blob: 1f93ce0e7828c775832976397e4fe9902d026925 [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",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700112 "src/f32-spmm/1x1-scalar.c",
113 "src/f32-spmm/2x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700114 "src/f32-spmm/2x1-scalar.c",
115 "src/f32-spmm/4x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700116 "src/f32-spmm/4x1-scalar.c",
117 "src/f32-spmm/8x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700118 "src/f32-spmm/8x1-scalar.c",
Erich Elsenc6afd9b2019-10-24 16:10:53 -0700119 "src/f32-spmm/8x2-scalar.c",
120 "src/f32-spmm/8x4-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700121 "src/f32-vadd/scalar.c",
122 "src/f32-vmul/scalar.c",
123 "src/f32-vmulcaddc/c1-scalar-x2.c",
124 "src/f32-vsub/scalar.c",
125 "src/q8-avgpool/mp9p8q-scalar.c",
126 "src/q8-avgpool/up9-scalar.c",
127 "src/q8-igemm/2x2-scalar.c",
128 "src/q8-dwconv/up1x9-scalar.c",
129 "src/q8-gavgpool/mp7p7q-scalar.c",
130 "src/q8-gavgpool/up7-scalar.c",
131 "src/q8-gemm/2x2-scalar.c",
132 "src/q8-vadd/scalar.c",
133 "src/u8-clamp/scalar.c",
134 "src/u8-lut32norm/scalar.c",
135 "src/u8-maxpool/9p8q-scalar.c",
136 "src/u8-rmax/scalar.c",
137 "src/x32-packx/x2-scalar.c",
138 "src/x32-packx/x3-scalar.c",
139 "src/x32-packx/x4-scalar.c",
140 "src/x32-pad/x2-scalar.c",
141 "src/x32-unpool/scalar.c",
142 "src/x32-zip/x2-scalar.c",
143 "src/x32-zip/x3-scalar.c",
144 "src/x32-zip/x4-scalar.c",
145 "src/x32-zip/xm-scalar.c",
146 "src/x8-lut/scalar.c",
147 "src/x8-zip/x2-scalar.c",
148 "src/x8-zip/x3-scalar.c",
149 "src/x8-zip/x4-scalar.c",
150 "src/x8-zip/xm-scalar.c",
151]
152
153PSIMD_UKERNELS = [
154 "src/f32-argmaxpool/mp9p8q-psimd.c",
155 "src/f32-argmaxpool/up4-psimd.c",
156 "src/f32-argmaxpool/up9-psimd.c",
157 "src/f32-avgpool/mp9p8q-psimd.c",
158 "src/f32-avgpool/up9-psimd.c",
159 "src/f32-clamp/psimd.c",
160 "src/f32-igemm/1x8-psimd-loadsplat.c",
161 "src/f32-igemm/1x8-psimd-splat.c",
162 "src/f32-igemm/1x8s4-psimd.c",
163 "src/f32-igemm/4x2c4-psimd.c",
164 "src/f32-igemm/4x8-psimd-loadsplat.c",
165 "src/f32-igemm/4x8-psimd-splat.c",
166 "src/f32-igemm/4x8s4-psimd.c",
167 "src/f32-igemm/6x8-psimd-loadsplat.c",
168 "src/f32-igemm/6x8-psimd-splat.c",
169 "src/f32-igemm/6x8s4-psimd.c",
170 "src/f32-dwconv/up4x25-psimd.c",
171 "src/f32-dwconv/up4x4-psimd.c",
172 "src/f32-dwconv/up4x9-psimd.c",
173 "src/f32-gavgpool/mp7p7q-psimd.c",
174 "src/f32-gavgpool/up7-psimd.c",
175 "src/f32-gemm/1x8-psimd-loadsplat.c",
176 "src/f32-gemm/1x8-psimd-splat.c",
177 "src/f32-gemm/1x8s4-psimd.c",
178 "src/f32-gemm/4x8-psimd-loadsplat.c",
179 "src/f32-gemm/4x8-psimd-splat.c",
180 "src/f32-gemm/4x8s4-psimd.c",
181 "src/f32-gemm/6x8-psimd-loadsplat.c",
182 "src/f32-gemm/6x8-psimd-splat.c",
183 "src/f32-gemm/6x8s4-psimd.c",
184 "src/f32-gemminc/1x8-psimd-loadsplat.c",
185 "src/f32-gemminc/1x8-psimd-splat.c",
186 "src/f32-gemminc/1x8s4-psimd.c",
187 "src/f32-gemminc/4x8-psimd-loadsplat.c",
188 "src/f32-gemminc/4x8-psimd-splat.c",
189 "src/f32-gemminc/4x8s4-psimd.c",
190 "src/f32-gemminc/6x8-psimd-loadsplat.c",
191 "src/f32-gemminc/6x8-psimd-splat.c",
192 "src/f32-gemminc/6x8s4-psimd.c",
193 "src/f32-hswish/psimd.c",
194 "src/f32-maxpool/9p8q-psimd.c",
195 "src/f32-pavgpool/mp9p8q-psimd.c",
196 "src/f32-pavgpool/up9-psimd.c",
197 "src/f32-ppmm/4x8-psimd.c",
198 "src/f32-prelu/x4-psimd.c",
199 "src/f32-vadd/psimd.c",
200 "src/f32-vmul/psimd.c",
201 "src/f32-vmulcaddc/c4-psimd-x2.c",
202 "src/f32-vsub/psimd.c",
203 "src/x32-packx/x4-psimd.c",
204 "src/x32-pad/x2-psimd.c",
205 "src/x32-unpool/psimd.c",
206 "src/x32-zip/x2-psimd.c",
207 "src/x32-zip/x3-psimd.c",
208 "src/x32-zip/x4-psimd.c",
209 "src/x32-zip/xm-psimd.c",
210]
211
212# ISA-specific micro-kernels
213NEON_UKERNELS = [
214 "src/f32-avgpool/mp9p8q-neon.c",
215 "src/f32-avgpool/up9-neon.c",
216 "src/f32-clamp/neon.c",
217 "src/f32-igemm/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700218 "src/f32-igemm/4x2-neon-ld64.c",
219 "src/f32-igemm/4x4-neon-ld64.c",
220 "src/f32-igemm/4x8-neon-ld128.c",
221 "src/f32-igemm/4x8-neon-ld64.c",
222 "src/f32-igemm/6x8-neon-ld64.c",
223 "src/f32-dwconv/up4x9-neon.c",
224 "src/f32-gavgpool-spchw/neon-x4.c",
225 "src/f32-gavgpool/mp7p7q-neon.c",
226 "src/f32-gavgpool/up7-neon.c",
227 "src/f32-gemm/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700228 "src/f32-gemm/4x2-neon-ld64.c",
229 "src/f32-gemm/4x8-neon-ld128.c",
230 "src/f32-gemm/4x8-neon-ld64.c",
231 "src/f32-gemm/5x8-neon-ld64.c",
232 "src/f32-gemm/6x8-neon-ld64.c",
233 "src/f32-gemminc/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700234 "src/f32-gemminc/4x8-neon-ld128.c",
235 "src/f32-gemminc/4x8-neon-ld64.c",
236 "src/f32-gemminc/5x8-neon-ld64.c",
237 "src/f32-gemminc/6x8-neon-ld64.c",
238 "src/f32-hswish/neon.c",
239 "src/f32-pavgpool/mp9p8q-neon.c",
240 "src/f32-pavgpool/up9-neon.c",
241 "src/f32-ppmm/4x8-neon.c",
242 "src/f32-ppmm/8x8-neon.c",
243 "src/f32-rmax/neon.c",
244 "src/f32-vmulcaddc/c4-neon-x2.c",
245 "src/q8-avgpool/mp9p8q-neon.c",
246 "src/q8-avgpool/up9-neon.c",
247 "src/q8-igemm/4x8-neon.c",
248 "src/q8-igemm/8x8-neon.c",
249 "src/q8-dwconv/up8x9-neon.c",
250 "src/q8-gavgpool/mp7p7q-neon.c",
251 "src/q8-gavgpool/up7-neon.c",
252 "src/q8-gemm/4x8-neon.c",
253 "src/q8-gemm/8x8-neon.c",
254 "src/q8-vadd/neon.c",
255 "src/u8-clamp/neon.c",
256 "src/u8-maxpool/9p8q-neon.c",
257 "src/u8-rmax/neon.c",
258 "src/x32-packx/x4-neon-st4.c",
259 "src/x32-pad/x2-neon.c",
260 "src/x32-zip/x2-neon.c",
261 "src/x32-zip/x3-neon.c",
262 "src/x32-zip/x4-neon.c",
263 "src/x32-zip/xm-neon.c",
264 "src/x8-zip/x2-neon.c",
265 "src/x8-zip/x3-neon.c",
266 "src/x8-zip/x4-neon.c",
267 "src/x8-zip/xm-neon.c",
268]
269
270NEONFMA_UKERNELS = [
Frank Barcharddb45b6a2019-10-09 16:42:45 -0700271 "src/f32-igemm/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700272 "src/f32-igemm/4x2-neonfma-ld64.c",
273 "src/f32-igemm/4x4-neonfma-ld64.c",
274 "src/f32-igemm/4x8-neonfma-ld128.c",
275 "src/f32-igemm/4x8-neonfma-ld64.c",
276 "src/f32-igemm/6x8-neonfma-ld64.c",
277 "src/f32-dwconv/up4x9-neonfma.c",
