_BENCHMARK_DEPS,
1023)
1024
1025xnnpack_benchmark(
1026 name = "sigmoid_bench",
1027 srcs = ["bench/sigmoid.cc"],
1028 deps = OPERATOR_BENCHMARK_DEPS,
1029)
1030
1031xnnpack_benchmark(
1032 name = "softargmax_bench",
1033 srcs = ["bench/softargmax.cc"],
1034 deps = OPERATOR_BENCHMARK_DEPS,
1035)
1036
Marat Dukhanc068bb62019-10-04 13:24:39 -07001037############################# End-to-end benchmarks ############################
1038
1039cc_library(
1040 name = "mobilenet_v1",
1041 srcs = ["models/mobilenet-v1.cc"],
1042 hdrs = ["models/models.h"],
1043 linkstatic = True,
1044 deps = [
1045 ":XNNPACK",
1046 "@pthreadpool",
1047 ],
1048)
1049
1050cc_library(
1051 name = "mobilenet_v2",
1052 srcs = ["models/mobilenet-v2.cc"],
1053 hdrs = ["models/models.h"],
1054 linkstatic = True,
1055 deps = [
1056 ":XNNPACK",
1057 "@pthreadpool",
1058 ],
1059)
1060
1061xnnpack_benchmark(
1062 name = "end2end_bench",
1063 srcs = ["bench/end2end.cc"],
1064 deps = [
1065 ":XNNPACK",
1066 ":mobilenet_v1",
1067 ":mobilenet_v2",
1068 "@pthreadpool",
1069 ],
1070)
1071
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001072#################### Accuracy evaluation for math functions ####################
1073
1074xnnpack_benchmark(
1075 name = "f32_exp_eval",
1076 srcs = [
1077 "eval/f32-exp.cc",
1078 "src/xnnpack/AlignedAllocator.h",
1079 ] + ACCURACY_EVAL_HDRS,
1080 deps = ACCURACY_EVAL_DEPS,
1081)
1082
Marat Dukhan515c9772019-10-17 18:07:57 -07001083xnnpack_benchmark(
1084 name = "f32_expminus_eval",
1085 srcs = [
1086 "eval/f32-expminus.cc",
1087 "src/xnnpack/AlignedAllocator.h",
1088 ] + ACCURACY_EVAL_HDRS,
1089 deps = ACCURACY_EVAL_DEPS,
1090)
1091
Marat Dukhan98ba4412019-10-23 02:14:28 -07001092xnnpack_benchmark(
1093 name = "f32_extexp_eval",
1094 srcs = [
1095 "eval/f32-extexp.cc",
1096 "src/xnnpack/AlignedAllocator.h",
1097 ] + ACCURACY_EVAL_HDRS,
1098 deps = ACCURACY_EVAL_DEPS,
1099)
1100
Marat Dukhan08c4a432019-10-03 09:29:21 -07001101######################### Unit tests for micro-kernels #########################
1102
1103xnnpack_unit_test(
1104 name = "f16_gemm_test",
1105 srcs = [
1106 "test/f16-gemm.cc",
1107 "test/gemm-microkernel-tester.h",
1108 "src/xnnpack/AlignedAllocator.h",
1109 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1110 deps = MICROKERNEL_TEST_DEPS,
1111)
1112
1113xnnpack_unit_test(
1114 name = "f32_argmaxpool_test",
1115 srcs = [
1116 "test/f32-argmaxpool.cc",
1117 "test/argmaxpool-microkernel-tester.h",
1118 "src/xnnpack/AlignedAllocator.h",
1119 ] + MICROKERNEL_TEST_HDRS,
1120 deps = MICROKERNEL_TEST_DEPS,
1121)
1122
1123xnnpack_unit_test(
1124 name = "f32_avgpool_test",
1125 srcs = [
1126 "test/f32-avgpool.cc",
1127 "test/avgpool-microkernel-tester.h",
1128 "src/xnnpack/AlignedAllocator.h",
1129 ] + MICROKERNEL_TEST_HDRS,
1130 deps = MICROKERNEL_TEST_DEPS,
1131)
1132
1133xnnpack_unit_test(
1134 name = "f32_clamp_test",
1135 srcs = [
