blob: 49d8dba851cf00e4032240e0a14ec4250fd67cbb [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",
81 "src/f32-igemm/1x4-scalar.c",
82 "src/f32-igemm/2x4-scalar.c",
83 "src/f32-igemm/4x2-scalar.c",
84 "src/f32-igemm/4x4-scalar.c",
85 "src/f32-dwconv/up1x25-scalar.c",
86 "src/f32-dwconv/up1x4-scalar.c",
87 "src/f32-dwconv/up1x9-scalar.c",
Erich Elsen0cc2c532019-10-15 04:44:18 -070088 "src/f32-dwconv-spchw/3x3p1-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -070089 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -070090 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070091 "src/f32-gavgpool/mp7p7q-scalar.c",
92 "src/f32-gavgpool/up7-scalar.c",
93 "src/f32-gemm/1x4-scalar.c",
94 "src/f32-gemm/2x4-scalar.c",
95 "src/f32-gemm/4x2-scalar.c",
96 "src/f32-gemm/4x4-scalar.c",
97 "src/f32-gemminc/1x4-scalar.c",
98 "src/f32-gemminc/2x4-scalar.c",
99 "src/f32-gemminc/4x4-scalar.c",
100 "src/f32-hswish/scalar.c",
101 "src/f32-maxpool/9p8q-scalar.c",
102 "src/f32-pavgpool/mp9p8q-scalar.c",
103 "src/f32-pavgpool/up9-scalar.c",
104 "src/f32-ppmm/2x4-scalar.c",
105 "src/f32-ppmm/3x3-scalar.c",
106 "src/f32-ppmm/4x2-scalar.c",
107 "src/f32-ppmm/4x4-scalar.c",
108 "src/f32-prelu/x4-scalar.c",
109 "src/f32-rmax/scalar.c",
110 "src/f32-spmm/1x1-scalar-pipelined.c",
111 "src/f32-spmm/1x1-scalar-unroll2.c",
112 "src/f32-spmm/1x1-scalar.c",
113 "src/f32-spmm/2x1-scalar-pipelined.c",
114 "src/f32-spmm/2x1-scalar-unroll2.c",
115 "src/f32-spmm/2x1-scalar.c",
116 "src/f32-spmm/4x1-scalar-pipelined.c",
117 "src/f32-spmm/4x1-scalar-unroll2.c",
118 "src/f32-spmm/4x1-scalar.c",
119 "src/f32-spmm/8x1-scalar-pipelined.c",
120 "src/f32-spmm/8x1-scalar-unroll2.c",
121 "src/f32-spmm/8x1-scalar.c",
122 "src/f32-vadd/scalar.c",
123 "src/f32-vmul/scalar.c",
124 "src/f32-vmulcaddc/c1-scalar-x2.c",
125 "src/f32-vsub/scalar.c",
126 "src/q8-avgpool/mp9p8q-scalar.c",
127 "src/q8-avgpool/up9-scalar.c",
128 "src/q8-igemm/2x2-scalar.c",
129 "src/q8-dwconv/up1x9-scalar.c",
130 "src/q8-gavgpool/mp7p7q-scalar.c",
131 "src/q8-gavgpool/up7-scalar.c",
132 "src/q8-gemm/2x2-scalar.c",
133 "src/q8-vadd/scalar.c",
134 "src/u8-clamp/scalar.c",
135 "src/u8-lut32norm/scalar.c",
136 "src/u8-maxpool/9p8q-scalar.c",
137 "src/u8-rmax/scalar.c",
138 "src/x32-packx/x2-scalar.c",
139 "src/x32-packx/x3-scalar.c",
140 "src/x32-packx/x4-scalar.c",
141 "src/x32-pad/x2-scalar.c",
142 "src/x32-unpool/scalar.c",
143 "src/x32-zip/x2-scalar.c",
144 "src/x32-zip/x3-scalar.c",
145 "src/x32-zip/x4-scalar.c",
146 "src/x32-zip/xm-scalar.c",
147 "src/x8-lut/scalar.c",
148 "src/x8-zip/x2-scalar.c",
149 "src/x8-zip/x3-scalar.c",
150 "src/x8-zip/x4-scalar.c",
151 "src/x8-zip/xm-scalar.c",
152]
153
154PSIMD_UKERNELS = [
155 "src/f32-argmaxpool/mp9p8q-psimd.c",
156 "src/f32-argmaxpool/up4-psimd.c",
157 "src/f32-argmaxpool/up9-psimd.c",
158 "src/f32-avgpool/mp9p8q-psimd.c",
159 "src/f32-avgpool/up9-psimd.c",
160 "src/f32-clamp/psimd.c",
161 "src/f32-igemm/1x8-psimd-loadsplat.c",
162 "src/f32-igemm/1x8-psimd-splat.c",
163 "src/f32-igemm/1x8s4-psimd.c",
164 "src/f32-igemm/4x2c4-psimd.c",
165 "src/f32-igemm/4x8-psimd-loadsplat.c",
166 "src/f32-igemm/4x8-psimd-splat.c",
167 "src/f32-igemm/4x8s4-psimd.c",
168 "src/f32-igemm/6x8-psimd-loadsplat.c",
169 "src/f32-igemm/6x8-psimd-splat.c",
170 "src/f32-igemm/6x8s4-psimd.c",
171 "src/f32-dwconv/up4x25-psimd.c",
172 "src/f32-dwconv/up4x4-psimd.c",
173 "src/f32-dwconv/up4x9-psimd.c",
174 "src/f32-gavgpool/mp7p7q-psimd.c",
175 "src/f32-gavgpool/up7-psimd.c",
176 "src/f32-gemm/1x8-psimd-loadsplat.c",
177 "src/f32-gemm/1x8-psimd-splat.c",
178 "src/f32-gemm/1x8s4-psimd.c",
179 "src/f32-gemm/4x8-psimd-loadsplat.c",
180 "src/f32-gemm/4x8-psimd-splat.c",
181 "src/f32-gemm/4x8s4-psimd.c",
182 "src/f32-gemm/6x8-psimd-loadsplat.c",
183 "src/f32-gemm/6x8-psimd-splat.c",
184 "src/f32-gemm/6x8s4-psimd.c",
185 "src/f32-gemminc/1x8-psimd-loadsplat.c",
186 "src/f32-gemminc/1x8-psimd-splat.c",
187 "src/f32-gemminc/1x8s4-psimd.c",
188 "src/f32-gemminc/4x8-psimd-loadsplat.c",
189 "src/f32-gemminc/4x8-psimd-splat.c",
190 "src/f32-gemminc/4x8s4-psimd.c",
191 "src/f32-gemminc/6x8-psimd-loadsplat.c",
192 "src/f32-gemminc/6x8-psimd-splat.c",
193 "src/f32-gemminc/6x8s4-psimd.c",
194 "src/f32-hswish/psimd.c",
195 "src/f32-maxpool/9p8q-psimd.c",
196 "src/f32-pavgpool/mp9p8q-psimd.c",
197 "src/f32-pavgpool/up9-psimd.c",
198 "src/f32-ppmm/4x8-psimd.c",
199 "src/f32-prelu/x4-psimd.c",
200 "src/f32-vadd/psimd.c",
201 "src/f32-vmul/psimd.c",
202 "src/f32-vmulcaddc/c4-psimd-x2.c",
203 "src/f32-vsub/psimd.c",
204 "src/x32-packx/x4-psimd.c",
205 "src/x32-pad/x2-psimd.c",
206 "src/x32-unpool/psimd.c",
207 "src/x32-zip/x2-psimd.c",
208 "src/x32-zip/x3-psimd.c",
209 "src/x32-zip/x4-psimd.c",
210 "src/x32-zip/xm-psimd.c",
211]
212
213# ISA-specific micro-kernels
214NEON_UKERNELS = [
215 "src/f32-avgpool/mp9p8q-neon.c",
216 "src/f32-avgpool/up9-neon.c",
217 "src/f32-clamp/neon.c",
218 "src/f32-igemm/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700219 "src/f32-igemm/4x2-neon-ld64.c",
220 "src/f32-igemm/4x4-neon-ld64.c",
221 "src/f32-igemm/4x8-neon-ld128.c",
222 "src/f32-igemm/4x8-neon-ld64.c",
223 "src/f32-igemm/6x8-neon-ld64.c",
224 "src/f32-dwconv/up4x9-neon.c",
225 "src/f32-gavgpool-spchw/neon-x4.c",
226 "src/f32-gavgpool/mp7p7q-neon.c",
227 "src/f32-gavgpool/up7-neon.c",
228 "src/f32-gemm/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700229 "src/f32-gemm/4x2-neon-ld64.c",
230 "src/f32-gemm/4x8-neon-ld128.c",
231 "src/f32-gemm/4x8-neon-ld64.c",
232 "src/f32-gemm/5x8-neon-ld64.c",
233 "src/f32-gemm/6x8-neon-ld64.c",
234 "src/f32-gemminc/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700235 "src/f32-gemminc/4x8-neon-ld128.c",
236 "src/f32-gemminc/4x8-neon-ld64.c",
237 "src/f32-gemminc/5x8-neon-ld64.c",
238 "src/f32-gemminc/6x8-neon-ld64.c",
239 "src/f32-hswish/neon.c",
240 "src/f32-pavgpool/mp9p8q-neon.c",
241 "src/f32-pavgpool/up9-neon.c",
