blob: b450b7238caafa3ce20218f63ac4f8a27d1edb65 [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
9licenses(["notice"])
10
11exports_files(["LICENSE"])
12
13load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_min_size_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_copts", "xnnpack_unit_test", "xnnpack_visibility")
14
15OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
19 "@pthreadpool",
20]
21
22MICROKERNEL_BENCHMARK_DEPS = [
23 ":ukernels",
24 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070025 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070026 "@cpuinfo",
27 "@FP16",
28 "@pthreadpool",
29]
30
Marat Dukhan6adff4e2019-10-14 18:32:07 -070031ACCURACY_EVAL_DEPS = [
32 ":XNNPACK",
33 ":ukernels",
34 "@FP16",
35 "@pthreadpool",
36]
37
Marat Dukhan08c4a432019-10-03 09:29:21 -070038MICROKERNEL_TEST_DEPS = [
39 ":ukernels",
Frank Barchard7e955972019-10-11 10:34:25 -070040 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070041 "@cpuinfo",
42 "@FP16",
43 "@pthreadpool",
44]
45
46OPERATOR_TEST_DEPS = [
47 ":XNNPACK",
48 "@pthreadpool",
49 "@FP16",
50]
51
52OPERATOR_SRCS = [
53 "src/add.c",
54 "src/argmax-pooling.c",
55 "src/average-pooling.c",
56 "src/channel-pad.c",
57 "src/channel-shuffle.c",
58 "src/clamp.c",
59 "src/convolution-spnchw.c",
60 "src/convolution.c",
61 "src/deconvolution.c",
62 "src/fully-connected.c",
63 "src/global-average-pooling-spnchw.c",
64 "src/global-average-pooling.c",
65 "src/hardswish.c",
66 "src/leaky-relu.c",
67 "src/max-pooling.c",
68 "src/prelu.c",
69 "src/sigmoid.c",
70 "src/softargmax.c",
71 "src/unpooling.c",
72]
73
74SCALAR_UKERNELS = [
75 "src/f32-argmaxpool/mp9p8q-scalar.c",
76 "src/f32-argmaxpool/up4-scalar.c",
77 "src/f32-argmaxpool/up9-scalar.c",
78 "src/f32-avgpool/mp9p8q-scalar.c",
79 "src/f32-avgpool/up9-scalar.c",
80 "src/f32-clamp/scalar.c",
Erich Elsen563df5f2019-10-23 08:02:21 -070081 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070082 "src/f32-igemm/1x4-scalar.c",
83 "src/f32-igemm/2x4-scalar.c",
84 "src/f32-igemm/4x2-scalar.c",
85 "src/f32-igemm/4x4-scalar.c",
86 "src/f32-dwconv/up1x25-scalar.c",
87 "src/f32-dwconv/up1x4-scalar.c",
88 "src/f32-dwconv/up1x9-scalar.c",
Erich Elsen0cc2c532019-10-15 04:44:18 -070089 "src/f32-dwconv-spchw/3x3p1-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -070090 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -070091 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070092 "src/f32-gavgpool/mp7p7q-scalar.c",
93 "src/f32-gavgpool/up7-scalar.c",
94 "src/f32-gemm/1x4-scalar.c",
95 "src/f32-gemm/2x4-scalar.c",
96 "src/f32-gemm/4x2-scalar.c",
97 "src/f32-gemm/4x4-scalar.c",
98 "src/f32-gemminc/1x4-scalar.c",
99 "src/f32-gemminc/2x4-scalar.c",
100 "src/f32-gemminc/4x4-scalar.c",
101 "src/f32-hswish/scalar.c",
102 "src/f32-maxpool/9p8q-scalar.c",
103 "src/f32-pavgpool/mp9p8q-scalar.c",
104 "src/f32-pavgpool/up9-scalar.c",
105 "src/f32-ppmm/2x4-scalar.c",
106 "src/f32-ppmm/3x3-scalar.c",
107 "src/f32-ppmm/4x2-scalar.c",
108 "src/f32-ppmm/4x4-scalar.c",
109 "src/f32-prelu/x4-scalar.c",
110 "src/f32-rmax/scalar.c",
111 "src/f32-spmm/1x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700112 "src/f32-spmm/1x1-scalar.c",
113 "src/f32-spmm/2x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700114 "src/f32-spmm/2x1-scalar.c",
115 "src/f32-spmm/4x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700116 "src/f32-spmm/4x1-scalar.c",
117 "src/f32-spmm/8x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700118 "src/f32-spmm/8x1-scalar.c",
Erich Elsenc6afd9b2019-10-24 16:10:53 -0700119 "src/f32-spmm/8x2-scalar.c",
120 "src/f32-spmm/8x4-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700121 "src/f32-vadd/scalar.c",
122 "src/f32-vmul/scalar.c",
123 "src/f32-vmulcaddc/c1-scalar-x2.c",
124 "src/f32-vsub/scalar.c",
125 "src/q8-avgpool/mp9p8q-scalar.c",
126 "src/q8-avgpool/up9-scalar.c",
127 "src/q8-igemm/2x2-scalar.c",
128 "src/q8-dwconv/up1x9-scalar.c",
129 "src/q8-gavgpool/mp7p7q-scalar.c",
130 "src/q8-gavgpool/up7-scalar.c",
131 "src/q8-gemm/2x2-scalar.c",
132 "src/q8-vadd/scalar.c",
133 "src/u8-clamp/scalar.c",
134 "src/u8-lut32norm/scalar.c",
135 "src/u8-maxpool/9p8q-scalar.c",
136 "src/u8-rmax/scalar.c",
137 "src/x32-packx/x2-scalar.c",
138 "src/x32-packx/x3-scalar.c",
139 "src/x32-packx/x4-scalar.c",
140 "src/x32-pad/x2-scalar.c",
141 "src/x32-unpool/scalar.c",
142 "src/x32-zip/x2-scalar.c",
143 "src/x32-zip/x3-scalar.c",
144 "src/x32-zip/x4-scalar.c",
145 "src/x32-zip/xm-scalar.c",
146 "src/x8-lut/scalar.c",
147 "src/x8-zip/x2-scalar.c",
148 "src/x8-zip/x3-scalar.c",
149 "src/x8-zip/x4-scalar.c",
150 "src/x8-zip/xm-scalar.c",
151]
152
153PSIMD_UKERNELS = [
154 "src/f32-argmaxpool/mp9p8q-psimd.c",
155 "src/f32-argmaxpool/up4-psimd.c",
156 "src/f32-argmaxpool/up9-psimd.c",
157 "src/f32-avgpool/mp9p8q-psimd.c",
158 "src/f32-avgpool/up9-psimd.c",
159 "src/f32-clamp/psimd.c",
160 "src/f32-igemm/1x8-psimd-loadsplat.c",
161 "src/f32-igemm/1x8-psimd-splat.c",
162 "src/f32-igemm/1x8s4-psimd.c",
163 "src/f32-igemm/4x2c4-psimd.c",
164 "src/f32-igemm/4x8-psimd-loadsplat.c",
165 "src/f32-igemm/4x8-psimd-splat.c",
166 "src/f32-igemm/4x8s4-psimd.c",
167 "src/f32-igemm/6x8-psimd-loadsplat.c",
168 "src/f32-igemm/6x8-psimd-splat.c",
169 "src/f32-igemm/6x8s4-psimd.c",
170 "src/f32-dwconv/up4x25-psimd.c",
171 "src/f32-dwconv/up4x4-psimd.c",
172 "src/f32-dwconv/up4x9-psimd.c",
173 "src/f32-gavgpool/mp7p7q-psimd.c",
174 "src/f32-gavgpool/up7-psimd.c",
175 "src/f32-gemm/1x8-psimd-loadsplat.c",
176 "src/f32-gemm/1x8-psimd-splat.c",
177 "src/f32-gemm/1x8s4-psimd.c",
178 "src/f32-gemm/4x8-psimd-loadsplat.c",
179 "src/f32-gemm/4x8-psimd-splat.c",
180 "src/f32-gemm/4x8s4-psimd.c",
181 "src/f32-gemm/6x8-psimd-loadsplat.c",
182 "src/f32-gemm/6x8-psimd-splat.c",
183 "src/f32-gemm/6x8s4-psimd.c",
184 "src/f32-gemminc/1x8-psimd-loadsplat.c",
185 "src/f32-gemminc/1x8-psimd-splat.c",
186 "src/f32-gemminc/1x8s4-psimd.c",
187 "src/f32-gemminc/4x8-psimd-loadsplat.c",
188 "src/f32-gemminc/4x8-psimd-splat.c",
189 "src/f32-gemminc/4x8s4-psimd.c",
190 "src/f32-gemminc/6x8-psimd-loadsplat.c",
191 "src/f32-gemminc/6x8-psimd-splat.c",
192 "src/f32-gemminc/6x8s4-psimd.c",
193 "src/f32-hswish/psimd.c",
194 "src/f32-maxpool/9p8q-psimd.c",
195 "src/f32-pavgpool/mp9p8q-psimd.c",
196 "src/f32-pavgpool/up9-psimd.c",
197 "src/f32-ppmm/4x8-psimd.c",
198 "src/f32-prelu/x4-psimd.c",
199 "src/f32-vadd/psimd.c",
200 "src/f32-vmul/psimd.c",
201 "src/f32-vmulcaddc/c4-psimd-x2.c",
202 "src/f32-vsub/psimd.c",
203 "src/x32-packx/x4-psimd.c",
204 "src/x32-pad/x2-psimd.c",
205 "src/x32-unpool/psimd.c",
206 "src/x32-zip/x2-psimd.c",
207 "src/x32-zip/x3-psimd.c",
208 "src/x32-zip/x4-psimd.c",
209 "src/x32-zip/xm-psimd.c",
210]
211
212# ISA-specific micro-kernels
213NEON_UKERNELS = [
214 "src/f32-avgpool/mp9p8q-neon.c",
215 "src/f32-avgpool/up9-neon.c",
216 "src/f32-clamp/neon.c",