1136 "test/f32-clamp.cc",
1137 "test/clamp-microkernel-tester.h",
1138 ] + MICROKERNEL_TEST_HDRS,
1139 deps = MICROKERNEL_TEST_DEPS,
1140)
1141
1142xnnpack_unit_test(
1143 name = "f32_igemm_test",
1144 srcs = [
1145 "test/f32-igemm.cc",
1146 "test/gemm-microkernel-tester.h",
1147 "src/xnnpack/AlignedAllocator.h",
1148 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1149 deps = MICROKERNEL_TEST_DEPS,
1150)
1151
1152xnnpack_unit_test(
1153 name = "f32_conv_hwc_test",
1154 srcs = [
1155 "test/f32-conv-hwc.cc",
1156 "test/conv-hwc-microkernel-tester.h",
1157 "src/xnnpack/AlignedAllocator.h",
1158 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1159 deps = MICROKERNEL_TEST_DEPS,
1160)
1161
1162xnnpack_unit_test(
1163 name = "f32_conv_hwc2spchw_test",
1164 srcs = [
1165 "test/f32-conv-hwc2spchw.cc",
1166 "test/conv-hwc2spchw-microkernel-tester.h",
1167 "src/xnnpack/AlignedAllocator.h",
1168 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1169 deps = MICROKERNEL_TEST_DEPS,
1170)
1171
1172xnnpack_unit_test(
1173 name = "f32_dwconv_test",
1174 srcs = [
1175 "test/f32-dwconv.cc",
1176 "test/dwconv-microkernel-tester.h",
1177 "src/xnnpack/AlignedAllocator.h",
1178 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1179 deps = MICROKERNEL_TEST_DEPS,
1180)
1181
1182xnnpack_unit_test(
1183 name = "f32_dwconv_spchw_test",
1184 srcs = [
1185 "test/f32-dwconv-spchw.cc",
1186 "test/dwconv-spchw-microkernel-tester.h",
1187 "src/xnnpack/AlignedAllocator.h",
1188 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1189 deps = MICROKERNEL_TEST_DEPS,
1190)
1191
1192xnnpack_unit_test(
1193 name = "f32_gavgpool_test",
1194 srcs = [
1195 "test/f32-gavgpool.cc",
1196 "test/gavgpool-microkernel-tester.h",
1197 "src/xnnpack/AlignedAllocator.h",
1198 ] + MICROKERNEL_TEST_HDRS,
1199 deps = MICROKERNEL_TEST_DEPS,
1200)
1201
1202xnnpack_unit_test(
1203 name = "f32_gavgpool_spchw_test",
1204 srcs = [
1205 "test/f32-gavgpool-spchw.cc",
1206 "test/gavgpool-spchw-microkernel-tester.h",
1207 "src/xnnpack/AlignedAllocator.h",
1208 ] + MICROKERNEL_TEST_HDRS,
1209 deps = MICROKERNEL_TEST_DEPS,
1210)
1211
1212xnnpack_unit_test(
1213 name = "f32_gemm_test",
1214 srcs = [
1215 "test/f32-gemm.cc",
1216 "test/gemm-microkernel-tester.h",
1217 "src/xnnpack/AlignedAllocator.h",
1218 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1219 deps = MICROKERNEL_TEST_DEPS,
1220)
1221
1222xnnpack_unit_test(
1223 name = "f32_gemminc_test",
1224 srcs = [
1225 "test/f32-gemminc.cc",
1226 "test/gemm-microkernel-tester.h",
1227 "src/xnnpack/AlignedAllocator.h",
1228 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1229 deps = MICROKERNEL_TEST_DEPS,
1230)
1231
1232xnnpack_unit_test(
1233 name = "f32_hswish_test",
1234 srcs = [
1235 "test/f32-hswish.cc",
1236 "test/hswish-microkernel-tester.h",
1237 ] + MICROKERNEL_TEST_HDRS,
1238 deps = MICROKERNEL_TEST_DEPS,
1239)
1240
1241xnnpack_unit_test(