242 "src/f32-ppmm/4x8-neon.c",
243 "src/f32-ppmm/8x8-neon.c",
244 "src/f32-rmax/neon.c",
245 "src/f32-vmulcaddc/c4-neon-x2.c",
246 "src/q8-avgpool/mp9p8q-neon.c",
247 "src/q8-avgpool/up9-neon.c",
248 "src/q8-igemm/4x8-neon.c",
249 "src/q8-igemm/8x8-neon.c",
250 "src/q8-dwconv/up8x9-neon.c",
251 "src/q8-gavgpool/mp7p7q-neon.c",
252 "src/q8-gavgpool/up7-neon.c",
253 "src/q8-gemm/4x8-neon.c",
254 "src/q8-gemm/8x8-neon.c",
255 "src/q8-vadd/neon.c",
256 "src/u8-clamp/neon.c",
257 "src/u8-maxpool/9p8q-neon.c",
258 "src/u8-rmax/neon.c",
259 "src/x32-packx/x4-neon-st4.c",
260 "src/x32-pad/x2-neon.c",
261 "src/x32-zip/x2-neon.c",
262 "src/x32-zip/x3-neon.c",
263 "src/x32-zip/x4-neon.c",
264 "src/x32-zip/xm-neon.c",
265 "src/x8-zip/x2-neon.c",
266 "src/x8-zip/x3-neon.c",
267 "src/x8-zip/x4-neon.c",
268 "src/x8-zip/xm-neon.c",
269]
270
271NEONFMA_UKERNELS = [
Frank Barcharddb45b6a2019-10-09 16:42:45 -0700272 "src/f32-igemm/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700273 "src/f32-igemm/4x2-neonfma-ld64.c",
274 "src/f32-igemm/4x4-neonfma-ld64.c",
275 "src/f32-igemm/4x8-neonfma-ld128.c",
276 "src/f32-igemm/4x8-neonfma-ld64.c",
277 "src/f32-igemm/6x8-neonfma-ld64.c",
278 "src/f32-dwconv/up4x9-neonfma.c",
279 "src/f32-dwconv/up8x9-neonfma.c",
280 "src/f32-gemm/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700281 "src/f32-gemm/4x2-neonfma-ld64.c",
282 "src/f32-gemm/4x8-neonfma-ld128.c",
283 "src/f32-gemm/4x8-neonfma-ld64.c",
284 "src/f32-gemm/5x8-neonfma-ld64.c",
285 "src/f32-gemm/6x8-neonfma-ld64.c",
286 "src/f32-gemminc/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700287 "src/f32-gemminc/4x8-neonfma-ld128.c",
288 "src/f32-gemminc/4x8-neonfma-ld64.c",
289 "src/f32-gemminc/5x8-neonfma-ld64.c",
290 "src/f32-gemminc/6x8-neonfma-ld64.c",
291 "src/f32-hswish/neonfma.c",
292 "src/f32-ppmm/4x8-neonfma.c",
293 "src/f32-ppmm/8x8-neonfma.c",
294 "src/f32-vmulcaddc/c4-neonfma-x2.c",
295]
296
297AARCH64_NEONFMA_UKERNELS = [
298 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
299 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
300 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
301 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
302 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
303 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
304 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
305 "src/f32-spmm/12x1-neonfma.c",
306 "src/f32-spmm/12x2-neonfma.c",
307 "src/f32-spmm/12x4-neonfma.c",
308 "src/f32-spmm/16x1-neonfma-pipelined.c",
309 "src/f32-spmm/16x1-neonfma-unroll2.c",
310 "src/f32-spmm/16x1-neonfma.c",
311 "src/f32-spmm/16x2-neonfma.c",
312 "src/f32-spmm/16x4-neonfma.c",
313 "src/f32-spmm/4x1-neonfma-pipelined.c",
314 "src/f32-spmm/4x1-neonfma-unroll2.c",
315 "src/f32-spmm/4x1-neonfma.c",
316 "src/f32-spmm/4x2-neonfma.c",
317 "src/f32-spmm/4x4-neonfma.c",
318 "src/f32-spmm/8x1-neonfma-pipelined.c",
319 "src/f32-spmm/8x1-neonfma-unroll2.c",
320 "src/f32-spmm/8x1-neonfma.c",
321 "src/f32-spmm/8x2-neonfma.c",
322 "src/f32-spmm/8x4-neonfma.c",
323]
324
325AARCH64_NEONFP16ARITH_UKERNELS = [
326 "src/f16-gemm/4x8-neonfp16arith-ld64.c",
327 "src/f16-gemm/6x8-neonfp16arith-ld64.c",
328 "src/f16-gemm/8x8-neonfp16arith-ld64.c",
329]
330
331SSE_UKERNELS = [
332 "src/f32-avgpool/mp9p8q-sse.c",
333 "src/f32-avgpool/up9-sse.c",
334 "src/f32-clamp/sse.c",
335 "src/f32-igemm/1x8-sse-dup.c",
336 "src/f32-igemm/1x8-sse-load1.c",
337 "src/f32-igemm/1x8s4-sse.c",
338 "src/f32-igemm/4x2c4-sse.c",
339 "src/f32-igemm/4x8-sse-dup.c",
340 "src/f32-igemm/4x8-sse-load1.c",
341 "src/f32-igemm/4x8s4-sse.c",
342 "src/f32-dwconv/up4x25-sse.c",
343 "src/f32-dwconv/up4x4-sse.c",
344 "src/f32-dwconv/up4x9-sse.c",
345 "src/f32-gavgpool-spchw/sse-x4.c",
346 "src/f32-gavgpool/mp7p7q-sse.c",
347 "src/f32-gavgpool/up7-sse.c",
348 "src/f32-gemm/1x8-sse-dup.c",
349 "src/f32-gemm/1x8-sse-load1.c",
350 "src/f32-gemm/1x8s4-sse.c",
351 "src/f32-gemm/4x8-sse-dup.c",
352 "src/f32-gemm/4x8-sse-load1.c",
353 "src/f32-gemm/4x8s4-sse.c",
354 "src/f32-gemminc/1x8-sse-dup.c",
355 "src/f32-gemminc/1x8-sse-load1.c",
356 "src/f32-gemminc/1x8s4-sse.c",
357 "src/f32-gemminc/4x8-sse-dup.c",
358 "src/f32-gemminc/4x8-sse-load1.c",
359 "src/f32-gemminc/4x8s4-sse.c",
360 "src/f32-hswish/sse.c",
361 "src/f32-maxpool/9p8q-sse.c",
362 "src/f32-pavgpool/mp9p8q-sse.c",
363 "src/f32-pavgpool/up9-sse.c",
364 "src/f32-dwconv-spchw/3x3p1-sse.c",
365 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
366 "src/f32-ppmm/4x8-sse.c",
367 "src/f32-prelu/x4-sse.c",
368 "src/f32-rmax/sse.c",
369 "src/f32-spmm/4x1-sse.c",
370 "src/f32-spmm/8x1-sse.c",
371 "src/f32-vadd/sse.c",
372 "src/f32-vmul/sse.c",
373 "src/f32-vmulcaddc/c4-sse-x2.c",
374 "src/f32-vsub/sse.c",
375 "src/x32-packx/x4-sse.c",
376]
377
378SSE2_UKERNELS = [
379 "src/f32-argmaxpool/mp9p8q-sse2.c",
380 "src/f32-argmaxpool/up4-sse2.c",
381 "src/f32-argmaxpool/up9-sse2.c",
382 "src/q8-avgpool/mp9p8q-sse2.c",
383 "src/q8-avgpool/up9-sse2.c",
384 "src/q8-igemm/4x4c2-sse2.c",
385 "src/q8-dwconv/up8x9-sse2.c",
386 "src/q8-gavgpool/mp7p7q-sse2.c",
387 "src/q8-gavgpool/up7-sse2.c",
388 "src/q8-gemm/2x4c8-sse2.c",
389 "src/q8-gemm/4x4c2-sse2.c",
390 "src/q8-vadd/sse2.c",
391 "src/u8-clamp/sse2.c",
392 "src/u8-maxpool/9p8q-sse2.c",
393 "src/u8-rmax/sse2.c",
394 "src/x32-pad/x2-sse2.c",
395 "src/x32-zip/x2-sse2.c",
396 "src/x32-zip/x3-sse2.c",
397 "src/x32-zip/x4-sse2.c",
398 "src/x32-zip/xm-sse2.c",
399 "src/x8-zip/x2-sse2.c",
400 "src/x8-zip/x3-sse2.c",
401 "src/x8-zip/x4-sse2.c",
402 "src/x8-zip/xm-sse2.c",
403]
404
405AVX_UKERNELS = [
406 "src/f32-rmax/avx.c",
407]
408
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700409AVX2_UKERNELS = [
410 "src/math/exp-avx2-p5.c",
411 "src/math/exp-avx2-perm-p3.c",
412 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -0700413 "src/math/expminus-avx2-p5.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700414]
415
