blob: 9bb91efd73198f4973bd36e7e5ac7bc976a62164 [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
31MICROKERNEL_TEST_DEPS = [
32 ":ukernels",
Frank Barchard7e955972019-10-11 10:34:25 -070033 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070034 "@cpuinfo",
35 "@FP16",
36 "@pthreadpool",
37]
38
39OPERATOR_TEST_DEPS = [
40 ":XNNPACK",
41 "@pthreadpool",
42 "@FP16",
43]
44
45OPERATOR_SRCS = [
46 "src/add.c",
47 "src/argmax-pooling.c",
48 "src/average-pooling.c",
49 "src/channel-pad.c",
50 "src/channel-shuffle.c",
51 "src/clamp.c",
52 "src/convolution-spnchw.c",
53 "src/convolution.c",
54 "src/deconvolution.c",
55 "src/fully-connected.c",
56 "src/global-average-pooling-spnchw.c",
57 "src/global-average-pooling.c",
58 "src/hardswish.c",
59 "src/leaky-relu.c",
60 "src/max-pooling.c",
61 "src/prelu.c",
62 "src/sigmoid.c",
63 "src/softargmax.c",
64 "src/unpooling.c",
65]
66
67SCALAR_UKERNELS = [
68 "src/f32-argmaxpool/mp9p8q-scalar.c",
69 "src/f32-argmaxpool/up4-scalar.c",
70 "src/f32-argmaxpool/up9-scalar.c",
71 "src/f32-avgpool/mp9p8q-scalar.c",
72 "src/f32-avgpool/up9-scalar.c",
73 "src/f32-clamp/scalar.c",
74 "src/f32-igemm/1x4-scalar.c",
75 "src/f32-igemm/2x4-scalar.c",
76 "src/f32-igemm/4x2-scalar.c",
77 "src/f32-igemm/4x4-scalar.c",
78 "src/f32-dwconv/up1x25-scalar.c",
79 "src/f32-dwconv/up1x4-scalar.c",
80 "src/f32-dwconv/up1x9-scalar.c",
81 "src/f32-gavgpool/mp7p7q-scalar.c",
82 "src/f32-gavgpool/up7-scalar.c",
83 "src/f32-gemm/1x4-scalar.c",
84 "src/f32-gemm/2x4-scalar.c",
85 "src/f32-gemm/4x2-scalar.c",
86 "src/f32-gemm/4x4-scalar.c",
87 "src/f32-gemminc/1x4-scalar.c",
88 "src/f32-gemminc/2x4-scalar.c",
89 "src/f32-gemminc/4x4-scalar.c",
90 "src/f32-hswish/scalar.c",
91 "src/f32-maxpool/9p8q-scalar.c",
92 "src/f32-pavgpool/mp9p8q-scalar.c",
93 "src/f32-pavgpool/up9-scalar.c",
94 "src/f32-ppmm/2x4-scalar.c",
95 "src/f32-ppmm/3x3-scalar.c",
96 "src/f32-ppmm/4x2-scalar.c",
97 "src/f32-ppmm/4x4-scalar.c",
98 "src/f32-prelu/x4-scalar.c",
99 "src/f32-rmax/scalar.c",
100 "src/f32-spmm/1x1-scalar-pipelined.c",
101 "src/f32-spmm/1x1-scalar-unroll2.c",
102 "src/f32-spmm/1x1-scalar.c",
103 "src/f32-spmm/2x1-scalar-pipelined.c",
104 "src/f32-spmm/2x1-scalar-unroll2.c",
105 "src/f32-spmm/2x1-scalar.c",
106 "src/f32-spmm/4x1-scalar-pipelined.c",
107 "src/f32-spmm/4x1-scalar-unroll2.c",
108 "src/f32-spmm/4x1-scalar.c",
109 "src/f32-spmm/8x1-scalar-pipelined.c",
110 "src/f32-spmm/8x1-scalar-unroll2.c",
111 "src/f32-spmm/8x1-scalar.c",
112 "src/f32-vadd/scalar.c",
113 "src/f32-vmul/scalar.c",
114 "src/f32-vmulcaddc/c1-scalar-x2.c",
115 "src/f32-vsub/scalar.c",
116 "src/q8-avgpool/mp9p8q-scalar.c",
117 "src/q8-avgpool/up9-scalar.c",
118 "src/q8-igemm/2x2-scalar.c",
119 "src/q8-dwconv/up1x9-scalar.c",
120 "src/q8-gavgpool/mp7p7q-scalar.c",
121 "src/q8-gavgpool/up7-scalar.c",
122 "src/q8-gemm/2x2-scalar.c",
123 "src/q8-vadd/scalar.c",
124 "src/u8-clamp/scalar.c",
125 "src/u8-lut32norm/scalar.c",
126 "src/u8-maxpool/9p8q-scalar.c",
127 "src/u8-rmax/scalar.c",
128 "src/x32-packx/x2-scalar.c",
129 "src/x32-packx/x3-scalar.c",
130 "src/x32-packx/x4-scalar.c",
131 "src/x32-pad/x2-scalar.c",
132 "src/x32-unpool/scalar.c",
133 "src/x32-zip/x2-scalar.c",
134 "src/x32-zip/x3-scalar.c",
135 "src/x32-zip/x4-scalar.c",
136 "src/x32-zip/xm-scalar.c",
137 "src/x8-lut/scalar.c",
138 "src/x8-zip/x2-scalar.c",
139 "src/x8-zip/x3-scalar.c",
140 "src/x8-zip/x4-scalar.c",
141 "src/x8-zip/xm-scalar.c",
142]
143
144PSIMD_UKERNELS = [
145 "src/f32-argmaxpool/mp9p8q-psimd.c",
146 "src/f32-argmaxpool/up4-psimd.c",
147 "src/f32-argmaxpool/up9-psimd.c",
148 "src/f32-avgpool/mp9p8q-psimd.c",
149 "src/f32-avgpool/up9-psimd.c",
150 "src/f32-clamp/psimd.c",
151 "src/f32-igemm/1x8-psimd-loadsplat.c",
152 "src/f32-igemm/1x8-psimd-splat.c",
153 "src/f32-igemm/1x8s4-psimd.c",
154 "src/f32-igemm/4x2c4-psimd.c",
155 "src/f32-igemm/4x8-psimd-loadsplat.c",
156 "src/f32-igemm/4x8-psimd-splat.c",
157 "src/f32-igemm/4x8s4-psimd.c",
158 "src/f32-igemm/6x8-psimd-loadsplat.c",
159 "src/f32-igemm/6x8-psimd-splat.c",
160 "src/f32-igemm/6x8s4-psimd.c",
161 "src/f32-dwconv/up4x25-psimd.c",
162 "src/f32-dwconv/up4x4-psimd.c",
163 "src/f32-dwconv/up4x9-psimd.c",
164 "src/f32-gavgpool/mp7p7q-psimd.c",
165 "src/f32-gavgpool/up7-psimd.c",
166 "src/f32-gemm/1x8-psimd-loadsplat.c",
167 "src/f32-gemm/1x8-psimd-splat.c",
168 "src/f32-gemm/1x8s4-psimd.c",
169 "src/f32-gemm/4x8-psimd-loadsplat.c",
170 "src/f32-gemm/4x8-psimd-splat.c",
171 "src/f32-gemm/4x8s4-psimd.c",
172 "src/f32-gemm/6x8-psimd-loadsplat.c",
173 "src/f32-gemm/6x8-psimd-splat.c",
174 "src/f32-gemm/6x8s4-psimd.c",
175 "src/f32-gemminc/1x8-psimd-loadsplat.c",
176 "src/f32-gemminc/1x8-psimd-splat.c",
177 "src/f32-gemminc/1x8s4-psimd.c",
178 "src/f32-gemminc/4x8-psimd-loadsplat.c",
179 "src/f32-gemminc/4x8-psimd-splat.c",
180 "src/f32-gemminc/4x8s4-psimd.c",
181 "src/f32-gemminc/6x8-psimd-loadsplat.c",
182 "src/f32-gemminc/6x8-psimd-splat.c",
183 "src/f32-gemminc/6x8s4-psimd.c",
184 "src/f32-hswish/psimd.c",
185 "src/f32-maxpool/9p8q-psimd.c",
186 "src/f32-pavgpool/mp9p8q-psimd.c",
187 "src/f32-pavgpool/up9-psimd.c",
188 "src/f32-ppmm/4x8-psimd.c",
189 "src/f32-prelu/x4-psimd.c",
190 "src/f32-vadd/psimd.c",
191 "src/f32-vmul/psimd.c",
192 "src/f32-vmulcaddc/c4-psimd-x2.c",
193 "src/f32-vsub/psimd.c",
194 "src/x32-packx/x4-psimd.c",
195 "src/x32-pad/x2-psimd.c",
196 "src/x32-unpool/psimd.c",
197 "src/x32-zip/x2-psimd.c",
198 "src/x32-zip/x3-psimd.c",
199 "src/x32-zip/x4-psimd.c",
200 "src/x32-zip/xm-psimd.c",
201]
202
203# ISA-specific micro-kernels
204NEON_UKERNELS = [
205 "src/f32-avgpool/mp9p8q-neon.c",
206 "src/f32-avgpool/up9-neon.c",
207 "src/f32-clamp/neon.c",
208 "src/f32-igemm/1x8-neon-ld64.c",
209 "src/f32-igemm/4x12-neon-ld64.c",
210 "src/f32-igemm/4x2-neon-ld64.c",
211 "src/f32-igemm/4x4-neon-ld64.c",
