blob: e230d566f28e4ff021943b020c162c34efe53298 [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 Dukhanef4416e2019-10-31 13:44:40 -07001082 name = "f32_dwconv_e2e_bench",
1083 srcs = ["bench/f32-dwconv-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 Dukhan5f18d262019-10-31 10:24:14 -07001093 name = "f32_gemm_e2e_bench",
1094 srcs = ["bench/f32-gemm-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1095 copts = ["-Wno-unused-function"],
1096 deps = MICROKERNEL_BENCHMARK_DEPS + [
1097 ":XNNPACK",
1098 ":mobilenet_v1",
1099 ":mobilenet_v2",
1100 ],
1101)
1102
1103xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07001104 name = "end2end_bench",
1105 srcs = ["bench/end2end.cc"],
1106 deps = [
1107 ":XNNPACK",
1108 ":mobilenet_v1",
1109 ":mobilenet_v2",
1110 "@pthreadpool",
1111 ],
1112)
1113
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001114#################### Accuracy evaluation for math functions ####################
1115
1116xnnpack_benchmark(
1117 name = "f32_exp_eval",
1118 srcs = [
1119 "eval/f32-exp.cc",
1120 "src/xnnpack/AlignedAllocator.h",
1121 ] + ACCURACY_EVAL_HDRS,
1122 deps = ACCURACY_EVAL_DEPS,
1123)
1124
Marat Dukhan515c9772019-10-17 18:07:57 -07001125xnnpack_benchmark(
1126 name = "f32_expminus_eval",
1127 srcs = [
1128 "eval/f32-expminus.cc",
1129 "src/xnnpack/AlignedAllocator.h",
1130 ] + ACCURACY_EVAL_HDRS,
1131 deps = ACCURACY_EVAL_DEPS,
1132)
1133
Marat Dukhan98ba4412019-10-23 02:14:28 -07001134xnnpack_benchmark(
1135 name = "f32_extexp_eval",
1136 srcs = [
1137 "eval/f32-extexp.cc",
1138 "src/xnnpack/AlignedAllocator.h",
1139 ] + ACCURACY_EVAL_HDRS,
1140 deps = ACCURACY_EVAL_DEPS,
1141)
1142
Marat Dukhan08c4a432019-10-03 09:29:21 -07001143######################### Unit tests for micro-kernels #########################
1144
1145xnnpack_unit_test(
1146 name = "f16_gemm_test",
1147 srcs = [
1148 "test/f16-gemm.cc",
1149 "test/gemm-microkernel-tester.h",
1150 "src/xnnpack/AlignedAllocator.h",
1151 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1152 deps = MICROKERNEL_TEST_DEPS,
1153)
1154
1155xnnpack_unit_test(
1156 name = "f32_argmaxpool_test",
1157 srcs = [
1158 "test/f32-argmaxpool.cc",
1159 "test/argmaxpool-microkernel-tester.h",
1160 "src/xnnpack/AlignedAllocator.h",
1161 ] + MICROKERNEL_TEST_HDRS,
1162 deps = MICROKERNEL_TEST_DEPS,
1163)
1164
1165xnnpack_unit_test(
1166 name = "f32_avgpool_test",
1167 srcs = [
1168 "test/f32-avgpool.cc",
1169 "test/avgpool-microkernel-tester.h",
1170 "src/xnnpack/AlignedAllocator.h",
1171 ] + MICROKERNEL_TEST_HDRS,
1172 deps = MICROKERNEL_TEST_DEPS,
1173)
1174
1175xnnpack_unit_test(
1176 name = "f32_clamp_test",
1177 srcs = [
1178 "test/f32-clamp.cc",
1179 "test/clamp-microkernel-tester.h",
1180 ] + MICROKERNEL_TEST_HDRS,
1181 deps = MICROKERNEL_TEST_DEPS,
1182)
1183
1184xnnpack_unit_test(