Marat Dukhan08c4a432019-10-03 09:29:21 -0700416AVX512F_UKERNELS = [
417 "src/f32-rmax/avx512f.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700418 "src/math/exp-avx512f-p5-scalef.c",
419 "src/math/exp-avx512f-p5.c",
420 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700421]
422
423AARCH32_ASM_UKERNELS = [
424 "src/q8-dwconv/up8x9-aarch32-neon.S",
425]
426
427AARCH64_ASM_UKERNELS = [
428 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
429 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
430 "src/f32-gemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700431 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700432 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a57.S",
433 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a75.S",
434 "src/f32-gemm/4x12-aarch64-neonfma-cortex-a53.S",
435 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a57.S",
436 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a75.S",
437 "src/f32-gemm/4x8-aarch64-neonfma-ld128.S",
438 "src/f32-gemm/4x8-aarch64-neonfma-ld64.S",
439 "src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S",
440 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a57.S",
441 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S",
442 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S",
443 "src/f32-gemm/6x8-aarch64-neonfma-ld128.S",
444 "src/f32-gemm/6x8-aarch64-neonfma-ld64.S",
445 "src/f32-gemminc/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700446 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700447 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a57.S",
448 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a75.S",
449 "src/f32-gemminc/4x12-aarch64-neonfma-cortex-a53.S",
450 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a57.S",
451 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a75.S",
452 "src/f32-gemminc/4x8-aarch64-neonfma-ld128.S",
453 "src/f32-gemminc/4x8-aarch64-neonfma-ld64.S",
454 "src/f32-gemminc/5x8-aarch64-neonfma-cortex-a75.S",
455 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a57.S",
456 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a73.S",
457 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a75.S",
458 "src/f32-gemminc/6x8-aarch64-neonfma-ld128.S",
459 "src/f32-gemminc/6x8-aarch64-neonfma-ld64.S",
460 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700461 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700462 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S",
463 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S",
464 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
465 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
466 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
467 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
468 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
469 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
470]
471
472INTERNAL_MICROKERNEL_HDRS = [
473 "src/xnnpack/argmaxpool.h",
474 "src/xnnpack/avgpool.h",
475 "src/xnnpack/clamp.h",
476 "src/xnnpack/common.h",
477 "src/xnnpack/conv.h",
478 "src/xnnpack/dwconv.h",
479 "src/xnnpack/gavgpool.h",
480 "src/xnnpack/gemm.h",
481 "src/xnnpack/hswish.h",
482 "src/xnnpack/igemm.h",
483 "src/xnnpack/lut.h",
484 "src/xnnpack/math.h",
485 "src/xnnpack/maxpool.h",
486 "src/xnnpack/packx.h",
487 "src/xnnpack/pad.h",
488 "src/xnnpack/params.h",
489 "src/xnnpack/pavgpool.h",
490 "src/xnnpack/ppmm.h",
491 "src/xnnpack/prelu.h",
492 "src/xnnpack/rmax.h",
493 "src/xnnpack/scalar-utils.h",
494 "src/xnnpack/spmm.h",
495 "src/xnnpack/unpool.h",
496 "src/xnnpack/vadd.h",
497 "src/xnnpack/vmul.h",
498 "src/xnnpack/vmulcaddc.h",
499 "src/xnnpack/vsub.h",
500 "src/xnnpack/zip.h",
501]
502
503INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
504 "include/xnnpack.h",
505 "src/xnnpack/allocator.h",
506 "src/xnnpack/compute.h",
507 "src/xnnpack/im2col.h",
508 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700509 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700510 "src/xnnpack/operator.h",
511 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700512 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700513 "src/xnnpack/requantization.h",
514]
515
516ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
517 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700518]
519
520MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
521 "src/xnnpack/requantization.h",
522 "include/xnnpack.h",
523]
524
525MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
526 "src/xnnpack/isa-checks.h",
527 "src/xnnpack/requantization.h",
528 "include/xnnpack.h",
529]
530
531OPERATOR_TEST_PARAMS_HDRS = [
532 "src/xnnpack/params.h",
533 "src/xnnpack/common.h",
534]
535
536WEIGHTS_PACK_HDRS = [
537 "src/xnnpack/pack.h",
538 "src/xnnpack/operator.h",
539 "src/xnnpack/compute.h",
540]
541
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700542LOGGING_COPTS = select({
543 # No logging in optimized mode
544 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
545 # Full logging in debug mode
546 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
547 # Error-only logging in default (fastbuild) mode
548 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
549})
550
551LOGGING_HDRS = [
552 "src/xnnpack/log.h",
553]
554
Marat Dukhan08c4a432019-10-03 09:29:21 -0700555xnnpack_cc_library(
556 name = "scalar_ukernels",
557 srcs = SCALAR_UKERNELS,
558 hdrs = INTERNAL_HDRS,
559 aarch32_copts = ["-marm"],
560 copts = xnnpack_std_copts(),
561 deps = [
562 "@FP16",
563 "@FXdiv",
564 ],
565)
566
567xnnpack_cc_library(
568 name = "psimd_ukernels",
569 srcs = PSIMD_UKERNELS,
570 hdrs = INTERNAL_HDRS,
571 aarch32_copts = [
572 "-marm",
573 "-mfpu=neon",
574 ],
575 copts = xnnpack_std_copts(),
576 optimized_copts = [
577 "-O3",
578 "-ffast-math",
579 ],
580 deps = [
581 "@FP16",
582 "@psimd",
583 ],
584)
585
586xnnpack_cc_library(
587 name = "neon_ukernels",