212 "src/f32-igemm/4x8-neon-ld128.c",
213 "src/f32-igemm/4x8-neon-ld64.c",
214 "src/f32-igemm/6x8-neon-ld64.c",
215 "src/f32-dwconv/up4x9-neon.c",
216 "src/f32-gavgpool-spchw/neon-x4.c",
217 "src/f32-gavgpool/mp7p7q-neon.c",
218 "src/f32-gavgpool/up7-neon.c",
219 "src/f32-gemm/1x8-neon-ld64.c",
220 "src/f32-gemm/4x12-neon-ld64.c",
221 "src/f32-gemm/4x2-neon-ld64.c",
222 "src/f32-gemm/4x8-neon-ld128.c",
223 "src/f32-gemm/4x8-neon-ld64.c",
224 "src/f32-gemm/5x8-neon-ld64.c",
225 "src/f32-gemm/6x8-neon-ld64.c",
226 "src/f32-gemminc/1x8-neon-ld64.c",
227 "src/f32-gemminc/4x12-neon-ld64.c",
228 "src/f32-gemminc/4x8-neon-ld128.c",
229 "src/f32-gemminc/4x8-neon-ld64.c",
230 "src/f32-gemminc/5x8-neon-ld64.c",
231 "src/f32-gemminc/6x8-neon-ld64.c",
232 "src/f32-hswish/neon.c",
233 "src/f32-pavgpool/mp9p8q-neon.c",
234 "src/f32-pavgpool/up9-neon.c",
235 "src/f32-ppmm/4x8-neon.c",
236 "src/f32-ppmm/8x8-neon.c",
237 "src/f32-rmax/neon.c",
238 "src/f32-vmulcaddc/c4-neon-x2.c",
239 "src/q8-avgpool/mp9p8q-neon.c",
240 "src/q8-avgpool/up9-neon.c",
241 "src/q8-igemm/4x8-neon.c",
242 "src/q8-igemm/8x8-neon.c",
243 "src/q8-dwconv/up8x9-neon.c",
244 "src/q8-gavgpool/mp7p7q-neon.c",
245 "src/q8-gavgpool/up7-neon.c",
246 "src/q8-gemm/4x8-neon.c",
247 "src/q8-gemm/8x8-neon.c",
248 "src/q8-vadd/neon.c",
249 "src/u8-clamp/neon.c",
250 "src/u8-maxpool/9p8q-neon.c",
251 "src/u8-rmax/neon.c",
252 "src/x32-packx/x4-neon-st4.c",
253 "src/x32-pad/x2-neon.c",
254 "src/x32-zip/x2-neon.c",
255 "src/x32-zip/x3-neon.c",
256 "src/x32-zip/x4-neon.c",
257 "src/x32-zip/xm-neon.c",
258 "src/x8-zip/x2-neon.c",
259 "src/x8-zip/x3-neon.c",
260 "src/x8-zip/x4-neon.c",
261 "src/x8-zip/xm-neon.c",
262]
263
264NEONFMA_UKERNELS = [
Frank Barcharddb45b6a2019-10-09 16:42:45 -0700265 "src/f32-igemm/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700266 "src/f32-igemm/4x12-neonfma-ld64.c",
267 "src/f32-igemm/4x2-neonfma-ld64.c",
268 "src/f32-igemm/4x4-neonfma-ld64.c",
269 "src/f32-igemm/4x8-neonfma-ld128.c",
270 "src/f32-igemm/4x8-neonfma-ld64.c",
271 "src/f32-igemm/6x8-neonfma-ld64.c",
272 "src/f32-dwconv/up4x9-neonfma.c",
273 "src/f32-dwconv/up8x9-neonfma.c",
274 "src/f32-gemm/1x8-neonfma-ld64.c",
275 "src/f32-gemm/4x12-neonfma-ld64.c",
276 "src/f32-gemm/4x2-neonfma-ld64.c",
277 "src/f32-gemm/4x8-neonfma-ld128.c",
278 "src/f32-gemm/4x8-neonfma-ld64.c",
279 "src/f32-gemm/5x8-neonfma-ld64.c",
280 "src/f32-gemm/6x8-neonfma-ld64.c",
281 "src/f32-gemminc/1x8-neonfma-ld64.c",
282 "src/f32-gemminc/4x12-neonfma-ld64.c",
283 "src/f32-gemminc/4x8-neonfma-ld128.c",
284 "src/f32-gemminc/4x8-neonfma-ld64.c",
285 "src/f32-gemminc/5x8-neonfma-ld64.c",
286 "src/f32-gemminc/6x8-neonfma-ld64.c",
287 "src/f32-hswish/neonfma.c",
288 "src/f32-ppmm/4x8-neonfma.c",
289 "src/f32-ppmm/8x8-neonfma.c",
290 "src/f32-vmulcaddc/c4-neonfma-x2.c",
291]
292
293AARCH64_NEONFMA_UKERNELS = [
294 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
295 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
296 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
297 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
298 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
299 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
300 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
301 "src/f32-spmm/12x1-neonfma.c",
302 "src/f32-spmm/12x2-neonfma.c",
303 "src/f32-spmm/12x4-neonfma.c",
304 "src/f32-spmm/16x1-neonfma-pipelined.c",
305 "src/f32-spmm/16x1-neonfma-unroll2.c",
306 "src/f32-spmm/16x1-neonfma.c",
307 "src/f32-spmm/16x2-neonfma.c",
308 "src/f32-spmm/16x4-neonfma.c",
309 "src/f32-spmm/4x1-neonfma-pipelined.c",
310 "src/f32-spmm/4x1-neonfma-unroll2.c",
311 "src/f32-spmm/4x1-neonfma.c",
312 "src/f32-spmm/4x2-neonfma.c",
313 "src/f32-spmm/4x4-neonfma.c",
314 "src/f32-spmm/8x1-neonfma-pipelined.c",
315 "src/f32-spmm/8x1-neonfma-unroll2.c",
316 "src/f32-spmm/8x1-neonfma.c",
317 "src/f32-spmm/8x2-neonfma.c",
318 "src/f32-spmm/8x4-neonfma.c",
319]
320
321AARCH64_NEONFP16ARITH_UKERNELS = [
322 "src/f16-gemm/4x8-neonfp16arith-ld64.c",
323 "src/f16-gemm/6x8-neonfp16arith-ld64.c",
324 "src/f16-gemm/8x8-neonfp16arith-ld64.c",
325]
326
327SSE_UKERNELS = [
328 "src/f32-avgpool/mp9p8q-sse.c",
329 "src/f32-avgpool/up9-sse.c",
330 "src/f32-clamp/sse.c",
331 "src/f32-igemm/1x8-sse-dup.c",
332 "src/f32-igemm/1x8-sse-load1.c",
333 "src/f32-igemm/1x8s4-sse.c",
334 "src/f32-igemm/4x2c4-sse.c",
335 "src/f32-igemm/4x8-sse-dup.c",
336 "src/f32-igemm/4x8-sse-load1.c",
337 "src/f32-igemm/4x8s4-sse.c",
338 "src/f32-dwconv/up4x25-sse.c",
339 "src/f32-dwconv/up4x4-sse.c",
340 "src/f32-dwconv/up4x9-sse.c",
341 "src/f32-gavgpool-spchw/sse-x4.c",
342 "src/f32-gavgpool/mp7p7q-sse.c",
343 "src/f32-gavgpool/up7-sse.c",
344 "src/f32-gemm/1x8-sse-dup.c",
345 "src/f32-gemm/1x8-sse-load1.c",
346 "src/f32-gemm/1x8s4-sse.c",
347 "src/f32-gemm/4x8-sse-dup.c",
348 "src/f32-gemm/4x8-sse-load1.c",
349 "src/f32-gemm/4x8s4-sse.c",
350 "src/f32-gemminc/1x8-sse-dup.c",
351 "src/f32-gemminc/1x8-sse-load1.c",
352 "src/f32-gemminc/1x8s4-sse.c",
353 "src/f32-gemminc/4x8-sse-dup.c",
354 "src/f32-gemminc/4x8-sse-load1.c",
355 "src/f32-gemminc/4x8s4-sse.c",
356 "src/f32-hswish/sse.c",
357 "src/f32-maxpool/9p8q-sse.c",
358 "src/f32-pavgpool/mp9p8q-sse.c",
359 "src/f32-pavgpool/up9-sse.c",
360 "src/f32-dwconv-spchw/3x3p1-sse.c",
361 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
362 "src/f32-ppmm/4x8-sse.c",
363 "src/f32-prelu/x4-sse.c",
364 "src/f32-rmax/sse.c",
365 "src/f32-spmm/4x1-sse.c",
366 "src/f32-spmm/8x1-sse.c",
367 "src/f32-vadd/sse.c",
368 "src/f32-vmul/sse.c",
369 "src/f32-vmulcaddc/c4-sse-x2.c",
370 "src/f32-vsub/sse.c",
371 "src/x32-packx/x4-sse.c",
372]
373
374SSE2_UKERNELS = [
375 "src/f32-argmaxpool/mp9p8q-sse2.c",
376 "src/f32-argmaxpool/up4-sse2.c",
377 "src/f32-argmaxpool/up9-sse2.c",
378 "src/q8-avgpool/mp9p8q-sse2.c",