1242 name = "f32_maxpool_test",
1243 srcs = [
1244 "test/f32-maxpool.cc",
1245 "test/maxpool-microkernel-tester.h",
1246 ] + MICROKERNEL_TEST_HDRS,
1247 deps = MICROKERNEL_TEST_DEPS,
1248)
1249
1250xnnpack_unit_test(
1251 name = "f32_pavgpool_test",
1252 srcs = [
1253 "test/f32-pavgpool.cc",
1254 "test/avgpool-microkernel-tester.h",
1255 "src/xnnpack/AlignedAllocator.h",
1256 ] + MICROKERNEL_TEST_HDRS,
1257 deps = MICROKERNEL_TEST_DEPS,
1258)
1259
1260xnnpack_unit_test(
1261 name = "f32_ppmm_test",
1262 srcs = [
1263 "test/f32-ppmm.cc",
1264 "test/gemm-microkernel-tester.h",
1265 "src/xnnpack/AlignedAllocator.h",
1266 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1267 deps = MICROKERNEL_TEST_DEPS,
1268)
1269
1270xnnpack_unit_test(
1271 name = "f32_prelu_test",
1272 srcs = [
1273 "test/f32-prelu.cc",
1274 "test/prelu-microkernel-tester.h",
1275 "src/xnnpack/AlignedAllocator.h",
1276 ] + MICROKERNEL_TEST_HDRS,
1277 deps = MICROKERNEL_TEST_DEPS,
1278)
1279
1280xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001281 name = "f32_raddexpminusmax_test",
1282 srcs = [
1283 "test/f32-raddexpminusmax.cc",
1284 "test/raddexpminusmax-microkernel-tester.h",
1285 ] + MICROKERNEL_TEST_HDRS,
1286 deps = MICROKERNEL_TEST_DEPS,
1287)
1288
1289xnnpack_unit_test(
1290 name = "f32_raddstoreexpminusmax_test",
1291 srcs = [
1292 "test/f32-raddstoreexpminusmax.cc",
1293 "test/raddstoreexpminusmax-microkernel-tester.h",
1294 ] + MICROKERNEL_TEST_HDRS,
1295 deps = MICROKERNEL_TEST_DEPS,
1296)
1297
1298xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001299 name = "f32_rmax_test",
1300 srcs = [
1301 "test/f32-rmax.cc",
1302 "test/rmax-microkernel-tester.h",
1303 ] + MICROKERNEL_TEST_HDRS,
1304 deps = MICROKERNEL_TEST_DEPS,
1305)
1306
1307xnnpack_unit_test(
1308 name = "f32_spmm_test",
1309 srcs = [
1310 "test/f32-spmm.cc",
1311 "test/spmm-microkernel-tester.h",
1312 "src/xnnpack/AlignedAllocator.h",
1313 ] + MICROKERNEL_TEST_HDRS,
1314 deps = MICROKERNEL_TEST_DEPS,
1315)
1316
1317xnnpack_unit_test(
1318 name = "f32_vadd_test",
1319 srcs = [
1320 "test/f32-vadd.cc",
1321 "test/vadd-microkernel-tester.h",
1322 ] + MICROKERNEL_TEST_HDRS,
1323 deps = MICROKERNEL_TEST_DEPS,
1324)
1325
1326xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001327 name = "f32_vmul_test",
1328 srcs = [
1329 "test/f32-vmul.cc",
1330 "test/vmul-microkernel-tester.h",
1331 ] + MICROKERNEL_TEST_HDRS,
1332 deps = MICROKERNEL_TEST_DEPS,
1333)
1334
1335xnnpack_unit_test(
1336 name = "f32_vmulcaddc_test",
1337 srcs = [
1338 "test/f32-vmulcaddc.cc",
1339 "test/vmulcaddc-microkernel-tester.h",
1340 "src/xnnpack/AlignedAllocator.h",
1341 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1342 deps = MICROKERNEL_TEST_DEPS,
1343)
1344
1345xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001346 name = "f32_vscale_test",
1347 srcs = [
1348 "test/f32-vscale.cc",
1349 "test/vscale-microkernel-tester.h",