588 hdrs = INTERNAL_HDRS,
589 aarch32_copts = [
590 "-marm",
591 "-mfpu=neon",
592 ],
593 aarch32_srcs = NEON_UKERNELS,
594 aarch64_srcs = NEON_UKERNELS,
595 copts = xnnpack_std_copts(),
596 deps = ["@FP16"],
597)
598
599xnnpack_cc_library(
600 name = "neonfma_ukernels",
601 hdrs = INTERNAL_HDRS,
602 aarch32_copts = [
603 "-marm",
604 "-mfpu=neon-vfpv4",
605 ],
606 aarch32_srcs = NEONFMA_UKERNELS,
607 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
608 copts = xnnpack_std_copts(),
609 deps = ["@FP16"],
610)
611
612xnnpack_cc_library(
613 name = "neonfp16arith_ukernels",
614 hdrs = INTERNAL_HDRS,
615 aarch64_copts = ["-march=armv8.2-a+fp16"],
616 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
617 copts = xnnpack_std_copts(),
618 deps = ["@FP16"],
619)
620
621xnnpack_cc_library(
622 name = "sse2_ukernels",
623 hdrs = INTERNAL_HDRS,
624 copts = xnnpack_std_copts(),
625 x86_copts = ["-msse2"],
626 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
627 deps = ["@FP16"],
628)
629
630xnnpack_cc_library(
631 name = "avx_ukernels",
632 hdrs = INTERNAL_HDRS,
633 copts = xnnpack_std_copts(),
634 x86_copts = ["-mavx"],
635 x86_srcs = AVX_UKERNELS,
636 deps = ["@FP16"],
637)
638
639xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700640 name = "avx2_ukernels",
641 hdrs = INTERNAL_HDRS,
642 copts = xnnpack_std_copts(),
643 x86_copts = [
644 "-mfma",
645 "-mavx2",
646 ],
647 x86_srcs = AVX2_UKERNELS,
648 deps = ["@FP16"],
649)
650
651xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700652 name = "avx512f_ukernels",
653 hdrs = INTERNAL_HDRS,
654 copts = xnnpack_std_copts(),
655 x86_copts = ["-mavx512f"],
656 x86_srcs = AVX512F_UKERNELS,
657 deps = ["@FP16"],
658)
659
660xnnpack_cc_library(
661 name = "asm_ukernels",
662 hdrs = ["src/xnnpack/assembly.h"],
663 aarch32_srcs = AARCH32_ASM_UKERNELS,
664 aarch64_srcs = AARCH64_ASM_UKERNELS,
665)
666
667xnnpack_aggregate_library(
668 name = "ukernels",
669 aarch32_deps = [
670 ":psimd_ukernels",
671 ":neon_ukernels",
672 ":neonfma_ukernels",
673 ":asm_ukernels",
674 ],
675 aarch64_deps = [
676 ":psimd_ukernels",
677 ":neon_ukernels",
678 ":neonfma_ukernels",
679 ":neonfp16arith_ukernels",
680 ":asm_ukernels",
681 ],
682 generic_deps = [":scalar_ukernels"],
683 wasmsimd_deps = [
684 ":psimd_ukernels",
685 ],
686 x86_deps = [
687 ":psimd_ukernels",
688 ":sse2_ukernels",
689 ":avx_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700690 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700691 ":avx512f_ukernels",
692 ],
693)
694
695xnnpack_cc_library(
696 name = "im2col",
697 srcs = ["src/im2col.c"],
698 hdrs = [
699 "src/xnnpack/common.h",
700 "src/xnnpack/im2col.h",
701 ],
702 copts = xnnpack_std_copts(),
703)
704
705xnnpack_cc_library(
706 name = "indirection",
707 srcs = ["src/indirection.c"],
708 hdrs = INTERNAL_HDRS,
709 copts = xnnpack_std_copts(),
710 deps = [
711 "@FP16",
712 "@FXdiv",
713 "@pthreadpool",
714 ],
715)
716
717xnnpack_cc_library(
718 name = "operator_run",
719 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700720 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
721 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700722 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
723 "-Wno-vla",
724 ],
725 deps = [
726 "@FP16",
727 "@FXdiv",
728 "@clog",
729 "@pthreadpool",
730 ],
731)
732
733cc_library(
734 name = "enable_assembly",
735 defines = select({
736 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
737 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -0700738 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700739 }),
740)
741
Marat Dukhancf056b22019-10-07 10:26:29 -0700742xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700743 name = "operators",
744 srcs = OPERATOR_SRCS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700745 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -0700746 ],
747 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700748 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700749 "-Isrc",
750 "-Iinclude",
751 ] + select({
752 ":debug_build": [],
753 "//conditions:default": xnnpack_min_size_copts(),
754 }),
Marat Dukhancf056b22019-10-07 10:26:29 -0700755 wasm_srcs = ["src/wasm-stubs.c"],
756 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700757 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700758 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700759 "@FP16",
760 "@FXdiv",
761 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700762 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700763 ],
764)
765
766cc_library(
767 name = "XNNPACK",
768 srcs = [
769 "src/init.c",
770 ],
771 copts = xnnpack_std_copts() + LOGGING_COPTS + [
772 "-Isrc",
773 "-Iinclude",
774 ] + select({
775 ":debug_build": [],
776 "//conditions:default": xnnpack_min_size_copts(),
777 }),
778 includes = ["include"],
779 linkstatic = True,
780 textual_hdrs = ["include/xnnpack.h"],
781 visibility = xnnpack_visibility(),
782 deps = [
783 ":enable_assembly",
784 ":ukernels",
785 ":operator_run",
786 ":operators",
787 "@clog",
788 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -0700789 ] + select({
790 ":emscripten": [],
791 "//conditions:default": ["@cpuinfo"],
792 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700793)
794
795cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700796 name = "xnnpack_operators_nhwc_f32",
797 srcs = [
798 "src/init.c",
799 ],
800 copts = xnnpack_std_copts() + LOGGING_COPTS + [
801 "-Isrc",
802 "-Iinclude",
803 ] + select({
804 ":debug_build": [],
805 "//conditions:default": xnnpack_min_size_copts(),
806 }),
807 defines = [
808 "XNN_NO_Q8_OPERATORS",
809 "XNN_NO_U8_OPERATORS",
810 "XNN_NO_X8_OPERATORS",
811 "XNN_NO_SPNCHW_OPERATORS",
812 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700813 includes = ["include"],
814 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700815 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700816 visibility = xnnpack_visibility(),