379 "src/q8-avgpool/up9-sse2.c",
380 "src/q8-igemm/4x4c2-sse2.c",
381 "src/q8-dwconv/up8x9-sse2.c",
382 "src/q8-gavgpool/mp7p7q-sse2.c",
383 "src/q8-gavgpool/up7-sse2.c",
384 "src/q8-gemm/2x4c8-sse2.c",
385 "src/q8-gemm/4x4c2-sse2.c",
386 "src/q8-vadd/sse2.c",
387 "src/u8-clamp/sse2.c",
388 "src/u8-maxpool/9p8q-sse2.c",
389 "src/u8-rmax/sse2.c",
390 "src/x32-pad/x2-sse2.c",
391 "src/x32-zip/x2-sse2.c",
392 "src/x32-zip/x3-sse2.c",
393 "src/x32-zip/x4-sse2.c",
394 "src/x32-zip/xm-sse2.c",
395 "src/x8-zip/x2-sse2.c",
396 "src/x8-zip/x3-sse2.c",
397 "src/x8-zip/x4-sse2.c",
398 "src/x8-zip/xm-sse2.c",
399]
400
401AVX_UKERNELS = [
402 "src/f32-rmax/avx.c",
403]
404
405AVX512F_UKERNELS = [
406 "src/f32-rmax/avx512f.c",
407]
408
409AARCH32_ASM_UKERNELS = [
410 "src/q8-dwconv/up8x9-aarch32-neon.S",
411]
412
413AARCH64_ASM_UKERNELS = [
414 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
415 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
416 "src/f32-gemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700417 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700418 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a57.S",
419 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a75.S",
420 "src/f32-gemm/4x12-aarch64-neonfma-cortex-a53.S",
421 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a57.S",
422 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a75.S",
423 "src/f32-gemm/4x8-aarch64-neonfma-ld128.S",
424 "src/f32-gemm/4x8-aarch64-neonfma-ld64.S",
425 "src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S",
426 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a57.S",
427 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S",
428 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S",
429 "src/f32-gemm/6x8-aarch64-neonfma-ld128.S",
430 "src/f32-gemm/6x8-aarch64-neonfma-ld64.S",
431 "src/f32-gemminc/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700432 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700433 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a57.S",
434 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a75.S",
435 "src/f32-gemminc/4x12-aarch64-neonfma-cortex-a53.S",
436 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a57.S",
437 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a75.S",
438 "src/f32-gemminc/4x8-aarch64-neonfma-ld128.S",
439 "src/f32-gemminc/4x8-aarch64-neonfma-ld64.S",
440 "src/f32-gemminc/5x8-aarch64-neonfma-cortex-a75.S",
441 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a57.S",
442 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a73.S",
443 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a75.S",
444 "src/f32-gemminc/6x8-aarch64-neonfma-ld128.S",
445 "src/f32-gemminc/6x8-aarch64-neonfma-ld64.S",
446 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700447 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700448 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S",
449 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S",
450 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
451 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
452 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
453 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
454 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
455 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
456]
457
458INTERNAL_MICROKERNEL_HDRS = [
459 "src/xnnpack/argmaxpool.h",
460 "src/xnnpack/avgpool.h",
461 "src/xnnpack/clamp.h",
462 "src/xnnpack/common.h",
463 "src/xnnpack/conv.h",
464 "src/xnnpack/dwconv.h",
465 "src/xnnpack/gavgpool.h",
466 "src/xnnpack/gemm.h",
467 "src/xnnpack/hswish.h",
468 "src/xnnpack/igemm.h",
469 "src/xnnpack/lut.h",
470 "src/xnnpack/math.h",
471 "src/xnnpack/maxpool.h",
472 "src/xnnpack/packx.h",
473 "src/xnnpack/pad.h",
474 "src/xnnpack/params.h",
475 "src/xnnpack/pavgpool.h",
476 "src/xnnpack/ppmm.h",
477 "src/xnnpack/prelu.h",
478 "src/xnnpack/rmax.h",
479 "src/xnnpack/scalar-utils.h",
480 "src/xnnpack/spmm.h",
481 "src/xnnpack/unpool.h",
482 "src/xnnpack/vadd.h",
483 "src/xnnpack/vmul.h",
484 "src/xnnpack/vmulcaddc.h",
485 "src/xnnpack/vsub.h",
486 "src/xnnpack/zip.h",
487]
488
489INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
490 "include/xnnpack.h",
491 "src/xnnpack/allocator.h",
492 "src/xnnpack/compute.h",
493 "src/xnnpack/im2col.h",
494 "src/xnnpack/indirection.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700495 "src/xnnpack/operator.h",
496 "src/xnnpack/pack.h",
497 "src/xnnpack/requantization.h",
498 "src/xnnpack/requantization-stubs.h",
499]
500
501MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
502 "src/xnnpack/requantization.h",
503 "include/xnnpack.h",
504]
505
506MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
507 "src/xnnpack/isa-checks.h",
508 "src/xnnpack/requantization.h",
509 "include/xnnpack.h",
510]
511
512OPERATOR_TEST_PARAMS_HDRS = [
513 "src/xnnpack/params.h",
514 "src/xnnpack/common.h",
515]
516
517WEIGHTS_PACK_HDRS = [
518 "src/xnnpack/pack.h",
519 "src/xnnpack/operator.h",
520 "src/xnnpack/compute.h",
521]
522
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700523LOGGING_COPTS = select({
524 # No logging in optimized mode
525 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
526 # Full logging in debug mode
527 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
528 # Error-only logging in default (fastbuild) mode
529 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
530})
531
532LOGGING_HDRS = [
533 "src/xnnpack/log.h",
534]
535
Marat Dukhan08c4a432019-10-03 09:29:21 -0700536xnnpack_cc_library(
537 name = "scalar_ukernels",
538 srcs = SCALAR_UKERNELS,
539 hdrs = INTERNAL_HDRS,
540 aarch32_copts = ["-marm"],
541 copts = xnnpack_std_copts(),
542 deps = [
543 "@FP16",
544 "@FXdiv",
545 ],
546)
547
548xnnpack_cc_library(
549 name = "psimd_ukernels",