217 "src/f32-igemm/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700218 "src/f32-igemm/4x2-neon-ld64.c",
219 "src/f32-igemm/4x4-neon-ld64.c",
220 "src/f32-igemm/4x8-neon-ld128.c",
221 "src/f32-igemm/4x8-neon-ld64.c",
222 "src/f32-igemm/6x8-neon-ld64.c",
223 "src/f32-dwconv/up4x9-neon.c",
224 "src/f32-gavgpool-spchw/neon-x4.c",
225 "src/f32-gavgpool/mp7p7q-neon.c",
226 "src/f32-gavgpool/up7-neon.c",
227 "src/f32-gemm/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700228 "src/f32-gemm/4x2-neon-ld64.c",
229 "src/f32-gemm/4x8-neon-ld128.c",
230 "src/f32-gemm/4x8-neon-ld64.c",
231 "src/f32-gemm/5x8-neon-ld64.c",
232 "src/f32-gemm/6x8-neon-ld64.c",
233 "src/f32-gemminc/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700234 "src/f32-gemminc/4x8-neon-ld128.c",
235 "src/f32-gemminc/4x8-neon-ld64.c",
236 "src/f32-gemminc/5x8-neon-ld64.c",
237 "src/f32-gemminc/6x8-neon-ld64.c",
238 "src/f32-hswish/neon.c",
239 "src/f32-pavgpool/mp9p8q-neon.c",
240 "src/f32-pavgpool/up9-neon.c",
241 "src/f32-ppmm/4x8-neon.c",
242 "src/f32-ppmm/8x8-neon.c",
243 "src/f32-rmax/neon.c",
244 "src/f32-vmulcaddc/c4-neon-x2.c",
245 "src/q8-avgpool/mp9p8q-neon.c",
246 "src/q8-avgpool/up9-neon.c",
247 "src/q8-igemm/4x8-neon.c",
248 "src/q8-igemm/8x8-neon.c",
249 "src/q8-dwconv/up8x9-neon.c",
250 "src/q8-gavgpool/mp7p7q-neon.c",
251 "src/q8-gavgpool/up7-neon.c",
252 "src/q8-gemm/4x8-neon.c",
253 "src/q8-gemm/8x8-neon.c",
254 "src/q8-vadd/neon.c",
255 "src/u8-clamp/neon.c",
256 "src/u8-maxpool/9p8q-neon.c",
257 "src/u8-rmax/neon.c",
258 "src/x32-packx/x4-neon-st4.c",
259 "src/x32-pad/x2-neon.c",
260 "src/x32-zip/x2-neon.c",
261 "src/x32-zip/x3-neon.c",
262 "src/x32-zip/x4-neon.c",
263 "src/x32-zip/xm-neon.c",
264 "src/x8-zip/x2-neon.c",
265 "src/x8-zip/x3-neon.c",
266 "src/x8-zip/x4-neon.c",
267 "src/x8-zip/xm-neon.c",
268]
269
270NEONFMA_UKERNELS = [
Frank Barcharddb45b6a2019-10-09 16:42:45 -0700271 "src/f32-igemm/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700272 "src/f32-igemm/4x2-neonfma-ld64.c",
273 "src/f32-igemm/4x4-neonfma-ld64.c",
274 "src/f32-igemm/4x8-neonfma-ld128.c",
275 "src/f32-igemm/4x8-neonfma-ld64.c",
276 "src/f32-igemm/6x8-neonfma-ld64.c",
277 "src/f32-dwconv/up4x9-neonfma.c",
278 "src/f32-dwconv/up8x9-neonfma.c",
279 "src/f32-gemm/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700280 "src/f32-gemm/4x2-neonfma-ld64.c",
281 "src/f32-gemm/4x8-neonfma-ld128.c",
282 "src/f32-gemm/4x8-neonfma-ld64.c",
283 "src/f32-gemm/5x8-neonfma-ld64.c",
284 "src/f32-gemm/6x8-neonfma-ld64.c",
285 "src/f32-gemminc/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700286 "src/f32-gemminc/4x8-neonfma-ld128.c",
287 "src/f32-gemminc/4x8-neonfma-ld64.c",
288 "src/f32-gemminc/5x8-neonfma-ld64.c",
289 "src/f32-gemminc/6x8-neonfma-ld64.c",
290 "src/f32-hswish/neonfma.c",
291 "src/f32-ppmm/4x8-neonfma.c",
292 "src/f32-ppmm/8x8-neonfma.c",
293 "src/f32-vmulcaddc/c4-neonfma-x2.c",
294]
295
296AARCH64_NEONFMA_UKERNELS = [
297 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
298 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
299 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
300 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
301 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
302 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
303 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
304 "src/f32-spmm/12x1-neonfma.c",
305 "src/f32-spmm/12x2-neonfma.c",
306 "src/f32-spmm/12x4-neonfma.c",
307 "src/f32-spmm/16x1-neonfma-pipelined.c",
308 "src/f32-spmm/16x1-neonfma-unroll2.c",
309 "src/f32-spmm/16x1-neonfma.c",
310 "src/f32-spmm/16x2-neonfma.c",
311 "src/f32-spmm/16x4-neonfma.c",
312 "src/f32-spmm/4x1-neonfma-pipelined.c",
313 "src/f32-spmm/4x1-neonfma-unroll2.c",
314 "src/f32-spmm/4x1-neonfma.c",
315 "src/f32-spmm/4x2-neonfma.c",
316 "src/f32-spmm/4x4-neonfma.c",
317 "src/f32-spmm/8x1-neonfma-pipelined.c",
318 "src/f32-spmm/8x1-neonfma-unroll2.c",
319 "src/f32-spmm/8x1-neonfma.c",
320 "src/f32-spmm/8x2-neonfma.c",
321 "src/f32-spmm/8x4-neonfma.c",
322]
323
324AARCH64_NEONFP16ARITH_UKERNELS = [
325 "src/f16-gemm/4x8-neonfp16arith-ld64.c",
326 "src/f16-gemm/6x8-neonfp16arith-ld64.c",
327 "src/f16-gemm/8x8-neonfp16arith-ld64.c",
328]
329
330SSE_UKERNELS = [
331 "src/f32-avgpool/mp9p8q-sse.c",
332 "src/f32-avgpool/up9-sse.c",
333 "src/f32-clamp/sse.c",
334 "src/f32-igemm/1x8-sse-dup.c",
335 "src/f32-igemm/1x8-sse-load1.c",
336 "src/f32-igemm/1x8s4-sse.c",
337 "src/f32-igemm/4x2c4-sse.c",
338 "src/f32-igemm/4x8-sse-dup.c",
339 "src/f32-igemm/4x8-sse-load1.c",
340 "src/f32-igemm/4x8s4-sse.c",
341 "src/f32-dwconv/up4x25-sse.c",
342 "src/f32-dwconv/up4x4-sse.c",
343 "src/f32-dwconv/up4x9-sse.c",
344 "src/f32-gavgpool-spchw/sse-x4.c",
345 "src/f32-gavgpool/mp7p7q-sse.c",
346 "src/f32-gavgpool/up7-sse.c",
347 "src/f32-gemm/1x8-sse-dup.c",
348 "src/f32-gemm/1x8-sse-load1.c",
349 "src/f32-gemm/1x8s4-sse.c",
350 "src/f32-gemm/4x8-sse-dup.c",
351 "src/f32-gemm/4x8-sse-load1.c",
352 "src/f32-gemm/4x8s4-sse.c",
353 "src/f32-gemminc/1x8-sse-dup.c",
354 "src/f32-gemminc/1x8-sse-load1.c",
355 "src/f32-gemminc/1x8s4-sse.c",
356 "src/f32-gemminc/4x8-sse-dup.c",
357 "src/f32-gemminc/4x8-sse-load1.c",
358 "src/f32-gemminc/4x8s4-sse.c",
359 "src/f32-hswish/sse.c",
360 "src/f32-maxpool/9p8q-sse.c",
361 "src/f32-pavgpool/mp9p8q-sse.c",
362 "src/f32-pavgpool/up9-sse.c",
363 "src/f32-dwconv-spchw/3x3p1-sse.c",
364 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
365 "src/f32-ppmm/4x8-sse.c",
366 "src/f32-prelu/x4-sse.c",
367 "src/f32-rmax/sse.c",
368 "src/f32-spmm/4x1-sse.c",
369 "src/f32-spmm/8x1-sse.c",
370 "src/f32-vadd/sse.c",
371 "src/f32-vmul/sse.c",
372 "src/f32-vmulcaddc/c4-sse-x2.c",
373 "src/f32-vsub/sse.c",
374 "src/x32-packx/x4-sse.c",
375]
376
377SSE2_UKERNELS = [
378 "src/f32-argmaxpool/mp9p8q-sse2.c",
379 "src/f32-argmaxpool/up4-sse2.c",
380 "src/f32-argmaxpool/up9-sse2.c",
381 "src/q8-avgpool/mp9p8q-sse2.c",
382 "src/q8-avgpool/up9-sse2.c",
383 "src/q8-igemm/4x4c2-sse2.c",
384 "src/q8-dwconv/up8x9-sse2.c",
385 "src/q8-gavgpool/mp7p7q-sse2.c",
386 "src/q8-gavgpool/up7-sse2.c",
387 "src/q8-gemm/2x4c8-sse2.c",
388 "src/q8-gemm/4x4c2-sse2.c",
389 "src/q8-vadd/sse2.c",
390 "src/u8-clamp/sse2.c",
391 "src/u8-maxpool/9p8q-sse2.c",
392 "src/u8-rmax/sse2.c",
393 "src/x32-pad/x2-sse2.c",
394 "src/x32-zip/x2-sse2.c",
395 "src/x32-zip/x3-sse2.c",
396 "src/x32-zip/x4-sse2.c",
397 "src/x32-zip/xm-sse2.c",
398 "src/x8-zip/x2-sse2.c",
399 "src/x8-zip/x3-sse2.c",
400 "src/x8-zip/x4-sse2.c",
401 "src/x8-zip/xm-sse2.c",
402]
403
404AVX_UKERNELS = [
405 "src/f32-rmax/avx.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700406 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700407]
408
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700409AVX2_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700410 "src/f32-raddexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700411 "src/f32-raddextexp/avx2-p5-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700412 "src/f32-raddstoreexpminusmax/avx2-p5-unroll64.c",