817 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700818 ":enable_assembly",
819 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700820 ":operator_run",
821 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700822 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700823 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700824 ] + select({
825 ":emscripten": [],
826 "//conditions:default": ["@cpuinfo"],
827 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700828)
829
Marat Dukhancf056b22019-10-07 10:26:29 -0700830xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700831 name = "bench_utils",
832 srcs = ["bench/utils.cc"],
833 hdrs = ["bench/utils.h"],
834 copts = ["-Wno-unused-result"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700835 deps = ["@cpuinfo"],
836)
837
Frank Barchard7e955972019-10-11 10:34:25 -0700838######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -0700839
840xnnpack_benchmark(
841 name = "q8_gemm_bench",
842 srcs = [
843 "bench/gemm.h",
844 "bench/q8-gemm.cc",
845 "src/xnnpack/AlignedAllocator.h",
846 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
847 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
848 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
849)
850
851xnnpack_benchmark(
852 name = "f16_gemm_bench",
853 srcs = [
854 "bench/f16-gemm.cc",
855 "bench/gemm.h",
856 "src/xnnpack/AlignedAllocator.h",
857 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
858 copts = ["-Wno-unused-function"],
859 deps = MICROKERNEL_BENCHMARK_DEPS,
860)
861
862xnnpack_benchmark(
863 name = "f32_igemm_bench",
864 srcs = [
865 "bench/f32-igemm.cc",
866 "bench/conv.h",
867 "src/xnnpack/AlignedAllocator.h",
868 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -0700869 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700870)
871
872xnnpack_benchmark(
873 name = "f32_conv_hwc_bench",
874 srcs = [
875 "bench/f32-conv-hwc.cc",
876 "bench/dconv.h",
877 "src/xnnpack/AlignedAllocator.h",
878 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
879 copts = ["-Wno-unused-function"],
880 deps = MICROKERNEL_BENCHMARK_DEPS,
881)
882
883xnnpack_benchmark(
884 name = "f32_dwconv_bench",
885 srcs = [
886 "bench/f32-dwconv.cc",
887 "bench/dwconv.h",
888 "src/xnnpack/AlignedAllocator.h",
889 ] + MICROKERNEL_BENCHMARK_HDRS,
890 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
891)
892
893xnnpack_benchmark(
894 name = "f32_dwconv_spchw_bench",
895 srcs = [
896 "bench/f32-dwconv-spchw.cc",
897 "bench/dwconv.h",
898 "src/xnnpack/AlignedAllocator.h",
899 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
900 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
901)
902
903xnnpack_benchmark(
904 name = "f32_gemm_bench",
905 srcs = [
906 "bench/f32-gemm.cc",
907 "bench/gemm.h",
908 "src/xnnpack/AlignedAllocator.h",
909 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
910 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -0700911 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700912)
913
914xnnpack_benchmark(
915 name = "f32_rmax_bench",
916 srcs = [
917 "bench/f32-rmax.cc",
918 "src/xnnpack/AlignedAllocator.h",
919 ] + MICROKERNEL_BENCHMARK_HDRS,
920 deps = MICROKERNEL_BENCHMARK_DEPS,
921)
922
923xnnpack_benchmark(
924 name = "f32_spmm_bench",
925 srcs = [
926 "bench/f32-spmm.cc",
927 "bench/gemm.h",
928 "src/xnnpack/AlignedAllocator.h",
929 ] + MICROKERNEL_BENCHMARK_HDRS,
930 copts = ["-Wno-unused-function"],
931 deps = MICROKERNEL_BENCHMARK_DEPS,
932)
933
934xnnpack_benchmark(
935 name = "f32_im2col_gemm_bench",
936 srcs = [
937 "bench/f32-im2col-gemm.cc",
938 "bench/conv.h",
939 "src/xnnpack/AlignedAllocator.h",
940 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
941 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
942)
943
944########################### Benchmarks for operators ###########################
945
946xnnpack_benchmark(
947 name = "add_bench",
948 srcs = ["bench/add.cc"],
949 deps = OPERATOR_BENCHMARK_DEPS,
950)
951
952xnnpack_benchmark(
953 name = "average_pooling_bench",
954 srcs = ["bench/average-pooling.cc"],
955 deps = OPERATOR_BENCHMARK_DEPS,
956)
957
958xnnpack_benchmark(
959 name = "channel_shuffle_bench",
960 srcs = ["bench/channel-shuffle.cc"],
961 deps = OPERATOR_BENCHMARK_DEPS,
962)
963
964xnnpack_benchmark(
965 name = "convolution_bench",
966 srcs = ["bench/convolution.cc"],
967 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
968 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
969)
970
971xnnpack_benchmark(
972 name = "deconvolution_bench",
973 srcs = ["bench/deconvolution.cc"],
974 copts = xnnpack_optional_tflite_copts(),
975 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
976)
977
978xnnpack_benchmark(
979 name = "global_average_pooling_bench",
980 srcs = ["bench/global-average-pooling.cc"],
981 deps = OPERATOR_BENCHMARK_DEPS,
982)
983
984xnnpack_benchmark(
985 name = "max_pooling_bench",
986 srcs = ["bench/max-pooling.cc"],
987 deps = OPERATOR_BENCHMARK_DEPS,
988)
989
990xnnpack_benchmark(
991 name = "sigmoid_bench",
992 srcs = ["bench/sigmoid.cc"],
993 deps = OPERATOR_BENCHMARK_DEPS,
994)
995
996xnnpack_benchmark(
997 name = "softargmax_bench",
998 srcs = ["bench/softargmax.cc"],
999 deps = OPERATOR_BENCHMARK_DEPS,
1000)
1001
Marat Dukhanc068bb62019-10-04 13:24:39 -07001002############################# End-to-end benchmarks ############################
1003
1004cc_library(
1005 name = "mobilenet_v1",
1006 srcs = ["models/mobilenet-v1.cc"],
1007 hdrs = ["models/models.h"],
1008 linkstatic = True,
1009 deps = [
1010 ":XNNPACK",
1011 "@pthreadpool",
1012 ],
1013)
1014
1015cc_library(
1016 name = "mobilenet_v2",
1017 srcs = ["models/mobilenet-v2.cc"],
1018 hdrs = ["models/models.h"],
1019 linkstatic = True,
1020 deps = [
1021 ":XNNPACK",
1022 "@pthreadpool",
1023 ],
1024)
1025
1026xnnpack_benchmark(