550 srcs = PSIMD_UKERNELS,
551 hdrs = INTERNAL_HDRS,
552 aarch32_copts = [
553 "-marm",
554 "-mfpu=neon",
555 ],
556 copts = xnnpack_std_copts(),
557 optimized_copts = [
558 "-O3",
559 "-ffast-math",
560 ],
561 deps = [
562 "@FP16",
563 "@psimd",
564 ],
565)
566
567xnnpack_cc_library(
568 name = "neon_ukernels",
569 hdrs = INTERNAL_HDRS,
570 aarch32_copts = [
571 "-marm",
572 "-mfpu=neon",
573 ],
574 aarch32_srcs = NEON_UKERNELS,
575 aarch64_srcs = NEON_UKERNELS,
576 copts = xnnpack_std_copts(),
577 deps = ["@FP16"],
578)
579
580xnnpack_cc_library(
581 name = "neonfma_ukernels",
582 hdrs = INTERNAL_HDRS,
583 aarch32_copts = [
584 "-marm",
585 "-mfpu=neon-vfpv4",
586 ],
587 aarch32_srcs = NEONFMA_UKERNELS,
588 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
589 copts = xnnpack_std_copts(),
590 deps = ["@FP16"],
591)
592
593xnnpack_cc_library(
594 name = "neonfp16arith_ukernels",
595 hdrs = INTERNAL_HDRS,
596 aarch64_copts = ["-march=armv8.2-a+fp16"],
597 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
598 copts = xnnpack_std_copts(),
599 deps = ["@FP16"],
600)
601
602xnnpack_cc_library(
603 name = "sse2_ukernels",
604 hdrs = INTERNAL_HDRS,
605 copts = xnnpack_std_copts(),
606 x86_copts = ["-msse2"],
607 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
608 deps = ["@FP16"],
609)
610
611xnnpack_cc_library(
612 name = "avx_ukernels",
613 hdrs = INTERNAL_HDRS,
614 copts = xnnpack_std_copts(),
615 x86_copts = ["-mavx"],
616 x86_srcs = AVX_UKERNELS,
617 deps = ["@FP16"],
618)
619
620xnnpack_cc_library(
621 name = "avx512f_ukernels",
622 hdrs = INTERNAL_HDRS,
623 copts = xnnpack_std_copts(),
624 x86_copts = ["-mavx512f"],
625 x86_srcs = AVX512F_UKERNELS,
626 deps = ["@FP16"],
627)
628
629xnnpack_cc_library(
630 name = "asm_ukernels",
631 hdrs = ["src/xnnpack/assembly.h"],
632 aarch32_srcs = AARCH32_ASM_UKERNELS,
633 aarch64_srcs = AARCH64_ASM_UKERNELS,
634)
635
636xnnpack_aggregate_library(
637 name = "ukernels",
638 aarch32_deps = [
639 ":psimd_ukernels",
640 ":neon_ukernels",
641 ":neonfma_ukernels",
642 ":asm_ukernels",
643 ],
644 aarch64_deps = [
645 ":psimd_ukernels",
646 ":neon_ukernels",
647 ":neonfma_ukernels",
648 ":neonfp16arith_ukernels",
649 ":asm_ukernels",
650 ],
651 generic_deps = [":scalar_ukernels"],
652 wasmsimd_deps = [
653 ":psimd_ukernels",
654 ],
655 x86_deps = [
656 ":psimd_ukernels",
657 ":sse2_ukernels",
658 ":avx_ukernels",
659 ":avx512f_ukernels",
660 ],
661)
662
663xnnpack_cc_library(
664 name = "im2col",
665 srcs = ["src/im2col.c"],
666 hdrs = [
667 "src/xnnpack/common.h",
668 "src/xnnpack/im2col.h",
669 ],
670 copts = xnnpack_std_copts(),
671)
672
673xnnpack_cc_library(
674 name = "indirection",
675 srcs = ["src/indirection.c"],
676 hdrs = INTERNAL_HDRS,
677 copts = xnnpack_std_copts(),
678 deps = [
679 "@FP16",
680 "@FXdiv",
681 "@pthreadpool",
682 ],
683)
684
685xnnpack_cc_library(
686 name = "operator_run",
687 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700688 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
689 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700690 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
691 "-Wno-vla",
692 ],
693 deps = [
694 "@FP16",
695 "@FXdiv",
696 "@clog",
697 "@pthreadpool",
698 ],
699)
700
701cc_library(
702 name = "enable_assembly",
703 defines = select({
704 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
705 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -0700706 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700707 }),
708)
709
Marat Dukhancf056b22019-10-07 10:26:29 -0700710xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700711 name = "operators",
712 srcs = OPERATOR_SRCS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700713 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -0700714 ],
715 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700716 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700717 "-Isrc",
718 "-Iinclude",
719 ] + select({
720 ":debug_build": [],
721 "//conditions:default": xnnpack_min_size_copts(),
722 }),
Marat Dukhancf056b22019-10-07 10:26:29 -0700723 wasm_srcs = ["src/wasm-stubs.c"],
724 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700725 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700726 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700727 "@FP16",
728 "@FXdiv",
729 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700730 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700731 ],
732)
733
734cc_library(
735 name = "XNNPACK",
736 srcs = [
737 "src/init.c",
738 ],
739 copts = xnnpack_std_copts() + LOGGING_COPTS + [
740 "-Isrc",
741 "-Iinclude",
742 ] + select({
743 ":debug_build": [],
744 "//conditions:default": xnnpack_min_size_copts(),
745 }),
746 includes = ["include"],
747 linkstatic = True,
748 textual_hdrs = ["include/xnnpack.h"],
749 visibility = xnnpack_visibility(),
750 deps = [
751 ":enable_assembly",
752 ":ukernels",
753 ":operator_run",
754 ":operators",
755 "@clog",
756 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -0700757 ] + select({
758 ":emscripten": [],
759 "//conditions:default": ["@cpuinfo"],
760 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700761)
762
763cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700764 name = "xnnpack_operators_nhwc_f32",
765 srcs = [
766 "src/init.c",
767 ],
768 copts = xnnpack_std_copts() + LOGGING_COPTS + [
769 "-Isrc",
770 "-Iinclude",
771 ] + select({
772 ":debug_build": [],
773 "//conditions:default": xnnpack_min_size_copts(),
774 }),
775 defines = [
776 "XNN_NO_Q8_OPERATORS",
777 "XNN_NO_U8_OPERATORS",
778 "XNN_NO_X8_OPERATORS",
779 "XNN_NO_SPNCHW_OPERATORS",
780 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700781 includes = ["include"],
782 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700783 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700784 visibility = xnnpack_visibility(),