1350 ] + MICROKERNEL_TEST_HDRS,
1351 deps = MICROKERNEL_TEST_DEPS,
1352)
1353
1354xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001355 name = "f32_vscaleexpminusmax_test",
1356 srcs = [
1357 "test/f32-vscaleexpminusmax.cc",
1358 "test/vscaleexpminusmax-microkernel-tester.h",
1359 ] + MICROKERNEL_TEST_HDRS,
1360 deps = MICROKERNEL_TEST_DEPS,
1361)
1362
1363xnnpack_unit_test(
1364 name = "f32_vsub_test",
1365 srcs = [
1366 "test/f32-vsub.cc",
1367 "test/vsub-microkernel-tester.h",
1368 ] + MICROKERNEL_TEST_HDRS,
1369 deps = MICROKERNEL_TEST_DEPS,
1370)
1371
1372xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001373 name = "q8_avgpool_test",
1374 srcs = [
1375 "test/q8-avgpool.cc",
1376 "test/avgpool-microkernel-tester.h",
1377 "src/xnnpack/AlignedAllocator.h",
1378 ] + MICROKERNEL_TEST_HDRS,
1379 deps = MICROKERNEL_TEST_DEPS,
1380)
1381
1382xnnpack_unit_test(
1383 name = "q8_igemm_test",
1384 srcs = [
1385 "test/q8-igemm.cc",
1386 "test/gemm-microkernel-tester.h",
1387 "src/xnnpack/AlignedAllocator.h",
1388 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1389 deps = MICROKERNEL_TEST_DEPS,
1390)
1391
1392xnnpack_unit_test(
1393 name = "q8_dwconv_test",
1394 srcs = [
1395 "test/q8-dwconv.cc",
1396 "test/dwconv-microkernel-tester.h",
1397 "src/xnnpack/AlignedAllocator.h",
1398 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1399 deps = MICROKERNEL_TEST_DEPS,
1400)
1401
1402xnnpack_unit_test(
1403 name = "q8_gavgpool_test",
1404 srcs = [
1405 "test/q8-gavgpool.cc",
1406 "test/gavgpool-microkernel-tester.h",
1407 "src/xnnpack/AlignedAllocator.h",
1408 ] + MICROKERNEL_TEST_HDRS,
1409 deps = MICROKERNEL_TEST_DEPS,
1410)
1411
1412xnnpack_unit_test(
1413 name = "q8_gemm_test",
1414 srcs = [
1415 "test/q8-gemm.cc",
1416 "test/gemm-microkernel-tester.h",
1417 "src/xnnpack/AlignedAllocator.h",
1418 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1419 deps = MICROKERNEL_TEST_DEPS,
1420)
1421
1422xnnpack_unit_test(
1423 name = "q8_vadd_test",
1424 srcs = [
1425 "test/q8-vadd.cc",
1426 "test/vadd-microkernel-tester.h",
1427 ] + MICROKERNEL_TEST_HDRS,
1428 deps = MICROKERNEL_TEST_DEPS,
1429)
1430
1431xnnpack_unit_test(
1432 name = "u8_clamp_test",
1433 srcs = [
1434 "test/u8-clamp.cc",
1435 "test/clamp-microkernel-tester.h",
1436 ] + MICROKERNEL_TEST_HDRS,
1437 deps = MICROKERNEL_TEST_DEPS,
1438)
1439
1440xnnpack_unit_test(
1441 name = "u8_lut32norm_test",
1442 srcs = [
1443 "test/u8-lut32norm.cc",
1444 "test/lut-norm-microkernel-tester.h",
1445 ] + MICROKERNEL_TEST_HDRS,
1446 deps = MICROKERNEL_TEST_DEPS,
1447)
1448
1449xnnpack_unit_test(
1450 name = "u8_maxpool_test",
1451 srcs = [
1452 "test/u8-maxpool.cc",
1453 "test/maxpool-microkernel-tester.h",
1454 ] + MICROKERNEL_TEST_HDRS,
1455 deps = MICROKERNEL_TEST_DEPS,
1456)
1457
1458xnnpack_unit_test(
1459 name = "u8_rmax_test",
1460 srcs = [
1461 "test/u8-rmax.cc",
1462 "test/rmax-microkernel-tester.h",