413 "src/f32-vscaleexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700414 "src/f32-vscaleextexp/avx2-p5-unroll64.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700415 "src/math/exp-avx2-p5.c",
416 "src/math/exp-avx2-perm-p3.c",
417 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -0700418 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700419 "src/math/extexp-avx2-p5.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700420]
421
Marat Dukhan08c4a432019-10-03 09:29:21 -0700422AVX512F_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700423 "src/f32-raddexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700424 "src/f32-raddextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700425 "src/f32-raddstoreexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700426 "src/f32-rmax/avx512f.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700427 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700428 "src/f32-vscaleexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700429 "src/f32-vscaleextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700430 "src/math/exp-avx512f-p5-scalef.c",
431 "src/math/exp-avx512f-p5.c",
432 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -0700433 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700434 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700435]
436
437AARCH32_ASM_UKERNELS = [
438 "src/q8-dwconv/up8x9-aarch32-neon.S",
439]
440
441AARCH64_ASM_UKERNELS = [
442 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
443 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
444 "src/f32-gemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700445 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700446 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a57.S",
447 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a75.S",
448 "src/f32-gemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700449 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700450 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a57.S",
451 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a75.S",
452 "src/f32-gemm/4x8-aarch64-neonfma-ld128.S",
453 "src/f32-gemm/4x8-aarch64-neonfma-ld64.S",
454 "src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700455 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700456 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a57.S",
457 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S",
458 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S",
459 "src/f32-gemm/6x8-aarch64-neonfma-ld128.S",
460 "src/f32-gemm/6x8-aarch64-neonfma-ld64.S",
461 "src/f32-gemminc/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700462 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700463 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a57.S",
464 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a75.S",
465 "src/f32-gemminc/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700466 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700467 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a57.S",
468 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a75.S",
469 "src/f32-gemminc/4x8-aarch64-neonfma-ld128.S",
470 "src/f32-gemminc/4x8-aarch64-neonfma-ld64.S",
471 "src/f32-gemminc/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700472 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700473 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a57.S",
474 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a73.S",
475 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a75.S",
476 "src/f32-gemminc/6x8-aarch64-neonfma-ld128.S",
477 "src/f32-gemminc/6x8-aarch64-neonfma-ld64.S",
478 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700479 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700480 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S",
481 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S",
482 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700483 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700484 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
485 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700486 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700487 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
488 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
489 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
490]
491
492INTERNAL_MICROKERNEL_HDRS = [
493 "src/xnnpack/argmaxpool.h",
494 "src/xnnpack/avgpool.h",
495 "src/xnnpack/clamp.h",
496 "src/xnnpack/common.h",
497 "src/xnnpack/conv.h",
498 "src/xnnpack/dwconv.h",
499 "src/xnnpack/gavgpool.h",
500 "src/xnnpack/gemm.h",
501 "src/xnnpack/hswish.h",
502 "src/xnnpack/igemm.h",
503 "src/xnnpack/lut.h",
504 "src/xnnpack/math.h",
505 "src/xnnpack/maxpool.h",
506 "src/xnnpack/packx.h",
507 "src/xnnpack/pad.h",
508 "src/xnnpack/params.h",
509 "src/xnnpack/pavgpool.h",
510 "src/xnnpack/ppmm.h",
511 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700512 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700513 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700514 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700515 "src/xnnpack/rmax.h",
516 "src/xnnpack/scalar-utils.h",
517 "src/xnnpack/spmm.h",
518 "src/xnnpack/unpool.h",
519 "src/xnnpack/vadd.h",
520 "src/xnnpack/vmul.h",
521 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700522 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700523 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700524 "src/xnnpack/vscaleextexp.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700525 "src/xnnpack/vsub.h",
526 "src/xnnpack/zip.h",
527]
528
529INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
530 "include/xnnpack.h",
531 "src/xnnpack/allocator.h",
532 "src/xnnpack/compute.h",
533 "src/xnnpack/im2col.h",
534 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700535 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700536 "src/xnnpack/operator.h",
537 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700538 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700539 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700540 "src/xnnpack/requantization.h",
541]
542
543ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
544 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700545]
546
547MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700548 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700549 "include/xnnpack.h",
550]
551
552MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
553 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700554 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700555 "src/xnnpack/requantization.h",
556 "include/xnnpack.h",
557]
558
559OPERATOR_TEST_PARAMS_HDRS = [