785 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700786 ":enable_assembly",
787 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700788 ":operator_run",
789 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700790 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700791 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700792 ] + select({
793 ":emscripten": [],
794 "//conditions:default": ["@cpuinfo"],
795 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700796)
797
Marat Dukhancf056b22019-10-07 10:26:29 -0700798xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700799 name = "bench_utils",
800 srcs = ["bench/utils.cc"],
801 hdrs = ["bench/utils.h"],
802 copts = ["-Wno-unused-result"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700803 deps = ["@cpuinfo"],
804)
805
Frank Barchard7e955972019-10-11 10:34:25 -0700806######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -0700807
808xnnpack_benchmark(
809 name = "q8_gemm_bench",
810 srcs = [
811 "bench/gemm.h",
812 "bench/q8-gemm.cc",
813 "src/xnnpack/AlignedAllocator.h",
814 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
815 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
816 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
817)
818
819xnnpack_benchmark(
820 name = "f16_gemm_bench",
821 srcs = [
822 "bench/f16-gemm.cc",
823 "bench/gemm.h",
824 "src/xnnpack/AlignedAllocator.h",
825 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
826 copts = ["-Wno-unused-function"],
827 deps = MICROKERNEL_BENCHMARK_DEPS,
828)
829
830xnnpack_benchmark(
831 name = "f32_igemm_bench",
832 srcs = [
833 "bench/f32-igemm.cc",
834 "bench/conv.h",
835 "src/xnnpack/AlignedAllocator.h",
836 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -0700837 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700838)
839
840xnnpack_benchmark(
841 name = "f32_conv_hwc_bench",
842 srcs = [
843 "bench/f32-conv-hwc.cc",
844 "bench/dconv.h",
845 "src/xnnpack/AlignedAllocator.h",
846 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
847 copts = ["-Wno-unused-function"],
848 deps = MICROKERNEL_BENCHMARK_DEPS,
849)
850
851xnnpack_benchmark(
852 name = "f32_dwconv_bench",
853 srcs = [
854 "bench/f32-dwconv.cc",
855 "bench/dwconv.h",
856 "src/xnnpack/AlignedAllocator.h",
857 ] + MICROKERNEL_BENCHMARK_HDRS,
858 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
859)
860
861xnnpack_benchmark(
862 name = "f32_dwconv_spchw_bench",
863 srcs = [
864 "bench/f32-dwconv-spchw.cc",
865 "bench/dwconv.h",
866 "src/xnnpack/AlignedAllocator.h",
867 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
868 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
869)
870
871xnnpack_benchmark(
872 name = "f32_gemm_bench",
873 srcs = [
874 "bench/f32-gemm.cc",
875 "bench/gemm.h",
876 "src/xnnpack/AlignedAllocator.h",
877 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
878 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -0700879 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700880)
881
882xnnpack_benchmark(
883 name = "f32_rmax_bench",
884 srcs = [
885 "bench/f32-rmax.cc",
886 "src/xnnpack/AlignedAllocator.h",
887 ] + MICROKERNEL_BENCHMARK_HDRS,
888 deps = MICROKERNEL_BENCHMARK_DEPS,
889)
890
891xnnpack_benchmark(
892 name = "f32_spmm_bench",
893 srcs = [
894 "bench/f32-spmm.cc",
895 "bench/gemm.h",
896 "src/xnnpack/AlignedAllocator.h",
897 ] + MICROKERNEL_BENCHMARK_HDRS,
898 copts = ["-Wno-unused-function"],
899 deps = MICROKERNEL_BENCHMARK_DEPS,
900)
901
902xnnpack_benchmark(
903 name = "f32_im2col_gemm_bench",
904 srcs = [
905 "bench/f32-im2col-gemm.cc",
906 "bench/conv.h",
907 "src/xnnpack/AlignedAllocator.h",
908 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
909 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
910)
911
912########################### Benchmarks for operators ###########################
913
914xnnpack_benchmark(
915 name = "add_bench",
916 srcs = ["bench/add.cc"],
917 deps = OPERATOR_BENCHMARK_DEPS,
918)
919
920xnnpack_benchmark(
921 name = "average_pooling_bench",
922 srcs = ["bench/average-pooling.cc"],
923 deps = OPERATOR_BENCHMARK_DEPS,
924)
925
926xnnpack_benchmark(
927 name = "channel_shuffle_bench",
928 srcs = ["bench/channel-shuffle.cc"],
929 deps = OPERATOR_BENCHMARK_DEPS,
930)
931
932xnnpack_benchmark(
933 name = "convolution_bench",
934 srcs = ["bench/convolution.cc"],
935 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
936 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
937)
938
939xnnpack_benchmark(
940 name = "deconvolution_bench",
941 srcs = ["bench/deconvolution.cc"],
942 copts = xnnpack_optional_tflite_copts(),
943 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
944)
945
946xnnpack_benchmark(
947 name = "global_average_pooling_bench",
948 srcs = ["bench/global-average-pooling.cc"],
949 deps = OPERATOR_BENCHMARK_DEPS,
950)
951
952xnnpack_benchmark(
953 name = "max_pooling_bench",
954 srcs = ["bench/max-pooling.cc"],
955 deps = OPERATOR_BENCHMARK_DEPS,
956)
957
958xnnpack_benchmark(
959 name = "sigmoid_bench",
960 srcs = ["bench/sigmoid.cc"],
961 deps = OPERATOR_BENCHMARK_DEPS,
962)
963
964xnnpack_benchmark(
965 name = "softargmax_bench",
966 srcs = ["bench/softargmax.cc"],
967 deps = OPERATOR_BENCHMARK_DEPS,
968)
969
Marat Dukhanc068bb62019-10-04 13:24:39 -0700970############################# End-to-end benchmarks ############################
971
972cc_library(
973 name = "mobilenet_v1",
974 srcs = ["models/mobilenet-v1.cc"],
975 hdrs = ["models/models.h"],
976 linkstatic = True,
977 deps = [
978 ":XNNPACK",
979 "@pthreadpool",
980 ],
981)
982
983cc_library(
984 name = "mobilenet_v2",
985 srcs = ["models/mobilenet-v2.cc"],
986 hdrs = ["models/models.h"],
987 linkstatic = True,
988 deps = [
989 ":XNNPACK",
990 "@pthreadpool",