1463 ] + MICROKERNEL_TEST_HDRS,
1464 deps = MICROKERNEL_TEST_DEPS,
1465)
1466
1467xnnpack_unit_test(
1468 name = "x32_packx_test",
1469 srcs = [
1470 "test/x32-packx.cc",
1471 "test/pack-microkernel-tester.h",
1472 "src/xnnpack/AlignedAllocator.h",
1473 ] + MICROKERNEL_TEST_HDRS,
1474 deps = MICROKERNEL_TEST_DEPS,
1475)
1476
1477xnnpack_unit_test(
1478 name = "x32_pad_test",
1479 srcs = [
1480 "test/x32-pad.cc",
1481 "test/pad-microkernel-tester.h",
1482 ] + MICROKERNEL_TEST_HDRS,
1483 deps = MICROKERNEL_TEST_DEPS,
1484)
1485
1486xnnpack_unit_test(
1487 name = "x32_unpool_test",
1488 srcs = [
1489 "test/x32-unpool.cc",
1490 "test/unpool-microkernel-tester.h",
1491 ] + MICROKERNEL_TEST_HDRS,
1492 deps = MICROKERNEL_TEST_DEPS,
1493)
1494
1495xnnpack_unit_test(
1496 name = "x32_zip_test",
1497 srcs = [
1498 "test/x32-zip.cc",
1499 "test/zip-microkernel-tester.h",
1500 ] + MICROKERNEL_TEST_HDRS,
1501 deps = MICROKERNEL_TEST_DEPS,
1502)
1503
1504xnnpack_unit_test(
1505 name = "x8_lut_test",
1506 srcs = [
1507 "test/x8-lut.cc",
1508 "test/lut-microkernel-tester.h",
1509 ] + MICROKERNEL_TEST_HDRS,
1510 deps = MICROKERNEL_TEST_DEPS,
1511)
1512
1513xnnpack_unit_test(
1514 name = "x8_zip_test",
1515 srcs = [
1516 "test/x8-zip.cc",
1517 "test/zip-microkernel-tester.h",
1518 ] + MICROKERNEL_TEST_HDRS,
1519 deps = MICROKERNEL_TEST_DEPS,
1520)
1521
1522########################### Size test for the library ##########################
1523
1524xnnpack_binary(
1525 name = "size_test",
1526 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001527 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001528)
1529
1530########################### Unit tests for operators ###########################
1531
1532xnnpack_unit_test(
1533 name = "add_test",
1534 srcs = [
1535 "test/add.cc",
1536 "test/add-operator-tester.h",
1537 ],
1538 deps = OPERATOR_TEST_DEPS,
1539)
1540
1541xnnpack_unit_test(
1542 name = "argmax_pooling_test",
1543 srcs = [
1544 "test/argmax-pooling.cc",
1545 "test/argmax-pooling-operator-tester.h",
1546 ] + OPERATOR_TEST_PARAMS_HDRS,
1547 deps = OPERATOR_TEST_DEPS,
1548)
1549
1550xnnpack_unit_test(
1551 name = "average_pooling_test",
1552 srcs = [
1553 "test/average-pooling.cc",
1554 "test/average-pooling-operator-tester.h",
1555 ] + OPERATOR_TEST_PARAMS_HDRS,
1556 deps = OPERATOR_TEST_DEPS,
1557)
1558
1559xnnpack_unit_test(
1560 name = "channel_pad_test",
1561 srcs = [
1562 "test/channel-pad.cc",
1563 "test/channel-pad-operator-tester.h",
1564 ] + OPERATOR_TEST_PARAMS_HDRS,
1565 deps = OPERATOR_TEST_DEPS,
1566)
1567
1568xnnpack_unit_test(
1569 name = "channel_shuffle_test",
1570 srcs = [
1571 "test/channel-shuffle.cc",
1572 "test/channel-shuffle-operator-tester.h",
1573 ],
1574 deps = OPERATOR_TEST_DEPS,
1575)
1576
1577xnnpack_unit_test(
1578 name = "clamp_test",
1579 srcs = [
1580 "test/clamp.cc",
1581 "test/clamp-operator-tester.h",
1582 ],