1185 name = "f32_igemm_test",
1186 srcs = [
1187 "test/f32-igemm.cc",
1188 "test/gemm-microkernel-tester.h",
1189 "src/xnnpack/AlignedAllocator.h",
1190 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1191 deps = MICROKERNEL_TEST_DEPS,
1192)
1193
1194xnnpack_unit_test(
1195 name = "f32_conv_hwc_test",
1196 srcs = [
1197 "test/f32-conv-hwc.cc",
1198 "test/conv-hwc-microkernel-tester.h",
1199 "src/xnnpack/AlignedAllocator.h",
1200 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1201 deps = MICROKERNEL_TEST_DEPS,
1202)
1203
1204xnnpack_unit_test(
1205 name = "f32_conv_hwc2spchw_test",
1206 srcs = [
1207 "test/f32-conv-hwc2spchw.cc",
1208 "test/conv-hwc2spchw-microkernel-tester.h",
1209 "src/xnnpack/AlignedAllocator.h",
1210 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1211 deps = MICROKERNEL_TEST_DEPS,
1212)
1213
1214xnnpack_unit_test(
1215 name = "f32_dwconv_test",
1216 srcs = [
1217 "test/f32-dwconv.cc",
1218 "test/dwconv-microkernel-tester.h",
1219 "src/xnnpack/AlignedAllocator.h",
1220 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1221 deps = MICROKERNEL_TEST_DEPS,
1222)
1223
1224xnnpack_unit_test(
1225 name = "f32_dwconv_spchw_test",
1226 srcs = [
1227 "test/f32-dwconv-spchw.cc",
1228 "test/dwconv-spchw-microkernel-tester.h",
1229 "src/xnnpack/AlignedAllocator.h",
1230 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1231 deps = MICROKERNEL_TEST_DEPS,
1232)
1233
1234xnnpack_unit_test(
1235 name = "f32_gavgpool_test",
1236 srcs = [
1237 "test/f32-gavgpool.cc",
1238 "test/gavgpool-microkernel-tester.h",
1239 "src/xnnpack/AlignedAllocator.h",
1240 ] + MICROKERNEL_TEST_HDRS,
1241 deps = MICROKERNEL_TEST_DEPS,
1242)
1243
1244xnnpack_unit_test(
1245 name = "f32_gavgpool_spchw_test",
1246 srcs = [
1247 "test/f32-gavgpool-spchw.cc",
1248 "test/gavgpool-spchw-microkernel-tester.h",
1249 "src/xnnpack/AlignedAllocator.h",
1250 ] + MICROKERNEL_TEST_HDRS,
1251 deps = MICROKERNEL_TEST_DEPS,
1252)
1253
1254xnnpack_unit_test(
1255 name = "f32_gemm_test",
1256 srcs = [
1257 "test/f32-gemm.cc",
1258 "test/gemm-microkernel-tester.h",
1259 "src/xnnpack/AlignedAllocator.h",
1260 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1261 deps = MICROKERNEL_TEST_DEPS,
1262)
1263
1264xnnpack_unit_test(
1265 name = "f32_gemminc_test",
1266 srcs = [
1267 "test/f32-gemminc.cc",
1268 "test/gemm-microkernel-tester.h",
1269 "src/xnnpack/AlignedAllocator.h",
1270 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1271 deps = MICROKERNEL_TEST_DEPS,
1272)
1273
1274xnnpack_unit_test(
1275 name = "f32_hswish_test",
1276 srcs = [
1277 "test/f32-hswish.cc",
1278 "test/hswish-microkernel-tester.h",
1279 ] + MICROKERNEL_TEST_HDRS,
1280 deps = MICROKERNEL_TEST_DEPS,
1281)
1282
1283xnnpack_unit_test(
1284 name = "f32_maxpool_test",
1285 srcs = [
1286 "test/f32-maxpool.cc",