278 "src/f32-dwconv/up8x9-neonfma.c",
279 "src/f32-gemm/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700280 "src/f32-gemm/4x2-neonfma-ld64.c",
281 "src/f32-gemm/4x8-neonfma-ld128.c",
282 "src/f32-gemm/4x8-neonfma-ld64.c",
283 "src/f32-gemm/5x8-neonfma-ld64.c",
284 "src/f32-gemm/6x8-neonfma-ld64.c",
285 "src/f32-gemminc/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700286 "src/f32-gemminc/4x8-neonfma-ld128.c",
287 "src/f32-gemminc/4x8-neonfma-ld64.c",
288 "src/f32-gemminc/5x8-neonfma-ld64.c",
289 "src/f32-gemminc/6x8-neonfma-ld64.c",
290 "src/f32-hswish/neonfma.c",
291 "src/f32-ppmm/4x8-neonfma.c",
292 "src/f32-ppmm/8x8-neonfma.c",
293 "src/f32-vmulcaddc/c4-neonfma-x2.c",
294]
295
296AARCH64_NEONFMA_UKERNELS = [
297 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
298 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
299 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
300 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
301 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
302 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
303 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
304 "src/f32-spmm/12x1-neonfma.c",
305 "src/f32-spmm/12x2-neonfma.c",
306 "src/f32-spmm/12x4-neonfma.c",
307 "src/f32-spmm/16x1-neonfma-pipelined.c",
308 "src/f32-spmm/16x1-neonfma-unroll2.c",
309 "src/f32-spmm/16x1-neonfma.c",
310 "src/f32-spmm/16x2-neonfma.c",
311 "src/f32-spmm/16x4-neonfma.c",
312 "src/f32-spmm/4x1-neonfma-pipelined.c",
313 "src/f32-spmm/4x1-neonfma-unroll2.c",
314 "src/f32-spmm/4x1-neonfma.c",
315 "src/f32-spmm/4x2-neonfma.c",
316 "src/f32-spmm/4x4-neonfma.c",
317 "src/f32-spmm/8x1-neonfma-pipelined.c",
318 "src/f32-spmm/8x1-neonfma-unroll2.c",
319 "src/f32-spmm/8x1-neonfma.c",
320 "src/f32-spmm/8x2-neonfma.c",
321 "src/f32-spmm/8x4-neonfma.c",
322]
323
324AARCH64_NEONFP16ARITH_UKERNELS = [
325 "src/f16-gemm/4x8-neonfp16arith-ld64.c",
326 "src/f16-gemm/6x8-neonfp16arith-ld64.c",
327 "src/f16-gemm/8x8-neonfp16arith-ld64.c",
328]
329
330SSE_UKERNELS = [
331 "src/f32-avgpool/mp9p8q-sse.c",
332 "src/f32-avgpool/up9-sse.c",
333 "src/f32-clamp/sse.c",
334 "src/f32-igemm/1x8-sse-dup.c",
335 "src/f32-igemm/1x8-sse-load1.c",
336 "src/f32-igemm/1x8s4-sse.c",
337 "src/f32-igemm/4x2c4-sse.c",
338 "src/f32-igemm/4x8-sse-dup.c",
339 "src/f32-igemm/4x8-sse-load1.c",
340 "src/f32-igemm/4x8s4-sse.c",
341 "src/f32-dwconv/up4x25-sse.c",
342 "src/f32-dwconv/up4x4-sse.c",
343 "src/f32-dwconv/up4x9-sse.c",
344 "src/f32-gavgpool-spchw/sse-x4.c",
345 "src/f32-gavgpool/mp7p7q-sse.c",
346 "src/f32-gavgpool/up7-sse.c",
347 "src/f32-gemm/1x8-sse-dup.c",
348 "src/f32-gemm/1x8-sse-load1.c",
349 "src/f32-gemm/1x8s4-sse.c",
350 "src/f32-gemm/4x8-sse-dup.c",
351 "src/f32-gemm/4x8-sse-load1.c",
352 "src/f32-gemm/4x8s4-sse.c",
353 "src/f32-gemminc/1x8-sse-dup.c",
354 "src/f32-gemminc/1x8-sse-load1.c",
355 "src/f32-gemminc/1x8s4-sse.c",
356 "src/f32-gemminc/4x8-sse-dup.c",
357 "src/f32-gemminc/4x8-sse-load1.c",
358 "src/f32-gemminc/4x8s4-sse.c",
359 "src/f32-hswish/sse.c",
360 "src/f32-maxpool/9p8q-sse.c",
361 "src/f32-pavgpool/mp9p8q-sse.c",
362 "src/f32-pavgpool/up9-sse.c",
363 "src/f32-dwconv-spchw/3x3p1-sse.c",
364 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
365 "src/f32-ppmm/4x8-sse.c",
366 "src/f32-prelu/x4-sse.c",
367 "src/f32-rmax/sse.c",
368 "src/f32-spmm/4x1-sse.c",
369 "src/f32-spmm/8x1-sse.c",
370 "src/f32-vadd/sse.c",
371 "src/f32-vmul/sse.c",
372 "src/f32-vmulcaddc/c4-sse-x2.c",
373 "src/f32-vsub/sse.c",
374 "src/x32-packx/x4-sse.c",
375]
376
377SSE2_UKERNELS = [
378 "src/f32-argmaxpool/mp9p8q-sse2.c",
379 "src/f32-argmaxpool/up4-sse2.c",
380 "src/f32-argmaxpool/up9-sse2.c",
381 "src/q8-avgpool/mp9p8q-sse2.c",
382 "src/q8-avgpool/up9-sse2.c",
383 "src/q8-igemm/4x4c2-sse2.c",
384 "src/q8-dwconv/up8x9-sse2.c",
385 "src/q8-gavgpool/mp7p7q-sse2.c",
386 "src/q8-gavgpool/up7-sse2.c",
387 "src/q8-gemm/2x4c8-sse2.c",
388 "src/q8-gemm/4x4c2-sse2.c",
389 "src/q8-vadd/sse2.c",
390 "src/u8-clamp/sse2.c",
391 "src/u8-maxpool/9p8q-sse2.c",
392 "src/u8-rmax/sse2.c",
393 "src/x32-pad/x2-sse2.c",
394 "src/x32-zip/x2-sse2.c",
395 "src/x32-zip/x3-sse2.c",
396 "src/x32-zip/x4-sse2.c",
397 "src/x32-zip/xm-sse2.c",
398 "src/x8-zip/x2-sse2.c",
399 "src/x8-zip/x3-sse2.c",
400 "src/x8-zip/x4-sse2.c",
401 "src/x8-zip/xm-sse2.c",
402]
403
404AVX_UKERNELS = [
405 "src/f32-rmax/avx.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700406 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700407]
408
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700409AVX2_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700410 "src/f32-raddexpminusmax/avx2-p5-unroll64.c",
411 "src/f32-raddstoreexpminusmax/avx2-p5-unroll64.c",
412 "src/f32-vscaleexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700413 "src/math/exp-avx2-p5.c",
414 "src/math/exp-avx2-perm-p3.c",
415 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -0700416 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700417 "src/math/extexp-avx2-p5.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700418]
419
Marat Dukhan08c4a432019-10-03 09:29:21 -0700420AVX512F_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700421 "src/f32-raddexpminusmax/avx512f-p5-scalef-unroll128.c",
422 "src/f32-raddstoreexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700423 "src/f32-rmax/avx512f.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700424 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700425 "src/f32-vscaleexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700426 "src/math/exp-avx512f-p5-scalef.c",
427 "src/math/exp-avx512f-p5.c",
428 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700429 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700430]
431
432AARCH32_ASM_UKERNELS = [
433 "src/q8-dwconv/up8x9-aarch32-neon.S",
434]
435
436AARCH64_ASM_UKERNELS = [
437 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
438 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
439 "src/f32-gemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700440 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700441 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a57.S",