560 "src/xnnpack/params.h",
561 "src/xnnpack/common.h",
562]
563
564WEIGHTS_PACK_HDRS = [
565 "src/xnnpack/pack.h",
566 "src/xnnpack/operator.h",
567 "src/xnnpack/compute.h",
568]
569
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700570LOGGING_COPTS = select({
571 # No logging in optimized mode
572 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
573 # Full logging in debug mode
574 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
575 # Error-only logging in default (fastbuild) mode
576 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
577})
578
579LOGGING_HDRS = [
580 "src/xnnpack/log.h",
581]
582
Marat Dukhan08c4a432019-10-03 09:29:21 -0700583xnnpack_cc_library(
584 name = "scalar_ukernels",
585 srcs = SCALAR_UKERNELS,
586 hdrs = INTERNAL_HDRS,
587 aarch32_copts = ["-marm"],
588 copts = xnnpack_std_copts(),
589 deps = [
590 "@FP16",
591 "@FXdiv",
592 ],
593)
594
595xnnpack_cc_library(
596 name = "psimd_ukernels",
597 srcs = PSIMD_UKERNELS,
598 hdrs = INTERNAL_HDRS,
599 aarch32_copts = [
600 "-marm",
601 "-mfpu=neon",
602 ],
603 copts = xnnpack_std_copts(),
604 optimized_copts = [
605 "-O3",
606 "-ffast-math",
607 ],
608 deps = [
609 "@FP16",
610 "@psimd",
611 ],
612)
613
614xnnpack_cc_library(
615 name = "neon_ukernels",
616 hdrs = INTERNAL_HDRS,
617 aarch32_copts = [
618 "-marm",
619 "-mfpu=neon",
620 ],
621 aarch32_srcs = NEON_UKERNELS,
622 aarch64_srcs = NEON_UKERNELS,
623 copts = xnnpack_std_copts(),
624 deps = ["@FP16"],
625)
626
627xnnpack_cc_library(
628 name = "neonfma_ukernels",
629 hdrs = INTERNAL_HDRS,
630 aarch32_copts = [
631 "-marm",
632 "-mfpu=neon-vfpv4",
633 ],
634 aarch32_srcs = NEONFMA_UKERNELS,
635 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
636 copts = xnnpack_std_copts(),
637 deps = ["@FP16"],
638)
639
640xnnpack_cc_library(
641 name = "neonfp16arith_ukernels",
642 hdrs = INTERNAL_HDRS,
643 aarch64_copts = ["-march=armv8.2-a+fp16"],
644 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
645 copts = xnnpack_std_copts(),
646 deps = ["@FP16"],
647)
648
649xnnpack_cc_library(
650 name = "sse2_ukernels",
651 hdrs = INTERNAL_HDRS,
652 copts = xnnpack_std_copts(),
653 x86_copts = ["-msse2"],
654 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
655 deps = ["@FP16"],
656)
657
658xnnpack_cc_library(
659 name = "avx_ukernels",
660 hdrs = INTERNAL_HDRS,
661 copts = xnnpack_std_copts(),
662 x86_copts = ["-mavx"],
663 x86_srcs = AVX_UKERNELS,
664 deps = ["@FP16"],
665)
666
667xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700668 name = "avx2_ukernels",
669 hdrs = INTERNAL_HDRS,
670 copts = xnnpack_std_copts(),
671 x86_copts = [
672 "-mfma",
673 "-mavx2",
674 ],
675 x86_srcs = AVX2_UKERNELS,
676 deps = ["@FP16"],
677)
678
679xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700680 name = "avx512f_ukernels",
681 hdrs = INTERNAL_HDRS,
682 copts = xnnpack_std_copts(),
683 x86_copts = ["-mavx512f"],
684 x86_srcs = AVX512F_UKERNELS,
685 deps = ["@FP16"],
686)
687
688xnnpack_cc_library(
689 name = "asm_ukernels",
690 hdrs = ["src/xnnpack/assembly.h"],
691 aarch32_srcs = AARCH32_ASM_UKERNELS,
692 aarch64_srcs = AARCH64_ASM_UKERNELS,
693)
694
695xnnpack_aggregate_library(
696 name = "ukernels",
697 aarch32_deps = [
698 ":psimd_ukernels",
699 ":neon_ukernels",
700 ":neonfma_ukernels",
701 ":asm_ukernels",
702 ],
703 aarch64_deps = [
704 ":psimd_ukernels",
705 ":neon_ukernels",
706 ":neonfma_ukernels",
707 ":neonfp16arith_ukernels",
708 ":asm_ukernels",
709 ],
710 generic_deps = [":scalar_ukernels"],
711 wasmsimd_deps = [
712 ":psimd_ukernels",
713 ],
714 x86_deps = [
715 ":psimd_ukernels",
716 ":sse2_ukernels",
717 ":avx_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700718 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700719 ":avx512f_ukernels",
720 ],
721)
722
723xnnpack_cc_library(
724 name = "im2col",
725 srcs = ["src/im2col.c"],
726 hdrs = [
727 "src/xnnpack/common.h",
728 "src/xnnpack/im2col.h",
729 ],
730 copts = xnnpack_std_copts(),
731)
732
733xnnpack_cc_library(
734 name = "indirection",
735 srcs = ["src/indirection.c"],
736 hdrs = INTERNAL_HDRS,
737 copts = xnnpack_std_copts(),
738 deps = [
739 "@FP16",
740 "@FXdiv",
741 "@pthreadpool",
742 ],
743)
744
745xnnpack_cc_library(
746 name = "operator_run",
747 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700748 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
749 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700750 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
751 "-Wno-vla",
752 ],
753 deps = [
754 "@FP16",
755 "@FXdiv",
756 "@clog",
757 "@pthreadpool",
758 ],
759)
760
761cc_library(
762 name = "enable_assembly",
763 defines = select({
764 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
765 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -0700766 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700767 }),
768)
769
Marat Dukhancf056b22019-10-07 10:26:29 -0700770xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700771 name = "operators",
772 srcs = OPERATOR_SRCS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700773 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -0700774 ],
775 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700776 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700777 "-Isrc",
778 "-Iinclude",
779 ] + select({
780 ":debug_build": [],
781 "//conditions:default": xnnpack_min_size_copts(),
782 }),
Marat Dukhancf056b22019-10-07 10:26:29 -0700783 wasm_srcs = ["src/wasm-stubs.c"],
784 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700785 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700786 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700787 "@FP16",
788 "@FXdiv",
789 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700790 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700791 ],
792)
793
794cc_library(
795 name = "XNNPACK",
796 srcs = [
797 "src/init.c",
798 ],
799 copts = xnnpack_std_copts() + LOGGING_COPTS + [
800 "-Isrc",
801 "-Iinclude",
802 ] + select({
803 ":debug_build": [],
804 "//conditions:default": xnnpack_min_size_copts(),
805 }),
806 includes = ["include"],
807 linkstatic = True,
808 textual_hdrs = ["include/xnnpack.h"],
809 visibility = xnnpack_visibility(),
810 deps = [
811 ":enable_assembly",
812 ":ukernels",
813 ":operator_run",
814 ":operators",
815 "@clog",
816 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -0700817 ] + select({
818 ":emscripten": [],
819 "//conditions:default": ["@cpuinfo"],
820 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700821)
822
823cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700824 name = "xnnpack_operators_nhwc_f32",
825 srcs = [
826 "src/init.c",
827 ],
828 copts = xnnpack_std_copts() + LOGGING_COPTS + [
829 "-Isrc",
830 "-Iinclude",
831 ] + select({
832 ":debug_build": [],
833 "//conditions:default": xnnpack_min_size_copts(),
834 }),
835 defines = [
836 "XNN_NO_Q8_OPERATORS",