1287 "test/maxpool-microkernel-tester.h",
1288 ] + MICROKERNEL_TEST_HDRS,
1289 deps = MICROKERNEL_TEST_DEPS,
1290)
1291
1292xnnpack_unit_test(
1293 name = "f32_pavgpool_test",
1294 srcs = [
1295 "test/f32-pavgpool.cc",
1296 "test/avgpool-microkernel-tester.h",
1297 "src/xnnpack/AlignedAllocator.h",
1298 ] + MICROKERNEL_TEST_HDRS,
1299 deps = MICROKERNEL_TEST_DEPS,
1300)
1301
1302xnnpack_unit_test(
1303 name = "f32_ppmm_test",
1304 srcs = [
1305 "test/f32-ppmm.cc",
1306 "test/gemm-microkernel-tester.h",
1307 "src/xnnpack/AlignedAllocator.h",
1308 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1309 deps = MICROKERNEL_TEST_DEPS,
1310)
1311
1312xnnpack_unit_test(
1313 name = "f32_prelu_test",
1314 srcs = [
1315 "test/f32-prelu.cc",
1316 "test/prelu-microkernel-tester.h",
1317 "src/xnnpack/AlignedAllocator.h",
1318 ] + MICROKERNEL_TEST_HDRS,
1319 deps = MICROKERNEL_TEST_DEPS,
1320)
1321
1322xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001323 name = "f32_raddexpminusmax_test",
1324 srcs = [
1325 "test/f32-raddexpminusmax.cc",
1326 "test/raddexpminusmax-microkernel-tester.h",
1327 ] + MICROKERNEL_TEST_HDRS,
1328 deps = MICROKERNEL_TEST_DEPS,
1329)
1330
1331xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001332 name = "f32_raddextexp_test",
1333 srcs = [
1334 "test/f32-raddextexp.cc",
1335 "test/raddextexp-microkernel-tester.h",
1336 ] + MICROKERNEL_TEST_HDRS,
1337 deps = MICROKERNEL_TEST_DEPS,
1338)
1339
1340xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001341 name = "f32_raddstoreexpminusmax_test",
1342 srcs = [
1343 "test/f32-raddstoreexpminusmax.cc",
1344 "test/raddstoreexpminusmax-microkernel-tester.h",
1345 ] + MICROKERNEL_TEST_HDRS,
1346 deps = MICROKERNEL_TEST_DEPS,
1347)
1348
1349xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001350 name = "f32_rmax_test",
1351 srcs = [
1352 "test/f32-rmax.cc",
1353 "test/rmax-microkernel-tester.h",
1354 ] + MICROKERNEL_TEST_HDRS,
1355 deps = MICROKERNEL_TEST_DEPS,
1356)
1357
1358xnnpack_unit_test(
1359 name = "f32_spmm_test",
1360 srcs = [
1361 "test/f32-spmm.cc",
1362 "test/spmm-microkernel-tester.h",
1363 "src/xnnpack/AlignedAllocator.h",
1364 ] + MICROKERNEL_TEST_HDRS,
1365 deps = MICROKERNEL_TEST_DEPS,
1366)
1367
1368xnnpack_unit_test(
1369 name = "f32_vadd_test",
1370 srcs = [
1371 "test/f32-vadd.cc",
1372 "test/vadd-microkernel-tester.h",
1373 ] + MICROKERNEL_TEST_HDRS,
1374 deps = MICROKERNEL_TEST_DEPS,
1375)
1376
1377xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001378 name = "f32_vmul_test",
1379 srcs = [
1380 "test/f32-vmul.cc",
1381 "test/vmul-microkernel-tester.h",
1382 ] + MICROKERNEL_TEST_HDRS,
1383 deps = MICROKERNEL_TEST_DEPS,
1384)
1385
1386xnnpack_unit_test(
1387 name = "f32_vmulcaddc_test",
1388 srcs = [
1389 "test/f32-vmulcaddc.cc",
1390 "test/vmulcaddc-microkernel-tester.h",