1027 name = "end2end_bench",
1028 srcs = ["bench/end2end.cc"],
1029 deps = [
1030 ":XNNPACK",
1031 ":mobilenet_v1",
1032 ":mobilenet_v2",
1033 "@pthreadpool",
1034 ],
1035)
1036
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001037#################### Accuracy evaluation for math functions ####################
1038
1039xnnpack_benchmark(
1040 name = "f32_exp_eval",
1041 srcs = [
1042 "eval/f32-exp.cc",
1043 "src/xnnpack/AlignedAllocator.h",
1044 ] + ACCURACY_EVAL_HDRS,
1045 deps = ACCURACY_EVAL_DEPS,
1046)
1047
Marat Dukhan515c9772019-10-17 18:07:57 -07001048xnnpack_benchmark(
1049 name = "f32_expminus_eval",
1050 srcs = [
1051 "eval/f32-expminus.cc",
1052 "src/xnnpack/AlignedAllocator.h",
1053 ] + ACCURACY_EVAL_HDRS,
1054 deps = ACCURACY_EVAL_DEPS,
1055)
1056
Marat Dukhan08c4a432019-10-03 09:29:21 -07001057######################### Unit tests for micro-kernels #########################
1058
1059xnnpack_unit_test(
1060 name = "f16_gemm_test",
1061 srcs = [
1062 "test/f16-gemm.cc",
1063 "test/gemm-microkernel-tester.h",
1064 "src/xnnpack/AlignedAllocator.h",
1065 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1066 deps = MICROKERNEL_TEST_DEPS,
1067)
1068
1069xnnpack_unit_test(
1070 name = "f32_argmaxpool_test",
1071 srcs = [
1072 "test/f32-argmaxpool.cc",
1073 "test/argmaxpool-microkernel-tester.h",
1074 "src/xnnpack/AlignedAllocator.h",
1075 ] + MICROKERNEL_TEST_HDRS,
1076 deps = MICROKERNEL_TEST_DEPS,
1077)
1078
1079xnnpack_unit_test(
1080 name = "f32_avgpool_test",
1081 srcs = [
1082 "test/f32-avgpool.cc",
1083 "test/avgpool-microkernel-tester.h",
1084 "src/xnnpack/AlignedAllocator.h",
1085 ] + MICROKERNEL_TEST_HDRS,
1086 deps = MICROKERNEL_TEST_DEPS,
1087)
1088
1089xnnpack_unit_test(
1090 name = "f32_clamp_test",
1091 srcs = [
1092 "test/f32-clamp.cc",
1093 "test/clamp-microkernel-tester.h",
1094 ] + MICROKERNEL_TEST_HDRS,
1095 deps = MICROKERNEL_TEST_DEPS,
1096)
1097
1098xnnpack_unit_test(
1099 name = "f32_igemm_test",
1100 srcs = [
1101 "test/f32-igemm.cc",
1102 "test/gemm-microkernel-tester.h",
1103 "src/xnnpack/AlignedAllocator.h",
1104 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1105 deps = MICROKERNEL_TEST_DEPS,
1106)
1107
1108xnnpack_unit_test(
1109 name = "f32_conv_hwc_test",
1110 srcs = [
1111 "test/f32-conv-hwc.cc",
1112 "test/conv-hwc-microkernel-tester.h",
1113 "src/xnnpack/AlignedAllocator.h",
1114 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1115 deps = MICROKERNEL_TEST_DEPS,
1116)
1117
1118xnnpack_unit_test(
1119 name = "f32_conv_hwc2spchw_test",
1120 srcs = [
1121 "test/f32-conv-hwc2spchw.cc",
1122 "test/conv-hwc2spchw-microkernel-tester.h",
1123 "src/xnnpack/AlignedAllocator.h",
1124 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1125 deps = MICROKERNEL_TEST_DEPS,
1126)
1127
1128xnnpack_unit_test(
1129 name = "f32_dwconv_test",
1130 srcs = [
1131 "test/f32-dwconv.cc",
1132 "test/dwconv-microkernel-tester.h",
1133 "src/xnnpack/AlignedAllocator.h",
1134 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1135 deps = MICROKERNEL_TEST_DEPS,
1136)
1137
1138xnnpack_unit_test(
1139 name = "f32_dwconv_spchw_test",
1140 srcs = [
1141 "test/f32-dwconv-spchw.cc",
1142 "test/dwconv-spchw-microkernel-tester.h",
1143 "src/xnnpack/AlignedAllocator.h",
1144 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1145 deps = MICROKERNEL_TEST_DEPS,
1146)
1147
1148xnnpack_unit_test(
1149 name = "f32_gavgpool_test",
1150 srcs = [
1151 "test/f32-gavgpool.cc",
1152 "test/gavgpool-microkernel-tester.h",
1153 "src/xnnpack/AlignedAllocator.h",
1154 ] + MICROKERNEL_TEST_HDRS,
1155 deps = MICROKERNEL_TEST_DEPS,
1156)
1157
1158xnnpack_unit_test(
1159 name = "f32_gavgpool_spchw_test",
1160 srcs = [
1161 "test/f32-gavgpool-spchw.cc",
1162 "test/gavgpool-spchw-microkernel-tester.h",
1163 "src/xnnpack/AlignedAllocator.h",
1164 ] + MICROKERNEL_TEST_HDRS,
1165 deps = MICROKERNEL_TEST_DEPS,
1166)
1167
1168xnnpack_unit_test(
1169 name = "f32_gemm_test",
1170 srcs = [
1171 "test/f32-gemm.cc",
1172 "test/gemm-microkernel-tester.h",
1173 "src/xnnpack/AlignedAllocator.h",
1174 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1175 deps = MICROKERNEL_TEST_DEPS,
1176)
1177
1178xnnpack_unit_test(
1179 name = "f32_gemminc_test",
1180 srcs = [
1181 "test/f32-gemminc.cc",
1182 "test/gemm-microkernel-tester.h",
1183 "src/xnnpack/AlignedAllocator.h",
1184 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1185 deps = MICROKERNEL_TEST_DEPS,
1186)
1187
1188xnnpack_unit_test(
1189 name = "f32_hswish_test",
1190 srcs = [
1191 "test/f32-hswish.cc",
1192 "test/hswish-microkernel-tester.h",
1193 ] + MICROKERNEL_TEST_HDRS,
1194 deps = MICROKERNEL_TEST_DEPS,
1195)
1196
1197xnnpack_unit_test(
1198 name = "f32_maxpool_test",
1199 srcs = [
1200 "test/f32-maxpool.cc",
1201 "test/maxpool-microkernel-tester.h",
1202 ] + MICROKERNEL_TEST_HDRS,
1203 deps = MICROKERNEL_TEST_DEPS,
1204)
1205
1206xnnpack_unit_test(
1207 name = "f32_pavgpool_test",
1208 srcs = [
1209 "test/f32-pavgpool.cc",
1210 "test/avgpool-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_ppmm_test",
1218 srcs = [
1219 "test/f32-ppmm.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_prelu_test",
1228 srcs = [
1229 "test/f32-prelu.cc",
1230 "test/prelu-microkernel-tester.h",
1231 "src/xnnpack/AlignedAllocator.h",
1232 ] + MICROKERNEL_TEST_HDRS,
1233 deps = MICROKERNEL_TEST_DEPS,
1234)
1235
1236xnnpack_unit_test(
1237 name = "f32_rmax_test",
1238 srcs = [
1239 "test/f32-rmax.cc",
1240 "test/rmax-microkernel-tester.h",
1241 ] + MICROKERNEL_TEST_HDRS,
1242 deps = MICROKERNEL_TEST_DEPS,
1243)
1244
1245xnnpack_unit_test(
1246 name = "f32_spmm_test",
1247 srcs = [
1248 "test/f32-spmm.cc",
1249 "test/spmm-microkernel-tester.h",
1250 "src/xnnpack/AlignedAllocator.h",
1251 ] + MICROKERNEL_TEST_HDRS,
1252 deps = MICROKERNEL_TEST_DEPS,
1253)
1254
1255xnnpack_unit_test(
1256 name = "f32_vadd_test",
1257 srcs = [
1258 "test/f32-vadd.cc",