837 "XNN_NO_U8_OPERATORS",
838 "XNN_NO_X8_OPERATORS",
839 "XNN_NO_SPNCHW_OPERATORS",
840 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700841 includes = ["include"],
842 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700843 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700844 visibility = xnnpack_visibility(),
845 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700846 ":enable_assembly",
847 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700848 ":operator_run",
849 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700850 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700851 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700852 ] + select({
853 ":emscripten": [],
854 "//conditions:default": ["@cpuinfo"],
855 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700856)
857
Marat Dukhancf056b22019-10-07 10:26:29 -0700858xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700859 name = "bench_utils",
860 srcs = ["bench/utils.cc"],
861 hdrs = ["bench/utils.h"],
862 copts = ["-Wno-unused-result"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700863 deps = ["@cpuinfo"],
864)
865
Frank Barchard7e955972019-10-11 10:34:25 -0700866######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -0700867
868xnnpack_benchmark(
869 name = "q8_gemm_bench",
870 srcs = [
871 "bench/gemm.h",
872 "bench/q8-gemm.cc",
873 "src/xnnpack/AlignedAllocator.h",
874 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
875 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
876 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
877)
878
879xnnpack_benchmark(
880 name = "f16_gemm_bench",
881 srcs = [
882 "bench/f16-gemm.cc",
883 "bench/gemm.h",
884 "src/xnnpack/AlignedAllocator.h",
885 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
886 copts = ["-Wno-unused-function"],
887 deps = MICROKERNEL_BENCHMARK_DEPS,
888)
889
890xnnpack_benchmark(
891 name = "f32_igemm_bench",
892 srcs = [
893 "bench/f32-igemm.cc",
894 "bench/conv.h",
895 "src/xnnpack/AlignedAllocator.h",
896 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -0700897 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700898)
899
900xnnpack_benchmark(
901 name = "f32_conv_hwc_bench",
902 srcs = [
903 "bench/f32-conv-hwc.cc",
904 "bench/dconv.h",
905 "src/xnnpack/AlignedAllocator.h",
906 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
907 copts = ["-Wno-unused-function"],
908 deps = MICROKERNEL_BENCHMARK_DEPS,
909)
910
911xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -0700912 name = "f32_conv_hwc2spchw_bench",
913 srcs = [
914 "bench/f32-conv-hwc2spchw.cc",
915 "bench/dconv.h",
916 "src/xnnpack/AlignedAllocator.h",
917 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
918 copts = ["-Wno-unused-function"],
919 deps = MICROKERNEL_BENCHMARK_DEPS,
920)
921
922xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700923 name = "f32_dwconv_bench",
924 srcs = [
925 "bench/f32-dwconv.cc",
926 "bench/dwconv.h",
927 "src/xnnpack/AlignedAllocator.h",
928 ] + MICROKERNEL_BENCHMARK_HDRS,
929 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
930)
931
932xnnpack_benchmark(
933 name = "f32_dwconv_spchw_bench",
934 srcs = [
935 "bench/f32-dwconv-spchw.cc",
936 "bench/dwconv.h",
937 "src/xnnpack/AlignedAllocator.h",
938 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
939 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
940)
941
942xnnpack_benchmark(
943 name = "f32_gemm_bench",
944 srcs = [
945 "bench/f32-gemm.cc",
946 "bench/gemm.h",
947 "src/xnnpack/AlignedAllocator.h",
948 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
949 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -0700950 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700951)
952
953xnnpack_benchmark(
954 name = "f32_rmax_bench",
955 srcs = [
956 "bench/f32-rmax.cc",
957 "src/xnnpack/AlignedAllocator.h",
958 ] + MICROKERNEL_BENCHMARK_HDRS,
959 deps = MICROKERNEL_BENCHMARK_DEPS,
960)
961
962xnnpack_benchmark(
963 name = "f32_spmm_bench",
964 srcs = [
965 "bench/f32-spmm.cc",
966 "bench/gemm.h",
967 "src/xnnpack/AlignedAllocator.h",
968 ] + MICROKERNEL_BENCHMARK_HDRS,
969 copts = ["-Wno-unused-function"],
970 deps = MICROKERNEL_BENCHMARK_DEPS,
971)
972
973xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -0700974 name = "f32_softargmax_bench",
975 srcs = [
976 "bench/f32-softargmax.cc",
977 ] + MICROKERNEL_BENCHMARK_HDRS,
978 copts = ["-Wno-unused-function"],
979 deps = MICROKERNEL_BENCHMARK_DEPS,
980)
981
982xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700983 name = "f32_im2col_gemm_bench",
984 srcs = [
985 "bench/f32-im2col-gemm.cc",
986 "bench/conv.h",
987 "src/xnnpack/AlignedAllocator.h",
988 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
989 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
990)
991
992########################### Benchmarks for operators ###########################
993
994xnnpack_benchmark(
995 name = "add_bench",
996 srcs = ["bench/add.cc"],
997 deps = OPERATOR_BENCHMARK_DEPS,
998)
999
1000xnnpack_benchmark(
1001 name = "average_pooling_bench",
1002 srcs = ["bench/average-pooling.cc"],
1003 deps = OPERATOR_BENCHMARK_DEPS,
1004)
1005
1006xnnpack_benchmark(
1007 name = "channel_shuffle_bench",
1008 srcs = ["bench/channel-shuffle.cc"],
1009 deps = OPERATOR_BENCHMARK_DEPS,
1010)
1011
1012xnnpack_benchmark(
1013 name = "convolution_bench",
1014 srcs = ["bench/convolution.cc"],
1015 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1016 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1017)
1018
1019xnnpack_benchmark(
1020 name = "deconvolution_bench",
1021 srcs = ["bench/deconvolution.cc"],
1022 copts = xnnpack_optional_tflite_copts(),
1023 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1024)
1025
1026xnnpack_benchmark(
1027 name = "global_average_pooling_bench",
1028 srcs = ["bench/global-average-pooling.cc"],
1029 deps = OPERATOR_BENCHMARK_DEPS,
1030)
1031
1032xnnpack_benchmark(
1033 name = "max_pooling_bench",
1034 srcs = ["bench/max-pooling.cc"],
1035 deps = OPERATOR_BENCHMARK_DEPS,
1036)
1037
1038xnnpack_benchmark(
1039 name = "sigmoid_bench",
1040 srcs = ["bench/sigmoid.cc"],
1041 deps = OPERATOR_BENCHMARK_DEPS,
1042)
1043
1044xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07001045 name = "prelu_bench",
1046 srcs = ["bench/prelu.cc"],
1047 copts = xnnpack_optional_tflite_copts(),
1048 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1049)
1050
1051xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001052 name = "softargmax_bench",
1053 srcs = ["bench/softargmax.cc"],
1054 deps = OPERATOR_BENCHMARK_DEPS,
1055)
1056
Marat Dukhanc068bb62019-10-04 13:24:39 -07001057############################# End-to-end benchmarks ############################
1058
1059cc_library(
1060 name = "mobilenet_v1",
1061 srcs = ["models/mobilenet-v1.cc"],
1062 hdrs = ["models/models.h"],
1063 linkstatic = True,
1064 deps = [
1065 ":XNNPACK",