1391 "src/xnnpack/AlignedAllocator.h",
1392 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1393 deps = MICROKERNEL_TEST_DEPS,
1394)
1395
1396xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001397 name = "f32_vscale_test",
1398 srcs = [
1399 "test/f32-vscale.cc",
1400 "test/vscale-microkernel-tester.h",
1401 ] + MICROKERNEL_TEST_HDRS,
1402 deps = MICROKERNEL_TEST_DEPS,
1403)
1404
1405xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001406 name = "f32_vscaleexpminusmax_test",
1407 srcs = [
1408 "test/f32-vscaleexpminusmax.cc",
1409 "test/vscaleexpminusmax-microkernel-tester.h",
1410 ] + MICROKERNEL_TEST_HDRS,
1411 deps = MICROKERNEL_TEST_DEPS,
1412)
1413
1414xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001415 name = "f32_vscaleextexp_test",
1416 srcs = [
1417 "test/f32-vscaleextexp.cc",
1418 "test/vscaleextexp-microkernel-tester.h",
1419 ] + MICROKERNEL_TEST_HDRS,
1420 deps = MICROKERNEL_TEST_DEPS,
1421)
1422
1423xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001424 name = "f32_vsub_test",
1425 srcs = [
1426 "test/f32-vsub.cc",
1427 "test/vsub-microkernel-tester.h",
1428 ] + MICROKERNEL_TEST_HDRS,
1429 deps = MICROKERNEL_TEST_DEPS,
1430)
1431
1432xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001433 name = "q8_avgpool_test",
1434 srcs = [
1435 "test/q8-avgpool.cc",
1436 "test/avgpool-microkernel-tester.h",
1437 "src/xnnpack/AlignedAllocator.h",
1438 ] + MICROKERNEL_TEST_HDRS,
1439 deps = MICROKERNEL_TEST_DEPS,
1440)
1441
1442xnnpack_unit_test(
1443 name = "q8_igemm_test",
1444 srcs = [
1445 "test/q8-igemm.cc",
1446 "test/gemm-microkernel-tester.h",
1447 "src/xnnpack/AlignedAllocator.h",
1448 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1449 deps = MICROKERNEL_TEST_DEPS,
1450)
1451
1452xnnpack_unit_test(
1453 name = "q8_dwconv_test",
1454 srcs = [
1455 "test/q8-dwconv.cc",
1456 "test/dwconv-microkernel-tester.h",
1457 "src/xnnpack/AlignedAllocator.h",
1458 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1459 deps = MICROKERNEL_TEST_DEPS,
1460)
1461
1462xnnpack_unit_test(
1463 name = "q8_gavgpool_test",
1464 srcs = [
1465 "test/q8-gavgpool.cc",
1466 "test/gavgpool-microkernel-tester.h",
1467 "src/xnnpack/AlignedAllocator.h",
1468 ] + MICROKERNEL_TEST_HDRS,
1469 deps = MICROKERNEL_TEST_DEPS,
1470)
1471
1472xnnpack_unit_test(
1473 name = "q8_gemm_test",
1474 srcs = [
1475 "test/q8-gemm.cc",
1476 "test/gemm-microkernel-tester.h",
1477 "src/xnnpack/AlignedAllocator.h",
1478 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1479 deps = MICROKERNEL_TEST_DEPS,
1480)
1481
1482xnnpack_unit_test(
1483 name = "q8_vadd_test",
1484 srcs = [
1485 "test/q8-vadd.cc",
1486 "test/vadd-microkernel-tester.h",
1487 ] + MICROKERNEL_TEST_HDRS,
1488 deps = MICROKERNEL_TEST_DEPS,
1489)
1490
1491xnnpack_unit_test(
1492 name = "u8_clamp_test",