991 ],
992)
993
994xnnpack_benchmark(
995 name = "end2end_bench",
996 srcs = ["bench/end2end.cc"],
997 deps = [
998 ":XNNPACK",
999 ":mobilenet_v1",
1000 ":mobilenet_v2",
1001 "@pthreadpool",
1002 ],
1003)
1004
Marat Dukhan08c4a432019-10-03 09:29:21 -07001005######################### Unit tests for micro-kernels #########################
1006
1007xnnpack_unit_test(
1008 name = "f16_gemm_test",
1009 srcs = [
1010 "test/f16-gemm.cc",
1011 "test/gemm-microkernel-tester.h",
1012 "src/xnnpack/AlignedAllocator.h",
1013 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1014 deps = MICROKERNEL_TEST_DEPS,
1015)
1016
1017xnnpack_unit_test(
1018 name = "f32_argmaxpool_test",
1019 srcs = [
1020 "test/f32-argmaxpool.cc",
1021 "test/argmaxpool-microkernel-tester.h",
1022 "src/xnnpack/AlignedAllocator.h",
1023 ] + MICROKERNEL_TEST_HDRS,
1024 deps = MICROKERNEL_TEST_DEPS,
1025)
1026
1027xnnpack_unit_test(
1028 name = "f32_avgpool_test",
1029 srcs = [
1030 "test/f32-avgpool.cc",
1031 "test/avgpool-microkernel-tester.h",
1032 "src/xnnpack/AlignedAllocator.h",
1033 ] + MICROKERNEL_TEST_HDRS,
1034 deps = MICROKERNEL_TEST_DEPS,
1035)
1036
1037xnnpack_unit_test(
1038 name = "f32_clamp_test",
1039 srcs = [
1040 "test/f32-clamp.cc",
1041 "test/clamp-microkernel-tester.h",
1042 ] + MICROKERNEL_TEST_HDRS,
1043 deps = MICROKERNEL_TEST_DEPS,
1044)
1045
1046xnnpack_unit_test(
1047 name = "f32_igemm_test",
1048 srcs = [
1049 "test/f32-igemm.cc",
1050 "test/gemm-microkernel-tester.h",
1051 "src/xnnpack/AlignedAllocator.h",
1052 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1053 deps = MICROKERNEL_TEST_DEPS,
1054)
1055
1056xnnpack_unit_test(
1057 name = "f32_conv_hwc_test",
1058 srcs = [
1059 "test/f32-conv-hwc.cc",
1060 "test/conv-hwc-microkernel-tester.h",
1061 "src/xnnpack/AlignedAllocator.h",
1062 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1063 deps = MICROKERNEL_TEST_DEPS,
1064)
1065
1066xnnpack_unit_test(
1067 name = "f32_conv_hwc2spchw_test",
1068 srcs = [
1069 "test/f32-conv-hwc2spchw.cc",
1070 "test/conv-hwc2spchw-microkernel-tester.h",
1071 "src/xnnpack/AlignedAllocator.h",
1072 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1073 deps = MICROKERNEL_TEST_DEPS,
1074)
1075
1076xnnpack_unit_test(
1077 name = "f32_dwconv_test",
1078 srcs = [
1079 "test/f32-dwconv.cc",
1080 "test/dwconv-microkernel-tester.h",
1081 "src/xnnpack/AlignedAllocator.h",
1082 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1083 deps = MICROKERNEL_TEST_DEPS,
1084)
1085
1086xnnpack_unit_test(
1087 name = "f32_dwconv_spchw_test",
1088 srcs = [
1089 "test/f32-dwconv-spchw.cc",
1090 "test/dwconv-spchw-microkernel-tester.h",
1091 "src/xnnpack/AlignedAllocator.h",
1092 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1093 deps = MICROKERNEL_TEST_DEPS,
1094)
1095
1096xnnpack_unit_test(
1097 name = "f32_gavgpool_test",
1098 srcs = [
1099 "test/f32-gavgpool.cc",
1100 "test/gavgpool-microkernel-tester.h",
1101 "src/xnnpack/AlignedAllocator.h",
1102 ] + MICROKERNEL_TEST_HDRS,
1103 deps = MICROKERNEL_TEST_DEPS,
1104)
1105
1106xnnpack_unit_test(
1107 name = "f32_gavgpool_spchw_test",
1108 srcs = [
1109 "test/f32-gavgpool-spchw.cc",
1110 "test/gavgpool-spchw-microkernel-tester.h",
1111 "src/xnnpack/AlignedAllocator.h",
1112 ] + MICROKERNEL_TEST_HDRS,
1113 deps = MICROKERNEL_TEST_DEPS,
1114)
1115
1116xnnpack_unit_test(
1117 name = "f32_gemm_test",
1118 srcs = [
1119 "test/f32-gemm.cc",
1120 "test/gemm-microkernel-tester.h",
1121 "src/xnnpack/AlignedAllocator.h",
1122 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1123 deps = MICROKERNEL_TEST_DEPS,
1124)
1125
1126xnnpack_unit_test(
1127 name = "f32_gemminc_test",
1128 srcs = [
1129 "test/f32-gemminc.cc",
1130 "test/gemm-microkernel-tester.h",
1131 "src/xnnpack/AlignedAllocator.h",
1132 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1133 deps = MICROKERNEL_TEST_DEPS,
1134)
1135
1136xnnpack_unit_test(
1137 name = "f32_hswish_test",
1138 srcs = [
1139 "test/f32-hswish.cc",
1140 "test/hswish-microkernel-tester.h",
1141 ] + MICROKERNEL_TEST_HDRS,
1142 deps = MICROKERNEL_TEST_DEPS,
1143)
1144
1145xnnpack_unit_test(
1146 name = "f32_maxpool_test",
1147 srcs = [
1148 "test/f32-maxpool.cc",
1149 "test/maxpool-microkernel-tester.h",
1150 ] + MICROKERNEL_TEST_HDRS,
1151 deps = MICROKERNEL_TEST_DEPS,
1152)
1153
1154xnnpack_unit_test(
1155 name = "f32_pavgpool_test",
1156 srcs = [
1157 "test/f32-pavgpool.cc",
1158 "test/avgpool-microkernel-tester.h",
1159 "src/xnnpack/AlignedAllocator.h",
1160 ] + MICROKERNEL_TEST_HDRS,
1161 deps = MICROKERNEL_TEST_DEPS,
1162)
1163
1164xnnpack_unit_test(
1165 name = "f32_ppmm_test",
1166 srcs = [
1167 "test/f32-ppmm.cc",
1168 "test/gemm-microkernel-tester.h",
1169 "src/xnnpack/AlignedAllocator.h",
1170 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1171 deps = MICROKERNEL_TEST_DEPS,
1172)
1173
1174xnnpack_unit_test(
1175 name = "f32_prelu_test",
1176 srcs = [
1177 "test/f32-prelu.cc",
1178 "test/prelu-microkernel-tester.h",
1179 "src/xnnpack/AlignedAllocator.h",
1180 ] + MICROKERNEL_TEST_HDRS,
1181 deps = MICROKERNEL_TEST_DEPS,
1182)
1183
1184xnnpack_unit_test(
1185 name = "f32_rmax_test",
1186 srcs = [
1187 "test/f32-rmax.cc",
1188 "test/rmax-microkernel-tester.h",
1189 ] + MICROKERNEL_TEST_HDRS,
1190 deps = MICROKERNEL_TEST_DEPS,
1191)
1192
1193xnnpack_unit_test(
1194 name = "f32_spmm_test",
1195 srcs = [
1196 "test/f32-spmm.cc",
1197 "test/spmm-microkernel-tester.h",
1198 "src/xnnpack/AlignedAllocator.h",
1199 ] + MICROKERNEL_TEST_HDRS,
1200 deps = MICROKERNEL_TEST_DEPS,
1201)
1202
1203xnnpack_unit_test(
1204 name = "f32_vadd_test",
1205 srcs = [
1206 "test/f32-vadd.cc",
1207 "test/vadd-microkernel-tester.h",
1208 ] + MICROKERNEL_TEST_HDRS,
1209 deps = MICROKERNEL_TEST_DEPS,
1210)
1211
1212xnnpack_unit_test(
1213 name = "f32_vsub_test",
1214 srcs = [