1583 deps = OPERATOR_TEST_DEPS,
1584)
1585
1586xnnpack_unit_test(
1587 name = "convolution_test",
1588 srcs = [
1589 "test/convolution.cc",
1590 "test/convolution-operator-tester.h",
1591 ],
1592 deps = OPERATOR_TEST_DEPS,
1593)
1594
1595xnnpack_unit_test(
1596 name = "convolution_spnchw_test",
1597 srcs = [
1598 "test/convolution-spnchw.cc",
1599 "test/convolution-spnchw-operator-tester.h",
1600 ],
1601 deps = OPERATOR_TEST_DEPS,
1602)
1603
1604xnnpack_unit_test(
1605 name = "deconvolution_test",
1606 srcs = [
1607 "test/deconvolution.cc",
1608 "test/deconvolution-operator-tester.h",
1609 ] + OPERATOR_TEST_PARAMS_HDRS,
1610 deps = OPERATOR_TEST_DEPS,
1611)
1612
1613xnnpack_unit_test(
1614 name = "fully_connected_test",
1615 srcs = [
1616 "test/fully-connected.cc",
1617 "test/fully-connected-operator-tester.h",
1618 ],
1619 deps = OPERATOR_TEST_DEPS,
1620)
1621
1622xnnpack_unit_test(
1623 name = "global_average_pooling_test",
1624 srcs = [
1625 "test/global-average-pooling.cc",
1626 "test/global-average-pooling-operator-tester.h",
1627 ] + OPERATOR_TEST_PARAMS_HDRS,
1628 deps = OPERATOR_TEST_DEPS,
1629)
1630
1631xnnpack_unit_test(
1632 name = "global_average_pooling_spnchw_test",
1633 srcs = [
1634 "test/global-average-pooling-spnchw.cc",
1635 "test/global-average-pooling-spnchw-operator-tester.h",
1636 ],
1637 deps = OPERATOR_TEST_DEPS,
1638)
1639
1640xnnpack_unit_test(
1641 name = "hardswish_test",
1642 srcs = [
1643 "test/hardswish.cc",
1644 "test/hardswish-operator-tester.h",
1645 ],
1646 deps = OPERATOR_TEST_DEPS,
1647)
1648
1649xnnpack_unit_test(
1650 name = "leaky_relu_test",
1651 srcs = [
1652 "test/leaky-relu.cc",
1653 "test/leaky-relu-operator-tester.h",
1654 ],
1655 deps = OPERATOR_TEST_DEPS,
1656)
1657
1658xnnpack_unit_test(
1659 name = "max_pooling_test",
1660 srcs = [
1661 "test/max-pooling.cc",
1662 "test/max-pooling-operator-tester.h",
1663 ] + OPERATOR_TEST_PARAMS_HDRS,
1664 deps = OPERATOR_TEST_DEPS,
1665)
1666
1667xnnpack_unit_test(
1668 name = "prelu_test",
1669 srcs = [
1670 "test/prelu.cc",
1671 "test/prelu-operator-tester.h",
1672 ] + OPERATOR_TEST_PARAMS_HDRS,
1673 deps = OPERATOR_TEST_DEPS,
1674)
1675
1676xnnpack_unit_test(
1677 name = "sigmoid_test",
1678 srcs = [
1679 "test/sigmoid.cc",
1680 "test/sigmoid-operator-tester.h",
1681 ],
1682 deps = OPERATOR_TEST_DEPS,
1683)
1684
1685xnnpack_unit_test(
1686 name = "softargmax_test",
1687 srcs = [
1688 "test/softargmax.cc",
1689 "test/softargmax-operator-tester.h",
1690 ],
1691 deps = OPERATOR_TEST_DEPS,
1692)
1693
1694xnnpack_unit_test(
1695 name = "unpooling_test",
1696 srcs = [
1697 "test/unpooling.cc",
1698 "test/unpooling-operator-tester.h",
1699 ],
1700 deps = OPERATOR_TEST_DEPS,
1701)
1702
1703############################# Build configurations #############################
1704
1705config_setting(
1706 name = "linux_k8",
1707 values = {
1708 "cpu": "k8",
1709 },
1710)
1711