1493 srcs = [
1494 "test/u8-clamp.cc",
1495 "test/clamp-microkernel-tester.h",
1496 ] + MICROKERNEL_TEST_HDRS,
1497 deps = MICROKERNEL_TEST_DEPS,
1498)
1499
1500xnnpack_unit_test(
1501 name = "u8_lut32norm_test",
1502 srcs = [
1503 "test/u8-lut32norm.cc",
1504 "test/lut-norm-microkernel-tester.h",
1505 ] + MICROKERNEL_TEST_HDRS,
1506 deps = MICROKERNEL_TEST_DEPS,
1507)
1508
1509xnnpack_unit_test(
1510 name = "u8_maxpool_test",
1511 srcs = [
1512 "test/u8-maxpool.cc",
1513 "test/maxpool-microkernel-tester.h",
1514 ] + MICROKERNEL_TEST_HDRS,
1515 deps = MICROKERNEL_TEST_DEPS,
1516)
1517
1518xnnpack_unit_test(
1519 name = "u8_rmax_test",
1520 srcs = [
1521 "test/u8-rmax.cc",
1522 "test/rmax-microkernel-tester.h",
1523 ] + MICROKERNEL_TEST_HDRS,
1524 deps = MICROKERNEL_TEST_DEPS,
1525)
1526
1527xnnpack_unit_test(
1528 name = "x32_packx_test",
1529 srcs = [
1530 "test/x32-packx.cc",
1531 "test/pack-microkernel-tester.h",
1532 "src/xnnpack/AlignedAllocator.h",
1533 ] + MICROKERNEL_TEST_HDRS,
1534 deps = MICROKERNEL_TEST_DEPS,
1535)
1536
1537xnnpack_unit_test(
1538 name = "x32_pad_test",
1539 srcs = [
1540 "test/x32-pad.cc",
1541 "test/pad-microkernel-tester.h",
1542 ] + MICROKERNEL_TEST_HDRS,
1543 deps = MICROKERNEL_TEST_DEPS,
1544)
1545
1546xnnpack_unit_test(
1547 name = "x32_unpool_test",
1548 srcs = [
1549 "test/x32-unpool.cc",
1550 "test/unpool-microkernel-tester.h",
1551 ] + MICROKERNEL_TEST_HDRS,
1552 deps = MICROKERNEL_TEST_DEPS,
1553)
1554
1555xnnpack_unit_test(
1556 name = "x32_zip_test",
1557 srcs = [
1558 "test/x32-zip.cc",
1559 "test/zip-microkernel-tester.h",
1560 ] + MICROKERNEL_TEST_HDRS,
1561 deps = MICROKERNEL_TEST_DEPS,
1562)
1563
1564xnnpack_unit_test(
1565 name = "x8_lut_test",
1566 srcs = [
1567 "test/x8-lut.cc",
1568 "test/lut-microkernel-tester.h",
1569 ] + MICROKERNEL_TEST_HDRS,
1570 deps = MICROKERNEL_TEST_DEPS,
1571)
1572
1573xnnpack_unit_test(
1574 name = "x8_zip_test",
1575 srcs = [
1576 "test/x8-zip.cc",
1577 "test/zip-microkernel-tester.h",
1578 ] + MICROKERNEL_TEST_HDRS,
1579 deps = MICROKERNEL_TEST_DEPS,
1580)
1581
1582########################### Size test for the library ##########################
1583
1584xnnpack_binary(
1585 name = "size_test",
1586 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001587 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001588)
1589
1590########################### Unit tests for operators ###########################
1591
1592xnnpack_unit_test(
1593 name = "add_test",
1594 srcs = [
1595 "test/add.cc",
1596 "test/add-operator-tester.h",
1597 ],
1598 deps = OPERATOR_TEST_DEPS,
1599)
1600
1601xnnpack_unit_test(
1602 name = "argmax_pooling_test",
1603 srcs = [
1604 "test/argmax-pooling.cc",
1605 "test/argmax-pooling-operator-tester.h",
1606 ] + OPERATOR_TEST_PARAMS_HDRS,
1607 deps = OPERATOR_TEST_DEPS,