1215 "test/f32-vsub.cc",
1216 "test/vsub-microkernel-tester.h",
1217 ] + MICROKERNEL_TEST_HDRS,
1218 deps = MICROKERNEL_TEST_DEPS,
1219)
1220
1221xnnpack_unit_test(
1222 name = "f32_vmul_test",
1223 srcs = [
1224 "test/f32-vmul.cc",
1225 "test/vmul-microkernel-tester.h",
1226 ] + MICROKERNEL_TEST_HDRS,
1227 deps = MICROKERNEL_TEST_DEPS,
1228)
1229
1230xnnpack_unit_test(
1231 name = "f32_vmulcaddc_test",
1232 srcs = [
1233 "test/f32-vmulcaddc.cc",
1234 "test/vmulcaddc-microkernel-tester.h",
1235 "src/xnnpack/AlignedAllocator.h",
1236 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1237 deps = MICROKERNEL_TEST_DEPS,
1238)
1239
1240xnnpack_unit_test(
1241 name = "q8_avgpool_test",
1242 srcs = [
1243 "test/q8-avgpool.cc",
1244 "test/avgpool-microkernel-tester.h",
1245 "src/xnnpack/AlignedAllocator.h",
1246 ] + MICROKERNEL_TEST_HDRS,
1247 deps = MICROKERNEL_TEST_DEPS,
1248)
1249
1250xnnpack_unit_test(
1251 name = "q8_igemm_test",
1252 srcs = [
1253 "test/q8-igemm.cc",
1254 "test/gemm-microkernel-tester.h",
1255 "src/xnnpack/AlignedAllocator.h",
1256 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1257 deps = MICROKERNEL_TEST_DEPS,
1258)
1259
1260xnnpack_unit_test(
1261 name = "q8_dwconv_test",
1262 srcs = [
1263 "test/q8-dwconv.cc",
1264 "test/dwconv-microkernel-tester.h",
1265 "src/xnnpack/AlignedAllocator.h",
1266 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1267 deps = MICROKERNEL_TEST_DEPS,
1268)
1269
1270xnnpack_unit_test(
1271 name = "q8_gavgpool_test",
1272 srcs = [
1273 "test/q8-gavgpool.cc",
1274 "test/gavgpool-microkernel-tester.h",
1275 "src/xnnpack/AlignedAllocator.h",
1276 ] + MICROKERNEL_TEST_HDRS,
1277 deps = MICROKERNEL_TEST_DEPS,
1278)
1279
1280xnnpack_unit_test(
1281 name = "q8_gemm_test",
1282 srcs = [
1283 "test/q8-gemm.cc",
1284 "test/gemm-microkernel-tester.h",
1285 "src/xnnpack/AlignedAllocator.h",
1286 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1287 deps = MICROKERNEL_TEST_DEPS,
1288)
1289
1290xnnpack_unit_test(
1291 name = "q8_vadd_test",
1292 srcs = [
1293 "test/q8-vadd.cc",
1294 "test/vadd-microkernel-tester.h",
1295 ] + MICROKERNEL_TEST_HDRS,
1296 deps = MICROKERNEL_TEST_DEPS,
1297)
1298
1299xnnpack_unit_test(
1300 name = "u8_clamp_test",
1301 srcs = [
1302 "test/u8-clamp.cc",
1303 "test/clamp-microkernel-tester.h",
1304 ] + MICROKERNEL_TEST_HDRS,
1305 deps = MICROKERNEL_TEST_DEPS,
1306)
1307
1308xnnpack_unit_test(
1309 name = "u8_lut32norm_test",
1310 srcs = [
1311 "test/u8-lut32norm.cc",
1312 "test/lut-norm-microkernel-tester.h",
1313 ] + MICROKERNEL_TEST_HDRS,
1314 deps = MICROKERNEL_TEST_DEPS,
1315)
1316
1317xnnpack_unit_test(
1318 name = "u8_maxpool_test",
1319 srcs = [
1320 "test/u8-maxpool.cc",
1321 "test/maxpool-microkernel-tester.h",
1322 ] + MICROKERNEL_TEST_HDRS,
1323 deps = MICROKERNEL_TEST_DEPS,
1324)
1325
1326xnnpack_unit_test(
1327 name = "u8_rmax_test",
1328 srcs = [
1329 "test/u8-rmax.cc",
1330 "test/rmax-microkernel-tester.h",
1331 ] + MICROKERNEL_TEST_HDRS,
1332 deps = MICROKERNEL_TEST_DEPS,
1333)
1334
1335xnnpack_unit_test(
1336 name = "x32_packx_test",
1337 srcs = [
1338 "test/x32-packx.cc",
1339 "test/pack-microkernel-tester.h",
1340 "src/xnnpack/AlignedAllocator.h",
1341 ] + MICROKERNEL_TEST_HDRS,
1342 deps = MICROKERNEL_TEST_DEPS,
1343)
1344
1345xnnpack_unit_test(
1346 name = "x32_pad_test",
1347 srcs = [
1348 "test/x32-pad.cc",
1349 "test/pad-microkernel-tester.h",
1350 ] + MICROKERNEL_TEST_HDRS,
1351 deps = MICROKERNEL_TEST_DEPS,
1352)
1353
1354xnnpack_unit_test(
1355 name = "x32_unpool_test",
1356 srcs = [
1357 "test/x32-unpool.cc",
1358 "test/unpool-microkernel-tester.h",
1359 ] + MICROKERNEL_TEST_HDRS,
1360 deps = MICROKERNEL_TEST_DEPS,
1361)
1362
1363xnnpack_unit_test(
1364 name = "x32_zip_test",
1365 srcs = [
1366 "test/x32-zip.cc",
1367 "test/zip-microkernel-tester.h",
1368 ] + MICROKERNEL_TEST_HDRS,
1369 deps = MICROKERNEL_TEST_DEPS,
1370)
1371
1372xnnpack_unit_test(
1373 name = "x8_lut_test",
1374 srcs = [
1375 "test/x8-lut.cc",
1376 "test/lut-microkernel-tester.h",
1377 ] + MICROKERNEL_TEST_HDRS,
1378 deps = MICROKERNEL_TEST_DEPS,
1379)
1380
1381xnnpack_unit_test(
1382 name = "x8_zip_test",
1383 srcs = [
1384 "test/x8-zip.cc",
1385 "test/zip-microkernel-tester.h",
1386 ] + MICROKERNEL_TEST_HDRS,
1387 deps = MICROKERNEL_TEST_DEPS,
1388)
1389
1390########################### Size test for the library ##########################
1391
1392xnnpack_binary(
1393 name = "size_test",
1394 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001395 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001396)
1397
1398########################### Unit tests for operators ###########################
1399
1400xnnpack_unit_test(
1401 name = "add_test",
1402 srcs = [
1403 "test/add.cc",
1404 "test/add-operator-tester.h",
1405 ],
1406 deps = OPERATOR_TEST_DEPS,
1407)
1408
1409xnnpack_unit_test(
1410 name = "argmax_pooling_test",
1411 srcs = [
1412 "test/argmax-pooling.cc",
1413 "test/argmax-pooling-operator-tester.h",
1414 ] + OPERATOR_TEST_PARAMS_HDRS,
1415 deps = OPERATOR_TEST_DEPS,
1416)
1417
1418xnnpack_unit_test(
1419 name = "average_pooling_test",
1420 srcs = [
1421 "test/average-pooling.cc",
1422 "test/average-pooling-operator-tester.h",
1423 ] + OPERATOR_TEST_PARAMS_HDRS,
1424 deps = OPERATOR_TEST_DEPS,
1425)
1426
1427xnnpack_unit_test(
1428 name = "channel_pad_test",
1429 srcs = [
1430 "test/channel-pad.cc",
1431 "test/channel-pad-operator-tester.h",
1432 ] + OPERATOR_TEST_PARAMS_HDRS,
1433 deps = OPERATOR_TEST_DEPS,
1434)
1435
1436xnnpack_unit_test(
1437 name = "channel_shuffle_test",
1438 srcs = [
1439 "test/channel-shuffle.cc",
1440 "test/channel-shuffle-operator-tester.h",
1441 ],
1442 deps = OPERATOR_TEST_DEPS,
1443)
1444
1445xnnpack_unit_test(
1446 name = "clamp_test",
1447 srcs = [
1448 "test/clamp.cc",
1449 "test/clamp-operator-tester.h",