1712config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07001713 name = "linux_aarch64",
1714 values = {
1715 "cpu": "aarch64",
1716 },
1717)
1718
1719config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001720 name = "android",
1721 values = {"crosstool_top": "//external:android/crosstool"},
1722)
1723
1724config_setting(
1725 name = "android_armv7",
1726 values = {
1727 "crosstool_top": "//external:android/crosstool",
1728 "cpu": "armeabi-v7a",
1729 },
1730)
1731
1732config_setting(
1733 name = "android_arm64",
1734 values = {
1735 "crosstool_top": "//external:android/crosstool",
1736 "cpu": "arm64-v8a",
1737 },
1738)
1739
1740config_setting(
1741 name = "android_x86",
1742 values = {
1743 "crosstool_top": "//external:android/crosstool",
1744 "cpu": "x86",
1745 },
1746)
1747
1748config_setting(
1749 name = "android_x86_64",
1750 values = {
1751 "crosstool_top": "//external:android/crosstool",
1752 "cpu": "x86_64",
1753 },
1754)
1755
1756config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07001757 name = "macos_x86_64",
1758 values = {
1759 "apple_platform_type": "macos",
1760 "cpu": "darwin",
1761 },
1762)
1763
1764config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001765 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001766 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07001767)
1768
1769config_setting(
1770 name = "emscripten_wasm",
1771 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001772 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001773 "cpu": "wasm",
1774 },
1775)
1776
1777config_setting(
1778 name = "emscripten_wasmsimd",
1779 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001780 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001781 "cpu": "wasm",
1782 "features": "wasmsimd",
1783 },
1784)
1785
1786config_setting(
1787 name = "emscripten_asmjs",
1788 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001789 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001790 "cpu": "asmjs",
1791 },
1792)
1793
Marat Dukhan174706e2019-10-09 16:33:53 -07001794# Builds with -c dbg
Marat Dukhan08c4a432019-10-03 09:29:21 -07001795config_setting(
1796 name = "debug_build",
1797 values = {
1798 "compilation_mode": "dbg",
1799 },
Marat Dukhan08c4a432019-10-03 09:29:21 -07001800)
1801
Marat Dukhan174706e2019-10-09 16:33:53 -07001802# Builds with -c opt
Marat Dukhan08c4a432019-10-03 09:29:21 -07001803config_setting(
1804 name = "optimized_build",
1805 values = {
1806 "compilation_mode": "opt",
1807 },
Marat Dukhan08c4a432019-10-03 09:29:21 -07001808)
1809
1810# Enables usage of assembly kernels.
1811config_setting(
1812 name = "xnn_enable_assembly_explicit_true",
1813 define_values = {"xnn_enable_assembly": "true"},
1814)
1815
1816# Disables usage of assembly kernels.
1817config_setting(
1818 name = "xnn_enable_assembly_explicit_false",
1819 define_values = {"xnn_enable_assembly": "false"},
1820)