1259 "test/vadd-microkernel-tester.h",
1260 ] + MICROKERNEL_TEST_HDRS,
1261 deps = MICROKERNEL_TEST_DEPS,
1262)
1263
1264xnnpack_unit_test(
1265 name = "f32_vsub_test",
1266 srcs = [
1267 "test/f32-vsub.cc",
1268 "test/vsub-microkernel-tester.h",
1269 ] + MICROKERNEL_TEST_HDRS,
1270 deps = MICROKERNEL_TEST_DEPS,
1271)
1272
1273xnnpack_unit_test(
1274 name = "f32_vmul_test",
1275 srcs = [
1276 "test/f32-vmul.cc",
1277 "test/vmul-microkernel-tester.h",
1278 ] + MICROKERNEL_TEST_HDRS,
1279 deps = MICROKERNEL_TEST_DEPS,
1280)
1281
1282xnnpack_unit_test(
1283 name = "f32_vmulcaddc_test",
1284 srcs = [
1285 "test/f32-vmulcaddc.cc",
1286 "test/vmulcaddc-microkernel-tester.h",
1287 "src/xnnpack/AlignedAllocator.h",
1288 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1289 deps = MICROKERNEL_TEST_DEPS,
1290)
1291
1292xnnpack_unit_test(
1293 name = "q8_avgpool_test",
1294 srcs = [
1295 "test/q8-avgpool.cc",
1296 "test/avgpool-microkernel-tester.h",
1297 "src/xnnpack/AlignedAllocator.h",
1298 ] + MICROKERNEL_TEST_HDRS,
1299 deps = MICROKERNEL_TEST_DEPS,
1300)
1301
1302xnnpack_unit_test(
1303 name = "q8_igemm_test",
1304 srcs = [
1305 "test/q8-igemm.cc",
1306 "test/gemm-microkernel-tester.h",
1307 "src/xnnpack/AlignedAllocator.h",
1308 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1309 deps = MICROKERNEL_TEST_DEPS,
1310)
1311
1312xnnpack_unit_test(
1313 name = "q8_dwconv_test",
1314 srcs = [
1315 "test/q8-dwconv.cc",
1316 "test/dwconv-microkernel-tester.h",
1317 "src/xnnpack/AlignedAllocator.h",
1318 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1319 deps = MICROKERNEL_TEST_DEPS,
1320)
1321
1322xnnpack_unit_test(
1323 name = "q8_gavgpool_test",
1324 srcs = [
1325 "test/q8-gavgpool.cc",
1326 "test/gavgpool-microkernel-tester.h",
1327 "src/xnnpack/AlignedAllocator.h",
1328 ] + MICROKERNEL_TEST_HDRS,
1329 deps = MICROKERNEL_TEST_DEPS,
1330)
1331
1332xnnpack_unit_test(
1333 name = "q8_gemm_test",
1334 srcs = [
1335 "test/q8-gemm.cc",
1336 "test/gemm-microkernel-tester.h",
1337 "src/xnnpack/AlignedAllocator.h",
1338 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1339 deps = MICROKERNEL_TEST_DEPS,
1340)
1341
1342xnnpack_unit_test(
1343 name = "q8_vadd_test",
1344 srcs = [
1345 "test/q8-vadd.cc",
1346 "test/vadd-microkernel-tester.h",
1347 ] + MICROKERNEL_TEST_HDRS,
1348 deps = MICROKERNEL_TEST_DEPS,
1349)
1350
1351xnnpack_unit_test(
1352 name = "u8_clamp_test",
1353 srcs = [
1354 "test/u8-clamp.cc",
1355 "test/clamp-microkernel-tester.h",
1356 ] + MICROKERNEL_TEST_HDRS,
1357 deps = MICROKERNEL_TEST_DEPS,
1358)
1359
1360xnnpack_unit_test(
1361 name = "u8_lut32norm_test",
1362 srcs = [
1363 "test/u8-lut32norm.cc",
1364 "test/lut-norm-microkernel-tester.h",
1365 ] + MICROKERNEL_TEST_HDRS,
1366 deps = MICROKERNEL_TEST_DEPS,
1367)
1368
1369xnnpack_unit_test(
1370 name = "u8_maxpool_test",
1371 srcs = [
1372 "test/u8-maxpool.cc",
1373 "test/maxpool-microkernel-tester.h",
1374 ] + MICROKERNEL_TEST_HDRS,
1375 deps = MICROKERNEL_TEST_DEPS,
1376)
1377
1378xnnpack_unit_test(
1379 name = "u8_rmax_test",
1380 srcs = [
1381 "test/u8-rmax.cc",
1382 "test/rmax-microkernel-tester.h",
1383 ] + MICROKERNEL_TEST_HDRS,
1384 deps = MICROKERNEL_TEST_DEPS,
1385)
1386
1387xnnpack_unit_test(
1388 name = "x32_packx_test",
1389 srcs = [
1390 "test/x32-packx.cc",
1391 "test/pack-microkernel-tester.h",
1392 "src/xnnpack/AlignedAllocator.h",
1393 ] + MICROKERNEL_TEST_HDRS,
1394 deps = MICROKERNEL_TEST_DEPS,
1395)
1396
1397xnnpack_unit_test(
1398 name = "x32_pad_test",
1399 srcs = [
1400 "test/x32-pad.cc",
1401 "test/pad-microkernel-tester.h",
1402 ] + MICROKERNEL_TEST_HDRS,
1403 deps = MICROKERNEL_TEST_DEPS,
1404)
1405
1406xnnpack_unit_test(
1407 name = "x32_unpool_test",
1408 srcs = [
1409 "test/x32-unpool.cc",
1410 "test/unpool-microkernel-tester.h",
1411 ] + MICROKERNEL_TEST_HDRS,
1412 deps = MICROKERNEL_TEST_DEPS,
1413)
1414
1415xnnpack_unit_test(
1416 name = "x32_zip_test",
1417 srcs = [
1418 "test/x32-zip.cc",
1419 "test/zip-microkernel-tester.h",
1420 ] + MICROKERNEL_TEST_HDRS,
1421 deps = MICROKERNEL_TEST_DEPS,
1422)
1423
1424xnnpack_unit_test(
1425 name = "x8_lut_test",
1426 srcs = [
1427 "test/x8-lut.cc",
1428 "test/lut-microkernel-tester.h",
1429 ] + MICROKERNEL_TEST_HDRS,
1430 deps = MICROKERNEL_TEST_DEPS,
1431)
1432
1433xnnpack_unit_test(
1434 name = "x8_zip_test",
1435 srcs = [
1436 "test/x8-zip.cc",
1437 "test/zip-microkernel-tester.h",
1438 ] + MICROKERNEL_TEST_HDRS,
1439 deps = MICROKERNEL_TEST_DEPS,
1440)
1441
1442########################### Size test for the library ##########################
1443
1444xnnpack_binary(
1445 name = "size_test",
1446 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001447 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001448)
1449
1450########################### Unit tests for operators ###########################
1451
1452xnnpack_unit_test(
1453 name = "add_test",
1454 srcs = [
1455 "test/add.cc",
1456 "test/add-operator-tester.h",
1457 ],
1458 deps = OPERATOR_TEST_DEPS,
1459)
1460
1461xnnpack_unit_test(
1462 name = "argmax_pooling_test",
1463 srcs = [
1464 "test/argmax-pooling.cc",
1465 "test/argmax-pooling-operator-tester.h",
1466 ] + OPERATOR_TEST_PARAMS_HDRS,
1467 deps = OPERATOR_TEST_DEPS,
1468)
1469
1470xnnpack_unit_test(
1471 name = "average_pooling_test",
1472 srcs = [
1473 "test/average-pooling.cc",
1474 "test/average-pooling-operator-tester.h",
1475 ] + OPERATOR_TEST_PARAMS_HDRS,
1476 deps = OPERATOR_TEST_DEPS,
1477)
1478
1479xnnpack_unit_test(
1480 name = "channel_pad_test",
1481 srcs = [
1482 "test/channel-pad.cc",
1483 "test/channel-pad-operator-tester.h",
1484 ] + OPERATOR_TEST_PARAMS_HDRS,
1485 deps = OPERATOR_TEST_DEPS,
1486)
1487
1488xnnpack_unit_test(
1489 name = "channel_shuffle_test",
1490 srcs = [
1491 "test/channel-shuffle.cc",
1492 "test/channel-shuffle-operator-tester.h",
1493 ],
1494 deps = OPERATOR_TEST_DEPS,
1495)
1496
1497xnnpack_unit_test(