442 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a75.S",
443 "src/f32-gemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700444 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700445 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a57.S",
446 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a75.S",
447 "src/f32-gemm/4x8-aarch64-neonfma-ld128.S",
448 "src/f32-gemm/4x8-aarch64-neonfma-ld64.S",
449 "src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700450 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700451 "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",
Frank Barchard46fb8072019-10-25 12:54:22 -0700461 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700462 "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",
Frank Barchard46fb8072019-10-25 12:54:22 -0700478 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700479 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
480 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700481 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700482 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
483 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
484 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
485]
486
487INTERNAL_MICROKERNEL_HDRS = [
488 "src/xnnpack/argmaxpool.h",
489 "src/xnnpack/avgpool.h",
490 "src/xnnpack/clamp.h",
491 "src/xnnpack/common.h",
492 "src/xnnpack/conv.h",
493 "src/xnnpack/dwconv.h",
494 "src/xnnpack/gavgpool.h",
495 "src/xnnpack/gemm.h",
496 "src/xnnpack/hswish.h",
497 "src/xnnpack/igemm.h",
498 "src/xnnpack/lut.h",
499 "src/xnnpack/math.h",
500 "src/xnnpack/maxpool.h",
501 "src/xnnpack/packx.h",
502 "src/xnnpack/pad.h",
503 "src/xnnpack/params.h",
504 "src/xnnpack/pavgpool.h",
505 "src/xnnpack/ppmm.h",
506 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700507 "src/xnnpack/raddexpminusmax.h",
508 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700509 "src/xnnpack/rmax.h",
510 "src/xnnpack/scalar-utils.h",
511 "src/xnnpack/spmm.h",
512 "src/xnnpack/unpool.h",
513 "src/xnnpack/vadd.h",
514 "src/xnnpack/vmul.h",
515 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700516 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700517 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700518 "src/xnnpack/vsub.h",
519 "src/xnnpack/zip.h",
520]
521
522INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
523 "include/xnnpack.h",
524 "src/xnnpack/allocator.h",
525 "src/xnnpack/compute.h",
526 "src/xnnpack/im2col.h",
527 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700528 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700529 "src/xnnpack/operator.h",
530 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700531 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700532 "src/xnnpack/requantization.h",
533]
534
535ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
536 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700537]
538
539MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
540 "src/xnnpack/requantization.h",
541 "include/xnnpack.h",
542]
543
544MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
545 "src/xnnpack/isa-checks.h",
546 "src/xnnpack/requantization.h",
547 "include/xnnpack.h",
548]
549
550OPERATOR_TEST_PARAMS_HDRS = [
551 "src/xnnpack/params.h",
552 "src/xnnpack/common.h",
553]
554
555WEIGHTS_PACK_HDRS = [
556 "src/xnnpack/pack.h",
557 "src/xnnpack/operator.h",
558 "src/xnnpack/compute.h",
559]
560
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700561LOGGING_COPTS = select({
562 # No logging in optimized mode
563 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
564 # Full logging in debug mode
565 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
566 # Error-only logging in default (fastbuild) mode
567 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
568})
569
570LOGGING_HDRS = [
571 "src/xnnpack/log.h",
572]
573
Marat Dukhan08c4a432019-10-03 09:29:21 -0700574xnnpack_cc_library(
575 name = "scalar_ukernels",
576 srcs = SCALAR_UKERNELS,
577 hdrs = INTERNAL_HDRS,
578 aarch32_copts = ["-marm"],
579 copts = xnnpack_std_copts(),
580 deps = [
581 "@FP16",
582 "@FXdiv",
583 ],
584)
585
586xnnpack_cc_library(
587 name = "psimd_ukernels",
588 srcs = PSIMD_UKERNELS,
589 hdrs = INTERNAL_HDRS,
590 aarch32_copts = [
591 "-marm",
592 "-mfpu=neon",
593 ],
594 copts = xnnpack_std_copts(),
595 optimized_copts = [
596 "-O3",
597 "-ffast-math",
598 ],
599 deps = [
600 "@FP16",
601 "@psimd",
602 ],
603)
604
605xnnpack_cc_library(
606 name = "neon_ukernels",
607 hdrs = INTERNAL_HDRS,
608 aarch32_copts = [
609 "-marm",
610 "-mfpu=neon",
611 ],
612 aarch32_srcs = NEON_UKERNELS,
613 aarch64_srcs = NEON_UKERNELS,
614 copts = xnnpack_std_copts(),
615 deps = ["@FP16"],
616)
617
618xnnpack_cc_library(
619 name = "neonfma_ukernels",
620 hdrs = INTERNAL_HDRS,
621 aarch32_copts = [
622 "-marm",
623 "-mfpu=neon-vfpv4",
624 ],
625 aarch32_srcs = NEONFMA_UKERNELS,
626 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
627 copts = xnnpack_std_copts(),
628 deps = ["@FP16"],
629)
630
631xnnpack_cc_library(
632 name = "neonfp16arith_ukernels",
633 hdrs = INTERNAL_HDRS,
634 aarch64_copts = ["-march=armv8.2-a+fp16"],
635 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
636 copts = xnnpack_std_copts(),
637 deps = ["@FP16"],
638)
639
640xnnpack_cc_library(
641 name = "sse2_ukernels",
642 hdrs = INTERNAL_HDRS,
643 copts = xnnpack_std_copts(),
644 x86_copts = ["-msse2"],
645 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
646 deps = ["@FP16"],
647)
648
649xnnpack_cc_library(
650 name = "avx_ukernels",
651 hdrs = INTERNAL_HDRS,
652 copts = xnnpack_std_copts(),
653 x86_copts = ["-mavx"],
654 x86_srcs = AVX_UKERNELS,
655 deps = ["@FP16"],
656)
657
658xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700659 name = "avx2_ukernels",
660 hdrs = INTERNAL_HDRS,
661 copts = xnnpack_std_copts(),
662 x86_copts = [
663 "-mfma",
664 "-mavx2",
665 ],
666 x86_srcs = AVX2_UKERNELS,
667 deps = ["@FP16"],
668)
669
670xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700671 name = "avx512f_ukernels",
672 hdrs = INTERNAL_HDRS,
673 copts = xnnpack_std_copts(),
674 x86_copts = ["-mavx512f"],
675 x86_srcs = AVX512F_UKERNELS,
676 deps = ["@FP16"],
677)
678
679xnnpack_cc_library(
680 name = "asm_ukernels",
681 hdrs = ["src/xnnpack/assembly.h"],
682 aarch32_srcs = AARCH32_ASM_UKERNELS,
683 aarch64_srcs = AARCH64_ASM_UKERNELS,
684)
685
686xnnpack_aggregate_library(
687 name = "ukernels",
688 aarch32_deps = [
689 ":psimd_ukernels",
690 ":neon_ukernels",
691 ":neonfma_ukernels",
692 ":asm_ukernels",
693 ],
694 aarch64_deps = [
695 ":psimd_ukernels",
696 ":neon_ukernels",
697 ":neonfma_ukernels",
698 ":neonfp16arith_ukernels",
699 ":asm_ukernels",
700 ],
701 generic_deps = [":scalar_ukernels"],
702 wasmsimd_deps = [
703 ":psimd_ukernels",
704 ],
705 x86_deps = [
706 ":psimd_ukernels",
707 ":sse2_ukernels",
708 ":avx_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700709 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700710 ":avx512f_ukernels",
711 ],
712)
713
714xnnpack_cc_library(
715 name = "im2col",
716 srcs = ["src/im2col.c"],
717 hdrs = [
718 "src/xnnpack/common.h",
719 "src/xnnpack/im2col.h",
720 ],
721 copts = xnnpack_std_copts(),
722)
723
724xnnpack_cc_library(
725 name = "indirection",
726 srcs = ["src/indirection.c"],
727 hdrs = INTERNAL_HDRS,
728 copts = xnnpack_std_copts(),
729 deps = [
730 "@FP16",
731 "@FXdiv",
732 "@pthreadpool",
733 ],
734)
735
736xnnpack_cc_library(
737 name = "operator_run",
738 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700739 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
740 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700741 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
742 "-Wno-vla",
743 ],
744 deps = [
745 "@FP16",
746 "@FXdiv",
747 "@clog",
748 "@pthreadpool",
749 ],
750)
751
752cc_library(
753 name = "enable_assembly",
754 defines = select({
755 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
756 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -0700757 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700758 }),
759)
760
Marat Dukhancf056b22019-10-07 10:26:29 -0700761xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700762 name = "operators",
763 srcs = OPERATOR_SRCS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700764 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -0700765 ],
766 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700767 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700768 "-Isrc",
769 "-Iinclude",
770 ] + select({
771 ":debug_build": [],
772 "//conditions:default": xnnpack_min_size_copts(),
773 }),
Marat Dukhancf056b22019-10-07 10:26:29 -0700774 wasm_srcs = ["src/wasm-stubs.c"],
775 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700776 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700777 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700778 "@FP16",
779 "@FXdiv",
780 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700781 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700782 ],
783)
784
785cc_library(
786 name = "XNNPACK",
787 srcs = [
788 "src/init.c",
789 ],
790 copts = xnnpack_std_copts() + LOGGING_COPTS + [
791 "-Isrc",
792 "-Iinclude",
793 ] + select({
794 ":debug_build": [],
795 "//conditions:default": xnnpack_min_size_copts(),
796 }),
797 includes = ["include"],
798 linkstatic = True,
799 textual_hdrs = ["include/xnnpack.h"],
800 visibility = xnnpack_visibility(),
801 deps = [
802 ":enable_assembly",
803 ":ukernels",
804 ":operator_run",
805 ":operators",
806 "@clog",
807 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -0700808 ] + select({
809 ":emscripten": [],
810 "//conditions:default": ["@cpuinfo"],
811 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700812)
813
814cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700815 name = "xnnpack_operators_nhwc_f32",
816 srcs = [
817 "src/init.c",
818 ],
819 copts = xnnpack_std_copts() + LOGGING_COPTS + [
820 "-Isrc",
821 "-Iinclude",
822 ] + select({
823 ":debug_build": [],
824 "//conditions:default": xnnpack_min_size_copts(),
825 }),
826 defines = [
827 "XNN_NO_Q8_OPERATORS",
828 "XNN_NO_U8_OPERATORS",
829 "XNN_NO_X8_OPERATORS",
830 "XNN_NO_SPNCHW_OPERATORS",
831 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700832 includes = ["include"],
833 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700834 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700835 visibility = xnnpack_visibility(),
836 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700837 ":enable_assembly",
838 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700839 ":operator_run",
840 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700841 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700842 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700843 ] + select({
844 ":emscripten": [],
845 "//conditions:default": ["@cpuinfo"],
846 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700847)
848
Marat Dukhancf056b22019-10-07 10:26:29 -0700849xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700850 name = "bench_utils",
851 srcs = ["bench/utils.cc"],
852 hdrs = ["bench/utils.h"],
853 copts = ["-Wno-unused-result"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700854 deps = ["@cpuinfo"],
855)
856
Frank Barchard7e955972019-10-11 10:34:25 -0700857######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -0700858
859xnnpack_benchmark(
860 name = "q8_gemm_bench",
861 srcs = [
862 "bench/gemm.h",
863 "bench/q8-gemm.cc",
864 "src/xnnpack/AlignedAllocator.h",
865 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
866 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
867 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
868)
869
870xnnpack_benchmark(
871 name = "f16_gemm_bench",
872 srcs = [
873 "bench/f16-gemm.cc",
874 "bench/gemm.h",
875 "src/xnnpack/AlignedAllocator.h",
876 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
877 copts = ["-Wno-unused-function"],
878 deps = MICROKERNEL_BENCHMARK_DEPS,
879)
880
881xnnpack_benchmark(
882 name = "f32_igemm_bench",
883 srcs = [
884 "bench/f32-igemm.cc",