1450 ],
1451 deps = OPERATOR_TEST_DEPS,
1452)
1453
1454xnnpack_unit_test(
1455 name = "convolution_test",
1456 srcs = [
1457 "test/convolution.cc",
1458 "test/convolution-operator-tester.h",
1459 ],
1460 deps = OPERATOR_TEST_DEPS,
1461)
1462
1463xnnpack_unit_test(
1464 name = "convolution_spnchw_test",
1465 srcs = [
1466 "test/convolution-spnchw.cc",
1467 "test/convolution-spnchw-operator-tester.h",
1468 ],
1469 deps = OPERATOR_TEST_DEPS,
1470)
1471
1472xnnpack_unit_test(
1473 name = "deconvolution_test",
1474 srcs = [
1475 "test/deconvolution.cc",
1476 "test/deconvolution-operator-tester.h",
1477 ] + OPERATOR_TEST_PARAMS_HDRS,
1478 deps = OPERATOR_TEST_DEPS,
1479)
1480
1481xnnpack_unit_test(
1482 name = "fully_connected_test",
1483 srcs = [
1484 "test/fully-connected.cc",
1485 "test/fully-connected-operator-tester.h",
1486 ],
1487 deps = OPERATOR_TEST_DEPS,
1488)
1489
1490xnnpack_unit_test(
1491 name = "global_average_pooling_test",
1492 srcs = [
1493 "test/global-average-pooling.cc",
1494 "test/global-average-pooling-operator-tester.h",
1495 ] + OPERATOR_TEST_PARAMS_HDRS,
1496 deps = OPERATOR_TEST_DEPS,
1497)
1498
1499xnnpack_unit_test(
1500 name = "global_average_pooling_spnchw_test",
1501 srcs = [
1502 "test/global-average-pooling-spnchw.cc",
1503 "test/global-average-pooling-spnchw-operator-tester.h",
1504 ],
1505 deps = OPERATOR_TEST_DEPS,
1506)
1507
1508xnnpack_unit_test(
1509 name = "hardswish_test",
1510 srcs = [
1511 "test/hardswish.cc",
1512 "test/hardswish-operator-tester.h",
1513 ],
1514 deps = OPERATOR_TEST_DEPS,
1515)
1516
1517xnnpack_unit_test(
1518 name = "leaky_relu_test",
1519 srcs = [
1520 "test/leaky-relu.cc",
1521 "test/leaky-relu-operator-tester.h",
1522 ],
1523 deps = OPERATOR_TEST_DEPS,
1524)
1525
1526xnnpack_unit_test(
1527 name = "max_pooling_test",
1528 srcs = [
1529 "test/max-pooling.cc",
1530 "test/max-pooling-operator-tester.h",
1531 ] + OPERATOR_TEST_PARAMS_HDRS,
1532 deps = OPERATOR_TEST_DEPS,
1533)
1534
1535xnnpack_unit_test(
1536 name = "prelu_test",
1537 srcs = [
1538 "test/prelu.cc",
1539 "test/prelu-operator-tester.h",
1540 ] + OPERATOR_TEST_PARAMS_HDRS,
1541 deps = OPERATOR_TEST_DEPS,
1542)
1543
1544xnnpack_unit_test(
1545 name = "sigmoid_test",
1546 srcs = [
1547 "test/sigmoid.cc",
1548 "test/sigmoid-operator-tester.h",
1549 ],
1550 deps = OPERATOR_TEST_DEPS,
1551)
1552
1553xnnpack_unit_test(
1554 name = "softargmax_test",
1555 srcs = [
1556 "test/softargmax.cc",
1557 "test/softargmax-operator-tester.h",
1558 ],
1559 deps = OPERATOR_TEST_DEPS,
1560)
1561
1562xnnpack_unit_test(
1563 name = "unpooling_test",
1564 srcs = [
1565 "test/unpooling.cc",
1566 "test/unpooling-operator-tester.h",
1567 ],
1568 deps = OPERATOR_TEST_DEPS,
1569)
1570
1571############################# Build configurations #############################
1572
1573config_setting(
1574 name = "linux_k8",
1575 values = {
1576 "cpu": "k8",
1577 },
1578)
1579
1580config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07001581 name = "linux_aarch64",
1582 values = {
1583 "cpu": "aarch64",
1584 },
1585)
1586
1587config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001588 name = "android",
1589 values = {"crosstool_top": "//external:android/crosstool"},
1590)
1591
1592config_setting(
1593 name = "android_armv7",
1594 values = {
1595 "crosstool_top": "//external:android/crosstool",
1596 "cpu": "armeabi-v7a",
1597 },
1598)
1599
1600config_setting(
1601 name = "android_arm64",
1602 values = {
1603 "crosstool_top": "//external:android/crosstool",
1604 "cpu": "arm64-v8a",
1605 },
1606)
1607
1608config_setting(
1609 name = "android_x86",
1610 values = {
1611 "crosstool_top": "//external:android/crosstool",
1612 "cpu": "x86",
1613 },
1614)
1615
1616config_setting(
1617 name = "android_x86_64",
1618 values = {
1619 "crosstool_top": "//external:android/crosstool",
1620 "cpu": "x86_64",
1621 },
1622)
1623
1624config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07001625 name = "macos_x86_64",
1626 values = {
1627 "apple_platform_type": "macos",
1628 "cpu": "darwin",
1629 },
1630)
1631
1632config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001633 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001634 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07001635)
1636
1637config_setting(
1638 name = "emscripten_wasm",
1639 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001640 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001641 "cpu": "wasm",
1642 },
1643)
1644
1645config_setting(
1646 name = "emscripten_wasmsimd",
1647 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001648 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001649 "cpu": "wasm",
1650 "features": "wasmsimd",
1651 },
1652)
1653
1654config_setting(
1655 name = "emscripten_asmjs",
1656 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001657 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001658 "cpu": "asmjs",
1659 },
1660)
1661
Marat Dukhan174706e2019-10-09 16:33:53 -07001662# Builds with -c dbg
Marat Dukhan08c4a432019-10-03 09:29:21 -07001663config_setting(
1664 name = "debug_build",
1665 values = {
1666 "compilation_mode": "dbg",
1667 },
Marat Dukhan08c4a432019-10-03 09:29:21 -07001668)
1669
Marat Dukhan174706e2019-10-09 16:33:53 -07001670# Builds with -c opt
Marat Dukhan08c4a432019-10-03 09:29:21 -07001671config_setting(
1672 name = "optimized_build",
1673 values = {
1674 "compilation_mode": "opt",
1675 },
Marat Dukhan08c4a432019-10-03 09:29:21 -07001676)
1677
1678# Enables usage of assembly kernels.
1679config_setting(
1680 name = "xnn_enable_assembly_explicit_true",
1681 define_values = {"xnn_enable_assembly": "true"},
1682)
1683
1684# Disables usage of assembly kernels.
1685config_setting(
1686 name = "xnn_enable_assembly_explicit_false",
1687 define_values = {"xnn_enable_assembly": "false"},
1688)