1066 "@pthreadpool",
1067 ],
1068)
1069
1070cc_library(
1071 name = "mobilenet_v2",
1072 srcs = ["models/mobilenet-v2.cc"],
1073 hdrs = ["models/models.h"],
1074 linkstatic = True,
1075 deps = [
1076 ":XNNPACK",
1077 "@pthreadpool",
1078 ],
1079)
1080
1081xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07001082 name = "f32_gemm_e2e_bench",
1083 srcs = ["bench/f32-gemm-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1084 copts = ["-Wno-unused-function"],
1085 deps = MICROKERNEL_BENCHMARK_DEPS + [
1086 ":XNNPACK",
1087 ":mobilenet_v1",
1088 ":mobilenet_v2",
1089 ],
1090)
1091
1092xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07001093 name = "end2end_bench",
1094 srcs = ["bench/end2end.cc"],
1095 deps = [
1096 ":XNNPACK",
1097 ":mobilenet_v1",
1098 ":mobilenet_v2",
1099 "@pthreadpool",
1100 ],
1101)
1102
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001103#################### Accuracy evaluation for math functions ####################
1104
1105xnnpack_benchmark(
1106 name = "f32_exp_eval",
1107 srcs = [
1108 "eval/f32-exp.cc",
1109 "src/xnnpack/AlignedAllocator.h",
1110 ] + ACCURACY_EVAL_HDRS,
1111 deps = ACCURACY_EVAL_DEPS,
1112)
1113
Marat Dukhan515c9772019-10-17 18:07:57 -07001114xnnpack_benchmark(
1115 name = "f32_expminus_eval",
1116 srcs = [
1117 "eval/f32-expminus.cc",
1118 "src/xnnpack/AlignedAllocator.h",
1119 ] + ACCURACY_EVAL_HDRS,
1120 deps = ACCURACY_EVAL_DEPS,
1121)
1122
Marat Dukhan98ba4412019-10-23 02:14:28 -07001123xnnpack_benchmark(
1124 name = "f32_extexp_eval",
1125 srcs = [
1126 "eval/f32-extexp.cc",
1127 "src/xnnpack/AlignedAllocator.h",
1128 ] + ACCURACY_EVAL_HDRS,
1129 deps = ACCURACY_EVAL_DEPS,
1130)
1131
Marat Dukhan08c4a432019-10-03 09:29:21 -07001132######################### Unit tests for micro-kernels #########################
1133
1134xnnpack_unit_test(
1135 name = "f16_gemm_test",
1136 srcs = [
1137 "test/f16-gemm.cc",
1138 "test/gemm-microkernel-tester.h",
1139 "src/xnnpack/AlignedAllocator.h",
1140 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1141 deps = MICROKERNEL_TEST_DEPS,
1142)
1143
1144xnnpack_unit_test(
1145 name = "f32_argmaxpool_test",
1146 srcs = [
1147 "test/f32-argmaxpool.cc",
1148 "test/argmaxpool-microkernel-tester.h",
1149 "src/xnnpack/AlignedAllocator.h",
1150 ] + MICROKERNEL_TEST_HDRS,
1151 deps = MICROKERNEL_TEST_DEPS,
1152)
1153
1154xnnpack_unit_test(
1155 name = "f32_avgpool_test",
1156 srcs = [
1157 "test/f32-avgpool.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_clamp_test",
1166 srcs = [
1167 "test/f32-clamp.cc",
1168 "test/clamp-microkernel-tester.h",
1169 ] + MICROKERNEL_TEST_HDRS,
1170 deps = MICROKERNEL_TEST_DEPS,
1171)
1172
1173xnnpack_unit_test(
1174 name = "f32_igemm_test",
1175 srcs = [
1176 "test/f32-igemm.cc",
1177 "test/gemm-microkernel-tester.h",
1178 "src/xnnpack/AlignedAllocator.h",
1179 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1180 deps = MICROKERNEL_TEST_DEPS,
1181)
1182
1183xnnpack_unit_test(
1184 name = "f32_conv_hwc_test",
1185 srcs = [
1186 "test/f32-conv-hwc.cc",
1187 "test/conv-hwc-microkernel-tester.h",
1188 "src/xnnpack/AlignedAllocator.h",
1189 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1190 deps = MICROKERNEL_TEST_DEPS,
1191)
1192
1193xnnpack_unit_test(
1194 name = "f32_conv_hwc2spchw_test",
1195 srcs = [
1196 "test/f32-conv-hwc2spchw.cc",
1197 "test/conv-hwc2spchw-microkernel-tester.h",
1198 "src/xnnpack/AlignedAllocator.h",
1199 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1200 deps = MICROKERNEL_TEST_DEPS,
1201)
1202
1203xnnpack_unit_test(
1204 name = "f32_dwconv_test",
1205 srcs = [
1206 "test/f32-dwconv.cc",
1207 "test/dwconv-microkernel-tester.h",
1208 "src/xnnpack/AlignedAllocator.h",
1209 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1210 deps = MICROKERNEL_TEST_DEPS,
1211)
1212
1213xnnpack_unit_test(
1214 name = "f32_dwconv_spchw_test",
1215 srcs = [
1216 "test/f32-dwconv-spchw.cc",
1217 "test/dwconv-spchw-microkernel-tester.h",
1218 "src/xnnpack/AlignedAllocator.h",
1219 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1220 deps = MICROKERNEL_TEST_DEPS,
1221)
1222
1223xnnpack_unit_test(
1224 name = "f32_gavgpool_test",
1225 srcs = [
1226 "test/f32-gavgpool.cc",
1227 "test/gavgpool-microkernel-tester.h",
1228 "src/xnnpack/AlignedAllocator.h",
1229 ] + MICROKERNEL_TEST_HDRS,
1230 deps = MICROKERNEL_TEST_DEPS,
1231)
1232
1233xnnpack_unit_test(
1234 name = "f32_gavgpool_spchw_test",
1235 srcs = [
1236 "test/f32-gavgpool-spchw.cc",
1237 "test/gavgpool-spchw-microkernel-tester.h",
1238 "src/xnnpack/AlignedAllocator.h",
1239 ] + MICROKERNEL_TEST_HDRS,
1240 deps = MICROKERNEL_TEST_DEPS,
1241)
1242
1243xnnpack_unit_test(
1244 name = "f32_gemm_test",
1245 srcs = [
1246 "test/f32-gemm.cc",
1247 "test/gemm-microkernel-tester.h",
1248 "src/xnnpack/AlignedAllocator.h",
1249 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1250 deps = MICROKERNEL_TEST_DEPS,
1251)
1252
1253xnnpack_unit_test(
1254 name = "f32_gemminc_test",
1255 srcs = [
1256 "test/f32-gemminc.cc",
1257 "test/gemm-microkernel-tester.h",
1258 "src/xnnpack/AlignedAllocator.h",
1259 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1260 deps = MICROKERNEL_TEST_DEPS,
1261)
1262
1263xnnpack_unit_test(
1264 name = "f32_hswish_test",
1265 srcs = [
1266 "test/f32-hswish.cc",
1267 "test/hswish-microkernel-tester.h",
1268 ] + MICROKERNEL_TEST_HDRS,
1269 deps = MICROKERNEL_TEST_DEPS,
1270)
1271
1272xnnpack_unit_test(
1273 name = "f32_maxpool_test",
1274 srcs = [
1275 "test/f32-maxpool.cc",
1276 "test/maxpool-microkernel-tester.h",
1277 ] + MICROKERNEL_TEST_HDRS,
1278 deps = MICROKERNEL_TEST_DEPS,
1279)
1280
1281xnnpack_unit_test(
1282 name = "f32_pavgpool_test",
1283 srcs = [
1284 "test/f32-pavgpool.cc",
1285 "test/avgpool-microkernel-tester.h",
1286 "src/xnnpack/AlignedAllocator.h",
1287 ] + MICROKERNEL_TEST_HDRS,
1288 deps = MICROKERNEL_TEST_DEPS,
1289)
1290
1291xnnpack_unit_test(
1292 name = "f32_ppmm_test",
1293 srcs = [
1294 "test/f32-ppmm.cc",
1295 "test/gemm-microkernel-tester.h",
1296 "src/xnnpack/AlignedAllocator.h",
1297 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1298 deps = MICROKERNEL_TEST_DEPS,
1299)
1300
1301xnnpack_unit_test(
1302 name = "f32_prelu_test",
1303 srcs = [
1304 "test/f32-prelu.cc",
1305 "test/prelu-microkernel-tester.h",
1306 "src/xnnpack/AlignedAllocator.h",
1307 ] + MICROKERNEL_TEST_HDRS,
1308 deps = MICROKERNEL_TEST_DEPS,
1309)
1310
1311xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001312 name = "f32_raddexpminusmax_test",
1313 srcs = [
1314 "test/f32-raddexpminusmax.cc",
1315 "test/raddexpminusmax-microkernel-tester.h",
1316 ] + MICROKERNEL_TEST_HDRS,
1317 deps = MICROKERNEL_TEST_DEPS,
1318)
1319
1320xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001321 name = "f32_raddextexp_test",
1322 srcs = [
1323 "test/f32-raddextexp.cc",
1324 "test/raddextexp-microkernel-tester.h",
1325 ] + MICROKERNEL_TEST_HDRS,