885 "bench/conv.h",
886 "src/xnnpack/AlignedAllocator.h",
887 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -0700888 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700889)
890
891xnnpack_benchmark(
892 name = "f32_conv_hwc_bench",
893 srcs = [
894 "bench/f32-conv-hwc.cc",
895 "bench/dconv.h",
896 "src/xnnpack/AlignedAllocator.h",
897 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
898 copts = ["-Wno-unused-function"],
899 deps = MICROKERNEL_BENCHMARK_DEPS,
900)
901
902xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -0700903 name = "f32_conv_hwc2spchw_bench",
904 srcs = [
905 "bench/f32-conv-hwc2spchw.cc",
906 "bench/dconv.h",
907 "src/xnnpack/AlignedAllocator.h",
908 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
909 copts = ["-Wno-unused-function"],
910 deps = MICROKERNEL_BENCHMARK_DEPS,
911)
912
913xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700914 name = "f32_dwconv_bench",
915 srcs = [
916 "bench/f32-dwconv.cc",
917 "bench/dwconv.h",
918 "src/xnnpack/AlignedAllocator.h",
919 ] + MICROKERNEL_BENCHMARK_HDRS,
920 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
921)
922
923xnnpack_benchmark(
924 name = "f32_dwconv_spchw_bench",
925 srcs = [
926 "bench/f32-dwconv-spchw.cc",
927 "bench/dwconv.h",
928 "src/xnnpack/AlignedAllocator.h",
929 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
930 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
931)
932
933xnnpack_benchmark(
934 name = "f32_gemm_bench",
935 srcs = [
936 "bench/f32-gemm.cc",
937 "bench/gemm.h",
938 "src/xnnpack/AlignedAllocator.h",
939 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
940 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -0700941 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700942)
943
944xnnpack_benchmark(
945 name = "f32_rmax_bench",
946 srcs = [
947 "bench/f32-rmax.cc",
948 "src/xnnpack/AlignedAllocator.h",
949 ] + MICROKERNEL_BENCHMARK_HDRS,
950 deps = MICROKERNEL_BENCHMARK_DEPS,
951)
952
953xnnpack_benchmark(
954 name = "f32_spmm_bench",
955 srcs = [
956 "bench/f32-spmm.cc",
957 "bench/gemm.h",
958 "src/xnnpack/AlignedAllocator.h",
959 ] + MICROKERNEL_BENCHMARK_HDRS,
960 copts = ["-Wno-unused-function"],
961 deps = MICROKERNEL_BENCHMARK_DEPS,
962)
963
964xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -0700965 name = "f32_softargmax_bench",
966 srcs = [
967 "bench/f32-softargmax.cc",
968 ] + MICROKERNEL_BENCHMARK_HDRS,
969 copts = ["-Wno-unused-function"],
970 deps = MICROKERNEL_BENCHMARK_DEPS,
971)
972
973xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700974 name = "f32_im2col_gemm_bench",
975 srcs = [
976 "bench/f32-im2col-gemm.cc",
977 "bench/conv.h",
978 "src/xnnpack/AlignedAllocator.h",
979 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
980 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
981)
982
983########################### Benchmarks for operators ###########################
984
985xnnpack_benchmark(
986 name = "add_bench",
987 srcs = ["bench/add.cc"],
988 deps = OPERATOR_BENCHMARK_DEPS,
989)
990
991xnnpack_benchmark(
992 name = "average_pooling_bench",
993 srcs = ["bench/average-pooling.cc"],
994 deps = OPERATOR_BENCHMARK_DEPS,
995)
996
997xnnpack_benchmark(
998 name = "channel_shuffle_bench",
999 srcs = ["bench/channel-shuffle.cc"],
1000 deps = OPERATOR_BENCHMARK_DEPS,
1001)
1002
1003xnnpack_benchmark(
1004 name = "convolution_bench",
1005 srcs = ["bench/convolution.cc"],
1006 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1007 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1008)
1009
1010xnnpack_benchmark(
1011 name = "deconvolution_bench",
1012 srcs = ["bench/deconvolution.cc"],
1013 copts = xnnpack_optional_tflite_copts(),
1014 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1015)
1016
1017xnnpack_benchmark(
1018 name = "global_average_pooling_bench",
1019 srcs = ["bench/global-average-pooling.cc"],
1020 deps = OPERATOR_BENCHMARK_DEPS,
1021)
1022
1023xnnpack_benchmark(
1024 name = "max_pooling_bench",
1025 srcs = ["bench/max-pooling.cc"],
1026 deps = OPERATOR_BENCHMARK_DEPS,
1027)
1028
1029xnnpack_benchmark(
1030 name = "sigmoid_bench",
1031 srcs = ["bench/sigmoid.cc"],
1032 deps = OPERATOR_BENCHMARK_DEPS,
1033)
1034
1035xnnpack_benchmark(
1036 name = "softargmax_bench",
1037 srcs = ["bench/softargmax.cc"],
1038 deps = OPERATOR_BENCHMARK_DEPS,
1039)
1040
Marat Dukhanc068bb62019-10-04 13:24:39 -07001041############################# End-to-end benchmarks ############################
1042
1043cc_library(
1044 name = "mobilenet_v1",
1045 srcs = ["models/mobilenet-v1.cc"],
1046 hdrs = ["models/models.h"],
1047 linkstatic = True,
1048 deps = [
1049 ":XNNPACK",
1050 "@pthreadpool",
1051 ],
1052)
1053
1054cc_library(
1055 name = "mobilenet_v2",
1056 srcs = ["models/mobilenet-v2.cc"],
1057 hdrs = ["models/models.h"],
1058 linkstatic = True,
1059 deps = [
1060 ":XNNPACK",
1061 "@pthreadpool",
1062 ],
1063)
1064
1065xnnpack_benchmark(
1066 name = "end2end_bench",
1067 srcs = ["bench/end2end.cc"],
1068 deps = [
1069 ":XNNPACK",
1070 ":mobilenet_v1",
1071 ":mobilenet_v2",
1072 "@pthreadpool",
1073 ],
1074)
1075
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001076#################### Accuracy evaluation for math functions ####################
1077
1078xnnpack_benchmark(
1079 name = "f32_exp_eval",
1080 srcs = [
1081 "eval/f32-exp.cc",
1082 "src/xnnpack/AlignedAllocator.h",
1083 ] + ACCURACY_EVAL_HDRS,
1084 deps = ACCURACY_EVAL_DEPS,
1085)
1086
Marat Dukhan515c9772019-10-17 18:07:57 -07001087xnnpack_benchmark(
1088 name = "f32_expminus_eval",
1089 srcs = [
1090 "eval/f32-expminus.cc",
1091 "src/xnnpack/AlignedAllocator.h",
1092 ] + ACCURACY_EVAL_HDRS,
1093 deps = ACCURACY_EVAL_DEPS,
1094)
1095
Marat Dukhan98ba4412019-10-23 02:14:28 -07001096xnnpack_benchmark(
1097 name = "f32_extexp_eval",
1098 srcs = [
1099 "eval/f32-extexp.cc",
1100 "src/xnnpack/AlignedAllocator.h",
1101 ] + ACCURACY_EVAL_HDRS,
1102 deps = ACCURACY_EVAL_DEPS,
1103)
1104
Marat Dukhan08c4a432019-10-03 09:29:21 -07001105######################### Unit tests for micro-kernels #########################
1106
1107xnnpack_unit_test(