1498 name = "clamp_test",
1499 srcs = [
1500 "test/clamp.cc",
1501 "test/clamp-operator-tester.h",
1502 ],
1503 deps = OPERATOR_TEST_DEPS,
1504)
1505
1506xnnpack_unit_test(
1507 name = "convolution_test",
1508 srcs = [
1509 "test/convolution.cc",
1510 "test/convolution-operator-tester.h",
1511 ],
1512 deps = OPERATOR_TEST_DEPS,
1513)
1514
1515xnnpack_unit_test(
1516 name = "convolution_spnchw_test",
1517 srcs = [
1518 "test/convolution-spnchw.cc",
1519 "test/convolution-spnchw-operator-tester.h",
1520 ],
1521 deps = OPERATOR_TEST_DEPS,
1522)
1523
1524xnnpack_unit_test(
1525 name = "deconvolution_test",
1526 srcs = [
1527 "test/deconvolution.cc",
1528 "test/deconvolution-operator-tester.h",
1529 ] + OPERATOR_TEST_PARAMS_HDRS,
1530 deps = OPERATOR_TEST_DEPS,
1531)
1532
1533xnnpack_unit_test(
1534 name = "fully_connected_test",
1535 srcs = [
1536 "test/fully-connected.cc",
1537 "test/fully-connected-operator-tester.h",
1538 ],
1539 deps = OPERATOR_TEST_DEPS,
1540)
1541
1542xnnpack_unit_test(
1543 name = "global_average_pooling_test",
1544 srcs = [
1545 "test/global-average-pooling.cc",
1546 "test/global-average-pooling-operator-tester.h",
1547 ] + OPERATOR_TEST_PARAMS_HDRS,
1548 deps = OPERATOR_TEST_DEPS,
1549)
1550
1551xnnpack_unit_test(
1552 name = "global_average_pooling_spnchw_test",
1553 srcs = [
1554 "test/global-average-pooling-spnchw.cc",
1555 "test/global-average-pooling-spnchw-operator-tester.h",
1556 ],
1557 deps = OPERATOR_TEST_DEPS,
1558)
1559
1560xnnpack_unit_test(
1561 name = "hardswish_test",
1562 srcs = [
1563 "test/hardswish.cc",
1564 "test/hardswish-operator-tester.h",
1565 ],
1566 deps = OPERATOR_TEST_DEPS,
1567)
1568
1569xnnpack_unit_test(
1570 name = "leaky_relu_test",
1571 srcs = [
1572 "test/leaky-relu.cc",
1573 "test/leaky-relu-operator-tester.h",
1574 ],
1575 deps = OPERATOR_TEST_DEPS,
1576)
1577
1578xnnpack_unit_test(
1579 name = "max_pooling_test",
1580 srcs = [
1581 "test/max-pooling.cc",
1582 "test/max-pooling-operator-tester.h",
1583 ] + OPERATOR_TEST_PARAMS_HDRS,
1584 deps = OPERATOR_TEST_DEPS,
1585)
1586
1587xnnpack_unit_test(
1588 name = "prelu_test",
1589 srcs = [
1590 "test/prelu.cc",
1591 "test/prelu-operator-tester.h",
1592 ] + OPERATOR_TEST_PARAMS_HDRS,
1593 deps = OPERATOR_TEST_DEPS,
1594)
1595
1596xnnpack_unit_test(
1597 name = "sigmoid_test",
1598 srcs = [
1599 "test/sigmoid.cc",
1600 "test/sigmoid-operator-tester.h",
1601 ],
1602 deps = OPERATOR_TEST_DEPS,
1603)
1604
1605xnnpack_unit_test(
1606 name = "softargmax_test",
1607 srcs = [
1608 "test/softargmax.cc",
1609 "test/softargmax-operator-tester.h",
1610 ],
1611 deps = OPERATOR_TEST_DEPS,
1612)
1613
1614xnnpack_unit_test(
1615 name = "unpooling_test",
1616 srcs = [
1617 "test/unpooling.cc",
1618 "test/unpooling-operator-tester.h",
1619 ],
1620 deps = OPERATOR_TEST_DEPS,
1621)
1622
1623############################# Build configurations #############################
1624
1625config_setting(
1626 name = "linux_k8",
1627 values = {
1628 "cpu": "k8",
1629 },
1630)
1631
1632config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07001633 name = "linux_aarch64",
1634 values = {
1635 "cpu": "aarch64",
1636 },
1637)
1638
1639config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001640 name = "android",
1641 values = {"crosstool_top": "//external:android/crosstool"},
1642)
1643
1644config_setting(
1645 name = "android_armv7",
1646 values = {
1647 "crosstool_top": "//external:android/crosstool",
1648 "cpu": "armeabi-v7a",
1649 },
1650)
1651
1652config_setting(
1653 name = "android_arm64",
1654 values = {
1655 "crosstool_top": "//external:android/crosstool",
1656 "cpu": "arm64-v8a",
1657 },
1658)
1659
1660config_setting(
1661 name = "android_x86",
1662 values = {
1663 "crosstool_top": "//external:android/crosstool",
1664 "cpu": "x86",
1665 },
1666)
1667
1668config_setting(
1669 name = "android_x86_64",
1670 values = {
1671 "crosstool_top": "//external:android/crosstool",
1672 "cpu": "x86_64",
1673 },
1674)
1675
1676config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07001677 name = "macos_x86_64",
1678 values = {
1679 "apple_platform_type": "macos",
1680 "cpu": "darwin",
1681 },
1682)
1683
1684config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001685 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001686 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07001687)
1688
1689config_setting(
1690 name = "emscripten_wasm",
1691 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001692 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001693 "cpu": "wasm",
1694 },
1695)
1696
1697config_setting(
1698 name = "emscripten_wasmsimd",
1699 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001700 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001701 "cpu": "wasm",
1702 "features": "wasmsimd",
1703 },
1704)
1705
1706config_setting(
1707 name = "emscripten_asmjs",
1708 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001709 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001710 "cpu": "asmjs",
1711 },
1712)
1713
Marat Dukhan174706e2019-10-09 16:33:53 -07001714# Builds with -c dbg
Marat Dukhan08c4a432019-10-03 09:29:21 -07001715config_setting(
1716 name = "debug_build",
1717 values = {
1718 "compilation_mode": "dbg",
1719 },
Marat Dukhan08c4a432019-10-03 09:29:21 -07001720)
1721
Marat Dukhan174706e2019-10-09 16:33:53 -07001722# Builds with -c opt
Marat Dukhan08c4a432019-10-03 09:29:21 -07001723config_setting(
1724 name = "optimized_build",
1725 values = {
1726 "compilation_mode": "opt",
1727 },
Marat Dukhan08c4a432019-10-03 09:29:21 -07001728)
1729
1730# Enables usage of assembly kernels.
1731config_setting(
1732 name = "xnn_enable_assembly_explicit_true",
1733 define_values = {"xnn_enable_assembly": "true"},
1734)
1735
1736# Disables usage of assembly kernels.
1737config_setting(
1738 name = "xnn_enable_assembly_explicit_false",
1739 define_values = {"xnn_enable_assembly": "false"},
1740)