1326 deps = MICROKERNEL_TEST_DEPS,
1327)
1328
1329xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001330 name = "f32_raddstoreexpminusmax_test",
1331 srcs = [
1332 "test/f32-raddstoreexpminusmax.cc",
1333 "test/raddstoreexpminusmax-microkernel-tester.h",
1334 ] + MICROKERNEL_TEST_HDRS,
1335 deps = MICROKERNEL_TEST_DEPS,
1336)
1337
1338xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001339 name = "f32_rmax_test",
1340 srcs = [
1341 "test/f32-rmax.cc",
1342 "test/rmax-microkernel-tester.h",
1343 ] + MICROKERNEL_TEST_HDRS,
1344 deps = MICROKERNEL_TEST_DEPS,
1345)
1346
1347xnnpack_unit_test(
1348 name = "f32_spmm_test",
1349 srcs = [
1350 "test/f32-spmm.cc",
1351 "test/spmm-microkernel-tester.h",
1352 "src/xnnpack/AlignedAllocator.h",
1353 ] + MICROKERNEL_TEST_HDRS,
1354 deps = MICROKERNEL_TEST_DEPS,
1355)
1356
1357xnnpack_unit_test(
1358 name = "f32_vadd_test",
1359 srcs = [
1360 "test/f32-vadd.cc",
1361 "test/vadd-microkernel-tester.h",
1362 ] + MICROKERNEL_TEST_HDRS,
1363 deps = MICROKERNEL_TEST_DEPS,
1364)
1365
1366xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001367 name = "f32_vmul_test",
1368 srcs = [
1369 "test/f32-vmul.cc",
1370 "test/vmul-microkernel-tester.h",
1371 ] + MICROKERNEL_TEST_HDRS,
1372 deps = MICROKERNEL_TEST_DEPS,
1373)
1374
1375xnnpack_unit_test(
1376 name = "f32_vmulcaddc_test",
1377 srcs = [
1378 "test/f32-vmulcaddc.cc",
1379 "test/vmulcaddc-microkernel-tester.h",
1380 "src/xnnpack/AlignedAllocator.h",
1381 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1382 deps = MICROKERNEL_TEST_DEPS,
1383)
1384
1385xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001386 name = "f32_vscale_test",
1387 srcs = [
1388 "test/f32-vscale.cc",
1389 "test/vscale-microkernel-tester.h",
1390 ] + MICROKERNEL_TEST_HDRS,
1391 deps = MICROKERNEL_TEST_DEPS,
1392)
1393
1394xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001395 name = "f32_vscaleexpminusmax_test",
1396 srcs = [
1397 "test/f32-vscaleexpminusmax.cc",
1398 "test/vscaleexpminusmax-microkernel-tester.h",
1399 ] + MICROKERNEL_TEST_HDRS,
1400 deps = MICROKERNEL_TEST_DEPS,
1401)
1402
1403xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001404 name = "f32_vscaleextexp_test",
1405 srcs = [
1406 "test/f32-vscaleextexp.cc",
1407 "test/vscaleextexp-microkernel-tester.h",
1408 ] + MICROKERNEL_TEST_HDRS,
1409 deps = MICROKERNEL_TEST_DEPS,
1410)
1411
1412xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001413 name = "f32_vsub_test",
1414 srcs = [
1415 "test/f32-vsub.cc",
1416 "test/vsub-microkernel-tester.h",
1417 ] + MICROKERNEL_TEST_HDRS,
1418 deps = MICROKERNEL_TEST_DEPS,
1419)
1420
1421xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001422 name = "q8_avgpool_test",
1423 srcs = [
1424 "test/q8-avgpool.cc",
1425 "test/avgpool-microkernel-tester.h",
1426 "src/xnnpack/AlignedAllocator.h",
1427 ] + MICROKERNEL_TEST_HDRS,
1428 deps = MICROKERNEL_TEST_DEPS,
1429)
1430
1431xnnpack_unit_test(
1432 name = "q8_igemm_test",
1433 srcs = [
1434 "test/q8-igemm.cc",
1435 "test/gemm-microkernel-tester.h",
1436 "src/xnnpack/AlignedAllocator.h",
1437 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1438 deps = MICROKERNEL_TEST_DEPS,
1439)
1440
1441xnnpack_unit_test(
1442 name = "q8_dwconv_test",
1443 srcs = [
1444 "test/q8-dwconv.cc",
1445 "test/dwconv-microkernel-tester.h",
1446 "src/xnnpack/AlignedAllocator.h",
1447 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1448 deps = MICROKERNEL_TEST_DEPS,
1449)
1450
1451xnnpack_unit_test(
1452 name = "q8_gavgpool_test",
1453 srcs = [
1454 "test/q8-gavgpool.cc",
1455 "test/gavgpool-microkernel-tester.h",
1456 "src/xnnpack/AlignedAllocator.h",
1457 ] + MICROKERNEL_TEST_HDRS,
1458 deps = MICROKERNEL_TEST_DEPS,
1459)
1460
1461xnnpack_unit_test(
1462 name = "q8_gemm_test",
1463 srcs = [
1464 "test/q8-gemm.cc",
1465 "test/gemm-microkernel-tester.h",
1466 "src/xnnpack/AlignedAllocator.h",
1467 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1468 deps = MICROKERNEL_TEST_DEPS,
1469)
1470
1471xnnpack_unit_test(
1472 name = "q8_vadd_test",
1473 srcs = [
1474 "test/q8-vadd.cc",
1475 "test/vadd-microkernel-tester.h",
1476 ] + MICROKERNEL_TEST_HDRS,
1477 deps = MICROKERNEL_TEST_DEPS,
1478)
1479
1480xnnpack_unit_test(
1481 name = "u8_clamp_test",
1482 srcs = [
1483 "test/u8-clamp.cc",
1484 "test/clamp-microkernel-tester.h",
1485 ] + MICROKERNEL_TEST_HDRS,
1486 deps = MICROKERNEL_TEST_DEPS,
1487)
1488
1489xnnpack_unit_test(
1490 name = "u8_lut32norm_test",
1491 srcs = [
1492 "test/u8-lut32norm.cc",
1493 "test/lut-norm-microkernel-tester.h",
1494 ] + MICROKERNEL_TEST_HDRS,
1495 deps = MICROKERNEL_TEST_DEPS,
1496)
1497
1498xnnpack_unit_test(
1499 name = "u8_maxpool_test",
1500 srcs = [
1501 "test/u8-maxpool.cc",
1502 "test/maxpool-microkernel-tester.h",
1503 ] + MICROKERNEL_TEST_HDRS,
1504 deps = MICROKERNEL_TEST_DEPS,
1505)
1506
1507xnnpack_unit_test(
1508 name = "u8_rmax_test",
1509 srcs = [
1510 "test/u8-rmax.cc",
1511 "test/rmax-microkernel-tester.h",
1512 ] + MICROKERNEL_TEST_HDRS,
1513 deps = MICROKERNEL_TEST_DEPS,
1514)
1515
1516xnnpack_unit_test(
1517 name = "x32_packx_test",
1518 srcs = [
1519 "test/x32-packx.cc",
1520 "test/pack-microkernel-tester.h",
1521 "src/xnnpack/AlignedAllocator.h",
1522 ] + MICROKERNEL_TEST_HDRS,
1523 deps = MICROKERNEL_TEST_DEPS,
1524)
1525
1526xnnpack_unit_test(
1527 name = "x32_pad_test",
1528 srcs = [
1529 "test/x32-pad.cc",
1530 "test/pad-microkernel-tester.h",
1531 ] + MICROKERNEL_TEST_HDRS,
1532 deps = MICROKERNEL_TEST_DEPS,
1533)
1534
1535xnnpack_unit_test(
1536 name = "x32_unpool_test",
1537 srcs = [
1538 "test/x32-unpool.cc",
1539 "test/unpool-microkernel-tester.h",
1540 ] + MICROKERNEL_TEST_HDRS,
1541 deps = MICROKERNEL_TEST_DEPS,
1542)
1543
1544xnnpack_unit_test(
1545 name = "x32_zip_test",
1546 srcs = [
1547 "test/x32-zip.cc",
1548 "test/zip-microkernel-tester.h",
1549 ] + MICROKERNEL_TEST_HDRS,
1550 deps = MICROKERNEL_TEST_DEPS,
1551)
1552
1553xnnpack_unit_test(
1554 name = "x8_lut_test",
1555 srcs = [
1556 "test/x8-lut.cc",
1557 "test/lut-microkernel-tester.h",
1558 ] + MICROKERNEL_TEST_HDRS,
1559 deps = MICROKERNEL_TEST_DEPS,
1560)
1561
1562xnnpack_unit_test(
1563 name = "x8_zip_test",
1564 srcs = [
1565 "test/x8-zip.cc",
1566 "test/zip-microkernel-tester.h",
1567 ] + MICROKERNEL_TEST_HDRS,
1568 deps = MICROKERNEL_TEST_DEPS,
1569)
1570
1571########################### Size test for the library ##########################
1572
1573xnnpack_binary(
1574 name = "size_test",
1575 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001576 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001577)
1578
1579########################### Unit tests for operators ###########################
1580
1581xnnpack_unit_test(
1582 name = "add_test",
1583 srcs = [
1584 "test/add.cc",
1585 "test/add-operator-tester.h",
1586 ],
1587 deps = OPERATOR_TEST_DEPS,
1588)
1589
1590xnnpack_unit_test(