1108 name = "f16_gemm_test",
1109 srcs = [
1110 "test/f16-gemm.cc",
1111 "test/gemm-microkernel-tester.h",
1112 "src/xnnpack/AlignedAllocator.h",
1113 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1114 deps = MICROKERNEL_TEST_DEPS,
1115)
1116
1117xnnpack_unit_test(
1118 name = "f32_argmaxpool_test",
1119 srcs = [
1120 "test/f32-argmaxpool.cc",
1121 "test/argmaxpool-microkernel-tester.h",
1122 "src/xnnpack/AlignedAllocator.h",
1123 ] + MICROKERNEL_TEST_HDRS,
1124 deps = MICROKERNEL_TEST_DEPS,
1125)
1126
1127xnnpack_unit_test(
1128 name = "f32_avgpool_test",
1129 srcs = [
1130 "test/f32-avgpool.cc",
1131 "test/avgpool-microkernel-tester.h",
1132 "src/xnnpack/AlignedAllocator.h",
1133 ] + MICROKERNEL_TEST_HDRS,
1134 deps = MICROKERNEL_TEST_DEPS,
1135)
1136
1137xnnpack_unit_test(
1138 name = "f32_clamp_test",
1139 srcs = [
1140 "test/f32-clamp.cc",
1141 "test/clamp-microkernel-tester.h",
1142 ] + MICROKERNEL_TEST_HDRS,
1143 deps = MICROKERNEL_TEST_DEPS,
1144)
1145
1146xnnpack_unit_test(
1147 name = "f32_igemm_test",
1148 srcs = [
1149 "test/f32-igemm.cc",
1150 "test/gemm-microkernel-tester.h",
1151 "src/xnnpack/AlignedAllocator.h",
1152 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1153 deps = MICROKERNEL_TEST_DEPS,
1154)
1155
1156xnnpack_unit_test(
1157 name = "f32_conv_hwc_test",
1158 srcs = [
1159 "test/f32-conv-hwc.cc",
1160 "test/conv-hwc-microkernel-tester.h",
1161 "src/xnnpack/AlignedAllocator.h",
1162 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1163 deps = MICROKERNEL_TEST_DEPS,
1164)
1165
1166xnnpack_unit_test(
1167 name = "f32_conv_hwc2spchw_test",
1168 srcs = [
1169 "test/f32-conv-hwc2spchw.cc",
1170 "test/conv-hwc2spchw-microkernel-tester.h",
1171 "src/xnnpack/AlignedAllocator.h",
1172 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1173 deps = MICROKERNEL_TEST_DEPS,
1174)
1175
1176xnnpack_unit_test(
1177 name = "f32_dwconv_test",
1178 srcs = [
1179 "test/f32-dwconv.cc",
1180 "test/dwconv-microkernel-tester.h",
1181 "src/xnnpack/AlignedAllocator.h",
1182 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1183 deps = MICROKERNEL_TEST_DEPS,
1184)
1185
1186xnnpack_unit_test(
1187 name = "f32_dwconv_spchw_test",
1188 srcs = [
1189 "test/f32-dwconv-spchw.cc",
1190 "test/dwconv-spchw-microkernel-tester.h",
1191 "src/xnnpack/AlignedAllocator.h",
1192 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1193 deps = MICROKERNEL_TEST_DEPS,
1194)
1195
1196xnnpack_unit_test(
1197 name = "f32_gavgpool_test",
1198 srcs = [
1199 "test/f32-gavgpool.cc",
1200 "test/gavgpool-microkernel-tester.h",
1201 "src/xnnpack/AlignedAllocator.h",
1202 ] + MICROKERNEL_TEST_HDRS,
1203 deps = MICROKERNEL_TEST_DEPS,
1204)
1205
1206xnnpack_unit_test(
1207 name = "f32_gavgpool_spchw_test",
1208 srcs = [
1209 "test/f32-gavgpool-spchw.cc",
1210 "test/gavgpool-spchw-microkernel-tester.h",
1211 "src/xnnpack/AlignedAllocator.h",
1212 ] + MICROKERNEL_TEST_HDRS,
1213 deps = MICROKERNEL_TEST_DEPS,
1214)
1215
1216xnnpack_unit_test(
1217 name = "f32_gemm_test",
1218 srcs = [
1219 "test/f32-gemm.cc",
1220 "test/gemm-microkernel-tester.h",
1221 "src/xnnpack/AlignedAllocator.h",
1222 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1223 deps = MICROKERNEL_TEST_DEPS,
1224)
1225
1226xnnpack_unit_test(
1227 name = "f32_gemminc_test",
1228 srcs = [
1229 "test/f32-gemminc.cc",
1230 "test/gemm-microkernel-tester.h",
1231 "src/xnnpack/AlignedAllocator.h",
1232 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1233 deps = MICROKERNEL_TEST_DEPS,
1234)
1235
1236xnnpack_unit_test(
1237 name = "f32_hswish_test",
1238 srcs = [
1239 "test/f32-hswish.cc",
1240 "test/hswish-microkernel-tester.h",
1241 ] + MICROKERNEL_TEST_HDRS,
1242 deps = MICROKERNEL_TEST_DEPS,
1243)
1244
1245xnnpack_unit_test(
1246 name = "f32_maxpool_test",
1247 srcs = [
1248 "test/f32-maxpool.cc",
1249 "test/maxpool-microkernel-tester.h",
1250 ] + MICROKERNEL_TEST_HDRS,
1251 deps = MICROKERNEL_TEST_DEPS,
1252)
1253
1254xnnpack_unit_test(
1255 name = "f32_pavgpool_test",
1256 srcs = [
1257 "test/f32-pavgpool.cc",
1258 "test/avgpool-microkernel-tester.h",
1259 "src/xnnpack/AlignedAllocator.h",
1260 ] + MICROKERNEL_TEST_HDRS,
1261 deps = MICROKERNEL_TEST_DEPS,
1262)
1263
1264xnnpack_unit_test(
1265 name = "f32_ppmm_test",
1266 srcs = [
1267 "test/f32-ppmm.cc",
1268 "test/gemm-microkernel-tester.h",
1269 "src/xnnpack/AlignedAllocator.h",
1270 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1271 deps = MICROKERNEL_TEST_DEPS,
1272)
1273
1274xnnpack_unit_test(
1275 name = "f32_prelu_test",
1276 srcs = [
1277 "test/f32-prelu.cc",
1278 "test/prelu-microkernel-tester.h",
1279 "src/xnnpack/AlignedAllocator.h",
1280 ] + MICROKERNEL_TEST_HDRS,
1281 deps = MICROKERNEL_TEST_DEPS,
1282)
1283
1284xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001285 name = "f32_raddexpminusmax_test",
1286 srcs = [
1287 "test/f32-raddexpminusmax.cc",
1288 "test/raddexpminusmax-microkernel-tester.h",
1289 ] + MICROKERNEL_TEST_HDRS,
1290 deps = MICROKERNEL_TEST_DEPS,
1291)
1292
1293xnnpack_unit_test(
1294 name = "f32_raddstoreexpminusmax_test",
1295 srcs = [
1296 "test/f32-raddstoreexpminusmax.cc",
1297 "test/raddstoreexpminusmax-microkernel-tester.h",
1298 ] + MICROKERNEL_TEST_HDRS,
1299 deps = MICROKERNEL_TEST_DEPS,
1300)
1301
1302xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001303 name = "f32_rmax_test",
1304 srcs = [
1305 "test/f32-rmax.cc",
1306 "test/rmax-microkernel-tester.h",
1307 ] + MICROKERNEL_TEST_HDRS,
1308 deps = MICROKERNEL_TEST_DEPS,
1309)
1310
1311xnnpack_unit_test(
1312 name = "f32_spmm_test",
1313 srcs = [
1314 "test/f32-spmm.cc",
1315 "test/spmm-microkernel-tester.h",
1316 "src/xnnpack/AlignedAllocator.h",
1317 ] + MICROKERNEL_TEST_HDRS,
1318 deps = MICROKERNEL_TEST_DEPS,
1319)
1320
1321xnnpack_unit_test(
1322 name = "f32_vadd_test",
1323 srcs = [
1324 "test/f32-vadd.cc",
1325 "test/vadd-microkernel-tester.h",
1326 ] + MICROKERNEL_TEST_HDRS,
1327 deps = MICROKERNEL_TEST_DEPS,
1328)
1329
1330xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001331 name = "f32_vmul_test",
1332 srcs = [
1333 "test/f32-vmul.cc",
1334 "test/vmul-microkernel-tester.h",
1335 ] + MICROKERNEL_TEST_HDRS,
1336 deps = MICROKERNEL_TEST_DEPS,
1337)
1338
1339xnnpack_unit_test(
1340 name = "f32_vmulcaddc_test",
1341 srcs = [
1342 "test/f32-vmulcaddc.cc",
1343 "test/vmulcaddc-microkernel-tester.h",
1344 "src/xnnpack/AlignedAllocator.h",
1345 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1346 deps = MICROKERNEL_TEST_DEPS,
1347)
1348
1349xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001350 name = "f32_vscale_test",
1351 srcs = [
1352 "test/f32-vscale.cc",
1353 "test/vscale-microkernel-tester.h",
1354 ] + MICROKERNEL_TEST_HDRS,
1355 deps = MICROKERNEL_TEST_DEPS,
1356)
1357
1358xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001359 name = "f32_vscaleexpminusmax_test",
1360 srcs = [
1361 "test/f32-vscaleexpminusmax.cc",
1362 "test/vscaleexpminusmax-microkernel-tester.h",
1363 ] + MICROKERNEL_TEST_HDRS,
1364 deps = MICROKERNEL_TEST_DEPS,
1365)
1366
1367xnnpack_unit_test(
1368 name = "f32_vsub_test",
1369 srcs = [
1370 "test/f32-vsub.cc",
1371 "test/vsub-microkernel-tester.h",
1372 ] + MICROKERNEL_TEST_HDRS,
1373 deps = MICROKERNEL_TEST_DEPS,
1374)
1375
1376xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001377 name = "q8_avgpool_test",
1378 srcs = [
1379 "test/q8-avgpool.cc",
1380 "test/avgpool-microkernel-tester.h",
1381 "src/xnnpack/AlignedAllocator.h",
1382 ] + MICROKERNEL_TEST_HDRS,
1383 deps = MICROKERNEL_TEST_DEPS,
1384)
1385
1386xnnpack_unit_test(
1387 name = "q8_igemm_test",
1388 srcs = [
1389 "test/q8-igemm.cc",
1390 "test/gemm-microkernel-tester.h",
1391 "src/xnnpack/AlignedAllocator.h",
1392 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1393 deps = MICROKERNEL_TEST_DEPS,
1394)
1395
1396xnnpack_unit_test(
1397 name = "q8_dwconv_test",
1398 srcs = [
1399 "test/q8-dwconv.cc",
1400 "test/dwconv-microkernel-tester.h",
1401 "src/xnnpack/AlignedAllocator.h",
1402 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1403 deps = MICROKERNEL_TEST_DEPS,
1404)
1405
1406xnnpack_unit_test(
1407 name = "q8_gavgpool_test",
1408 srcs = [
1409 "test/q8-gavgpool.cc",
1410 "test/gavgpool-microkernel-tester.h",
1411 "src/xnnpack/AlignedAllocator.h",
1412 ] + MICROKERNEL_TEST_HDRS,
1413 deps = MICROKERNEL_TEST_DEPS,
1414)
1415
1416xnnpack_unit_test(
1417 name = "q8_gemm_test",
1418 srcs = [
1419 "test/q8-gemm.cc",
1420 "test/gemm-microkernel-tester.h",
1421 "src/xnnpack/AlignedAllocator.h",
1422 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1423 deps = MICROKERNEL_TEST_DEPS,
1424)
1425
1426xnnpack_unit_test(
1427 name = "q8_vadd_test",
1428 srcs = [
1429 "test/q8-vadd.cc",
1430 "test/vadd-microkernel-tester.h",
1431 ] + MICROKERNEL_TEST_HDRS,
1432 deps = MICROKERNEL_TEST_DEPS,
1433)
1434
1435xnnpack_unit_test(
1436 name = "u8_clamp_test",
1437 srcs = [
1438 "test/u8-clamp.cc",
1439 "test/clamp-microkernel-tester.h",
1440 ] + MICROKERNEL_TEST_HDRS,
1441 deps = MICROKERNEL_TEST_DEPS,
1442)
1443
1444xnnpack_unit_test(
1445 name = "u8_lut32norm_test",
1446 srcs = [
1447 "test/u8-lut32norm.cc",
1448 "test/lut-norm-microkernel-tester.h",
1449 ] + MICROKERNEL_TEST_HDRS,
1450 deps = MICROKERNEL_TEST_DEPS,
1451)
1452
1453xnnpack_unit_test(
1454 name = "u8_maxpool_test",
1455 srcs = [
1456 "test/u8-maxpool.cc",
1457 "test/maxpool-microkernel-tester.h",
1458 ] + MICROKERNEL_TEST_HDRS,
1459 deps = MICROKERNEL_TEST_DEPS,
1460)
1461
1462xnnpack_unit_test(
1463 name = "u8_rmax_test",
1464 srcs = [
1465 "test/u8-rmax.cc",
1466 "test/rmax-microkernel-tester.h",
1467 ] + MICROKERNEL_TEST_HDRS,
1468 deps = MICROKERNEL_TEST_DEPS,
1469)
1470
1471xnnpack_unit_test(
1472 name = "x32_packx_test",
1473 srcs = [
1474 "test/x32-packx.cc",
1475 "test/pack-microkernel-tester.h",
1476 "src/xnnpack/AlignedAllocator.h",
1477 ] + MICROKERNEL_TEST_HDRS,
1478 deps = MICROKERNEL_TEST_DEPS,
1479)
1480
1481xnnpack_unit_test(
1482 name = "x32_pad_test",
1483 srcs = [
1484 "test/x32-pad.cc",
1485 "test/pad-microkernel-tester.h",
1486 ] + MICROKERNEL_TEST_HDRS,
1487 deps = MICROKERNEL_TEST_DEPS,
1488)
1489
1490xnnpack_unit_test(
1491 name = "x32_unpool_test",
1492 srcs = [
1493 "test/x32-unpool.cc",
1494 "test/unpool-microkernel-tester.h",
1495 ] + MICROKERNEL_TEST_HDRS,
1496 deps = MICROKERNEL_TEST_DEPS,
1497)
1498
1499xnnpack_unit_test(
1500 name = "x32_zip_test",
1501 srcs = [
1502 "test/x32-zip.cc",
1503 "test/zip-microkernel-tester.h",
1504 ] + MICROKERNEL_TEST_HDRS,
1505 deps = MICROKERNEL_TEST_DEPS,
1506)
1507
1508xnnpack_unit_test(
1509 name = "x8_lut_test",
1510 srcs = [
1511 "test/x8-lut.cc",
1512 "test/lut-microkernel-tester.h",
1513 ] + MICROKERNEL_TEST_HDRS,
1514 deps = MICROKERNEL_TEST_DEPS,
1515)
1516
1517xnnpack_unit_test(
1518 name = "x8_zip_test",
1519 srcs = [
1520 "test/x8-zip.cc",
1521 "test/zip-microkernel-tester.h",
1522 ] + MICROKERNEL_TEST_HDRS,
1523 deps = MICROKERNEL_TEST_DEPS,
1524)
1525
1526########################### Size test for the library ##########################
1527
1528xnnpack_binary(
1529 name = "size_test",
1530 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001531 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001532)
1533
1534########################### Unit tests for operators ###########################
1535
1536xnnpack_unit_test(
1537 name = "add_test",
1538 srcs = [
1539 "test/add.cc",
1540 "test/add-operator-tester.h",
1541 ],
1542 deps = OPERATOR_TEST_DEPS,
1543)
1544
1545xnnpack_unit_test(
1546 name = "argmax_pooling_test",
1547 srcs = [
1548 "test/argmax-pooling.cc",
1549 "test/argmax-pooling-operator-tester.h",
1550 ] + OPERATOR_TEST_PARAMS_HDRS,
1551 deps = OPERATOR_TEST_DEPS,
1552)
1553
1554xnnpack_unit_test(
1555 name = "average_pooling_test",
1556 srcs = [
1557 "test/average-pooling.cc",
1558 "test/average-pooling-operator-tester.h",
1559 ] + OPERATOR_TEST_PARAMS_HDRS,
1560 deps = OPERATOR_TEST_DEPS,
1561)
1562
1563xnnpack_unit_test(
1564 name = "channel_pad_test",
1565 srcs = [
1566 "test/channel-pad.cc",
1567 "test/channel-pad-operator-tester.h",
1568 ] + OPERATOR_TEST_PARAMS_HDRS,
1569 deps = OPERATOR_TEST_DEPS,
1570)
1571
1572xnnpack_unit_test(
1573 name = "channel_shuffle_test",