1591 name = "argmax_pooling_test",
1592 srcs = [
1593 "test/argmax-pooling.cc",
1594 "test/argmax-pooling-operator-tester.h",
1595 ] + OPERATOR_TEST_PARAMS_HDRS,
1596 deps = OPERATOR_TEST_DEPS,
1597)
1598
1599xnnpack_unit_test(
1600 name = "average_pooling_test",
1601 srcs = [
1602 "test/average-pooling.cc",
1603 "test/average-pooling-operator-tester.h",
1604 ] + OPERATOR_TEST_PARAMS_HDRS,
1605 deps = OPERATOR_TEST_DEPS,
1606)
1607
1608xnnpack_unit_test(
1609 name = "channel_pad_test",
1610 srcs = [
1611 "test/channel-pad.cc",
1612 "test/channel-pad-operator-tester.h",
1613 ] + OPERATOR_TEST_PARAMS_HDRS,
1614 deps = OPERATOR_TEST_DEPS,
1615)
1616
1617xnnpack_unit_test(
1618 name = "channel_shuffle_test",
1619 srcs = [
1620 "test/channel-shuffle.cc",
1621 "test/channel-shuffle-operator-tester.h",
1622 ],
1623 deps = OPERATOR_TEST_DEPS,
1624)
1625
1626xnnpack_unit_test(
1627 name = "clamp_test",
1628 srcs = [
1629 "test/clamp.cc",
1630 "test/clamp-operator-tester.h",
1631 ],
1632 deps = OPERATOR_TEST_DEPS,
1633)
1634
1635xnnpack_unit_test(
1636 name = "convolution_test",
1637 srcs = [
1638 "test/convolution.cc",
1639 "test/convolution-operator-tester.h",
1640 ],
1641 deps = OPERATOR_TEST_DEPS,
1642)
1643
1644xnnpack_unit_test(
1645 name = "convolution_spnchw_test",
1646 srcs = [
1647 "test/convolution-spnchw.cc",
1648 "test/convolution-spnchw-operator-tester.h",
1649 ],
1650 deps = OPERATOR_TEST_DEPS,
1651)
1652
1653xnnpack_unit_test(
1654 name = "deconvolution_test",
1655 srcs = [
1656 "test/deconvolution.cc",
1657 "test/deconvolution-operator-tester.h",
1658 ] + OPERATOR_TEST_PARAMS_HDRS,
1659 deps = OPERATOR_TEST_DEPS,
1660)
1661
1662xnnpack_unit_test(
1663 name = "fully_connected_test",
1664 srcs = [
1665 "test/fully-connected.cc",
1666 "test/fully-connected-operator-tester.h",
1667 ],
1668 deps = OPERATOR_TEST_DEPS,
1669)
1670
1671xnnpack_unit_test(
1672 name = "global_average_pooling_test",
1673 srcs = [
1674 "test/global-average-pooling.cc",
1675 "test/global-average-pooling-operator-tester.h",
1676 ] + OPERATOR_TEST_PARAMS_HDRS,
1677 deps = OPERATOR_TEST_DEPS,
1678)
1679
1680xnnpack_unit_test(
1681 name = "global_average_pooling_spnchw_test",
1682 srcs = [
1683 "test/global-average-pooling-spnchw.cc",
1684 "test/global-average-pooling-spnchw-operator-tester.h",
1685 ],
1686 deps = OPERATOR_TEST_DEPS,
1687)
1688
1689xnnpack_unit_test(
1690 name = "hardswish_test",
1691 srcs = [
1692 "test/hardswish.cc",
1693 "test/hardswish-operator-tester.h",
1694 ],
1695 deps = OPERATOR_TEST_DEPS,
1696)
1697
1698xnnpack_unit_test(
1699 name = "leaky_relu_test",
1700 srcs = [
1701 "test/leaky-relu.cc",
1702 "test/leaky-relu-operator-tester.h",
1703 ],
1704 deps = OPERATOR_TEST_DEPS,
1705)
1706
1707xnnpack_unit_test(
1708 name = "max_pooling_test",
1709 srcs = [
1710 "test/max-pooling.cc",
1711 "test/max-pooling-operator-tester.h",
1712 ] + OPERATOR_TEST_PARAMS_HDRS,
1713 deps = OPERATOR_TEST_DEPS,
1714)
1715
1716xnnpack_unit_test(
1717 name = "prelu_test",
1718 srcs = [
1719 "test/prelu.cc",
1720 "test/prelu-operator-tester.h",
1721 ] + OPERATOR_TEST_PARAMS_HDRS,
1722 deps = OPERATOR_TEST_DEPS,
1723)
1724
1725xnnpack_unit_test(
1726 name = "sigmoid_test",
1727 srcs = [
1728 "test/sigmoid.cc",
1729 "test/sigmoid-operator-tester.h",
1730 ],
1731 deps = OPERATOR_TEST_DEPS,
1732)
1733
1734xnnpack_unit_test(
1735 name = "softargmax_test",
1736 srcs = [
1737 "test/softargmax.cc",
1738 "test/softargmax-operator-tester.h",
1739 ],
1740 deps = OPERATOR_TEST_DEPS,
1741)
1742
1743xnnpack_unit_test(
1744 name = "unpooling_test",
1745 srcs = [
1746 "test/unpooling.cc",
1747 "test/unpooling-operator-tester.h",
1748 ],
1749 deps = OPERATOR_TEST_DEPS,
1750)
1751
1752############################# Build configurations #############################
1753
Marat Dukhanb8642352019-10-30 15:43:02 -07001754# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07001755config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07001756 name = "xnn_enable_assembly_explicit_true",
1757 define_values = {"xnn_enable_assembly": "true"},
1758)
1759
1760# Disables usage of assembly kernels.
1761config_setting(
1762 name = "xnn_enable_assembly_explicit_false",
1763 define_values = {"xnn_enable_assembly": "false"},
1764)
1765
1766# Builds with -c dbg
1767config_setting(
1768 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001769 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07001770 "compilation_mode": "dbg",
1771 },
1772)
1773
1774# Builds with -c opt
1775config_setting(
1776 name = "optimized_build",
1777 values = {
1778 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001779 },
1780)
1781
1782config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07001783 name = "linux_k8",
1784 values = {"cpu": "k8"},
1785)
1786
1787config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07001788 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07001789 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07001790)
1791
1792config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001793 name = "android",
1794 values = {"crosstool_top": "//external:android/crosstool"},
1795)
1796
1797config_setting(
1798 name = "android_armv7",
1799 values = {
1800 "crosstool_top": "//external:android/crosstool",
1801 "cpu": "armeabi-v7a",
1802 },
1803)
1804
1805config_setting(
1806 name = "android_arm64",
1807 values = {
1808 "crosstool_top": "//external:android/crosstool",
1809 "cpu": "arm64-v8a",
1810 },
1811)
1812
1813config_setting(
1814 name = "android_x86",
1815 values = {
1816 "crosstool_top": "//external:android/crosstool",
1817 "cpu": "x86",
1818 },
1819)
1820
1821config_setting(
1822 name = "android_x86_64",
1823 values = {
1824 "crosstool_top": "//external:android/crosstool",
1825 "cpu": "x86_64",
1826 },
1827)
1828
1829config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07001830 name = "macos_x86_64",
1831 values = {
1832 "apple_platform_type": "macos",
1833 "cpu": "darwin",
1834 },
1835)
1836
1837config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001838 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001839 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07001840)
1841
1842config_setting(
1843 name = "emscripten_wasm",
1844 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001845 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001846 "cpu": "wasm",
1847 },
1848)
1849
1850config_setting(
1851 name = "emscripten_wasmsimd",
1852 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001853 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001854 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07001855 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001856 },
1857)
1858
1859config_setting(
1860 name = "emscripten_asmjs",
1861 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001862 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001863 "cpu": "asmjs",
1864 },
1865)