1574 srcs = [
1575 "test/channel-shuffle.cc",
1576 "test/channel-shuffle-operator-tester.h",
1577 ],
1578 deps = OPERATOR_TEST_DEPS,
1579)
1580
1581xnnpack_unit_test(
1582 name = "clamp_test",
1583 srcs = [
1584 "test/clamp.cc",
1585 "test/clamp-operator-tester.h",
1586 ],
1587 deps = OPERATOR_TEST_DEPS,
1588)
1589
1590xnnpack_unit_test(
1591 name = "convolution_test",
1592 srcs = [
1593 "test/convolution.cc",
1594 "test/convolution-operator-tester.h",
1595 ],
1596 deps = OPERATOR_TEST_DEPS,
1597)
1598
1599xnnpack_unit_test(
1600 name = "convolution_spnchw_test",
1601 srcs = [
1602 "test/convolution-spnchw.cc",
1603 "test/convolution-spnchw-operator-tester.h",
1604 ],
1605 deps = OPERATOR_TEST_DEPS,
1606)
1607
1608xnnpack_unit_test(
1609 name = "deconvolution_test",
1610 srcs = [
1611 "test/deconvolution.cc",
1612 "test/deconvolution-operator-tester.h",
1613 ] + OPERATOR_TEST_PARAMS_HDRS,
1614 deps = OPERATOR_TEST_DEPS,
1615)
1616
1617xnnpack_unit_test(
1618 name = "fully_connected_test",
1619 srcs = [
1620 "test/fully-connected.cc",
1621 "test/fully-connected-operator-tester.h",
1622 ],
1623 deps = OPERATOR_TEST_DEPS,
1624)
1625
1626xnnpack_unit_test(
1627 name = "global_average_pooling_test",
1628 srcs = [
1629 "test/global-average-pooling.cc",
1630 "test/global-average-pooling-operator-tester.h",
1631 ] + OPERATOR_TEST_PARAMS_HDRS,
1632 deps = OPERATOR_TEST_DEPS,
1633)
1634
1635xnnpack_unit_test(
1636 name = "global_average_pooling_spnchw_test",
1637 srcs = [
1638 "test/global-average-pooling-spnchw.cc",
1639 "test/global-average-pooling-spnchw-operator-tester.h",
1640 ],
1641 deps = OPERATOR_TEST_DEPS,
1642)
1643
1644xnnpack_unit_test(
1645 name = "hardswish_test",
1646 srcs = [
1647 "test/hardswish.cc",
1648 "test/hardswish-operator-tester.h",
1649 ],
1650 deps = OPERATOR_TEST_DEPS,
1651)
1652
1653xnnpack_unit_test(
1654 name = "leaky_relu_test",
1655 srcs = [
1656 "test/leaky-relu.cc",
1657 "test/leaky-relu-operator-tester.h",
1658 ],
1659 deps = OPERATOR_TEST_DEPS,
1660)
1661
1662xnnpack_unit_test(
1663 name = "max_pooling_test",
1664 srcs = [
1665 "test/max-pooling.cc",
1666 "test/max-pooling-operator-tester.h",
1667 ] + OPERATOR_TEST_PARAMS_HDRS,
1668 deps = OPERATOR_TEST_DEPS,
1669)
1670
1671xnnpack_unit_test(
1672 name = "prelu_test",
1673 srcs = [
1674 "test/prelu.cc",
1675 "test/prelu-operator-tester.h",
1676 ] + OPERATOR_TEST_PARAMS_HDRS,
1677 deps = OPERATOR_TEST_DEPS,
1678)
1679
1680xnnpack_unit_test(
1681 name = "sigmoid_test",
1682 srcs = [
1683 "test/sigmoid.cc",
1684 "test/sigmoid-operator-tester.h",
1685 ],
1686 deps = OPERATOR_TEST_DEPS,
1687)
1688
1689xnnpack_unit_test(
1690 name = "softargmax_test",
1691 srcs = [
1692 "test/softargmax.cc",
1693 "test/softargmax-operator-tester.h",
1694 ],
1695 deps = OPERATOR_TEST_DEPS,
1696)
1697
1698xnnpack_unit_test(
1699 name = "unpooling_test",
1700 srcs = [
1701 "test/unpooling.cc",
1702 "test/unpooling-operator-tester.h",
1703 ],
1704 deps = OPERATOR_TEST_DEPS,
1705)
1706
1707############################# Build configurations #############################
1708
1709config_setting(
1710 name = "linux_k8",
1711 values = {
1712 "cpu": "k8",
1713 },
1714)
1715
1716config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07001717 name = "linux_aarch64",
1718 values = {
1719 "cpu": "aarch64",
1720 },
1721)
1722
1723config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001724 name = "android",
1725 values = {"crosstool_top": "//external:android/crosstool"},
1726)
1727
1728config_setting(
1729 name = "android_armv7",
1730 values = {
1731 "crosstool_top": "//external:android/crosstool",
1732 "cpu": "armeabi-v7a",
1733 },
1734)
1735
1736config_setting(
1737 name = "android_arm64",
1738 values = {
1739 "crosstool_top": "//external:android/crosstool",
1740 "cpu": "arm64-v8a",
1741 },
1742)
1743
1744config_setting(
1745 name = "android_x86",
1746 values = {
1747 "crosstool_top": "//external:android/crosstool",
1748 "cpu": "x86",
1749 },
1750)
1751
1752config_setting(
1753 name = "android_x86_64",
1754 values = {
1755 "crosstool_top": "//external:android/crosstool",
1756 "cpu": "x86_64",
1757 },
1758)
1759
1760config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07001761 name = "macos_x86_64",
1762 values = {
1763 "apple_platform_type": "macos",
1764 "cpu": "darwin",
1765 },
1766)
1767
1768config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001769 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001770 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07001771)
1772
1773config_setting(
1774 name = "emscripten_wasm",
1775 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001776 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001777 "cpu": "wasm",
1778 },
1779)
1780
1781config_setting(
1782 name = "emscripten_wasmsimd",
1783 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001784 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001785 "cpu": "wasm",
1786 "features": "wasmsimd",
1787 },
1788)
1789
1790config_setting(
1791 name = "emscripten_asmjs",
1792 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001793 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001794 "cpu": "asmjs",
1795 },
1796)
1797
Marat Dukhan174706e2019-10-09 16:33:53 -07001798# Builds with -c dbg
Marat Dukhan08c4a432019-10-03 09:29:21 -07001799config_setting(
1800 name = "debug_build",
1801 values = {
1802 "compilation_mode": "dbg",
1803 },
Marat Dukhan08c4a432019-10-03 09:29:21 -07001804)
1805
Marat Dukhan174706e2019-10-09 16:33:53 -07001806# Builds with -c opt
Marat Dukhan08c4a432019-10-03 09:29:21 -07001807config_setting(
1808 name = "optimized_build",
1809 values = {
1810 "compilation_mode": "opt",
1811 },
Marat Dukhan08c4a432019-10-03 09:29:21 -07001812)
1813
1814# Enables usage of assembly kernels.
1815config_setting(
1816 name = "xnn_enable_assembly_explicit_true",
1817 define_values = {"xnn_enable_assembly": "true"},
1818)
1819
1820# Disables usage of assembly kernels.
1821config_setting(
1822 name = "xnn_enable_assembly_explicit_false",
1823 define_values = {"xnn_enable_assembly": "false"},
1824)