1608)
1609
1610xnnpack_unit_test(
1611 name = "average_pooling_test",
1612 srcs = [
1613 "test/average-pooling.cc",
1614 "test/average-pooling-operator-tester.h",
1615 ] + OPERATOR_TEST_PARAMS_HDRS,
1616 deps = OPERATOR_TEST_DEPS,
1617)
1618
1619xnnpack_unit_test(
1620 name = "channel_pad_test",
1621 srcs = [
1622 "test/channel-pad.cc",
1623 "test/channel-pad-operator-tester.h",
1624 ] + OPERATOR_TEST_PARAMS_HDRS,
1625 deps = OPERATOR_TEST_DEPS,
1626)
1627
1628xnnpack_unit_test(
1629 name = "channel_shuffle_test",
1630 srcs = [
1631 "test/channel-shuffle.cc",
1632 "test/channel-shuffle-operator-tester.h",
1633 ],
1634 deps = OPERATOR_TEST_DEPS,
1635)
1636
1637xnnpack_unit_test(
1638 name = "clamp_test",
1639 srcs = [
1640 "test/clamp.cc",
1641 "test/clamp-operator-tester.h",
1642 ],
1643 deps = OPERATOR_TEST_DEPS,
1644)
1645
1646xnnpack_unit_test(
1647 name = "convolution_test",
1648 srcs = [
1649 "test/convolution.cc",
1650 "test/convolution-operator-tester.h",
1651 ],
1652 deps = OPERATOR_TEST_DEPS,
1653)
1654
1655xnnpack_unit_test(
1656 name = "convolution_spnchw_test",
1657 srcs = [
1658 "test/convolution-spnchw.cc",
1659 "test/convolution-spnchw-operator-tester.h",
1660 ],
1661 deps = OPERATOR_TEST_DEPS,
1662)
1663
1664xnnpack_unit_test(
1665 name = "deconvolution_test",
1666 srcs = [
1667 "test/deconvolution.cc",
1668 "test/deconvolution-operator-tester.h",
1669 ] + OPERATOR_TEST_PARAMS_HDRS,
1670 deps = OPERATOR_TEST_DEPS,
1671)
1672
1673xnnpack_unit_test(
1674 name = "fully_connected_test",
1675 srcs = [
1676 "test/fully-connected.cc",
1677 "test/fully-connected-operator-tester.h",
1678 ],
1679 deps = OPERATOR_TEST_DEPS,
1680)
1681
1682xnnpack_unit_test(
1683 name = "global_average_pooling_test",
1684 srcs = [
1685 "test/global-average-pooling.cc",
1686 "test/global-average-pooling-operator-tester.h",
1687 ] + OPERATOR_TEST_PARAMS_HDRS,
1688 deps = OPERATOR_TEST_DEPS,
1689)
1690
1691xnnpack_unit_test(
1692 name = "global_average_pooling_spnchw_test",
1693 srcs = [
1694 "test/global-average-pooling-spnchw.cc",
1695 "test/global-average-pooling-spnchw-operator-tester.h",
1696 ],
1697 deps = OPERATOR_TEST_DEPS,
1698)
1699
1700xnnpack_unit_test(
1701 name = "hardswish_test",
1702 srcs = [
1703 "test/hardswish.cc",
1704 "test/hardswish-operator-tester.h",
1705 ],
1706 deps = OPERATOR_TEST_DEPS,
1707)
1708
1709xnnpack_unit_test(
1710 name = "leaky_relu_test",
1711 srcs = [
1712 "test/leaky-relu.cc",
1713 "test/leaky-relu-operator-tester.h",
1714 ],
1715 deps = OPERATOR_TEST_DEPS,
1716)
1717
1718xnnpack_unit_test(
1719 name = "max_pooling_test",
1720 srcs = [
1721 "test/max-pooling.cc",
1722 "test/max-pooling-operator-tester.h",
1723 ] + OPERATOR_TEST_PARAMS_HDRS,
1724 deps = OPERATOR_TEST_DEPS,
1725)
1726
1727xnnpack_unit_test(
1728 name = "prelu_test",
1729 srcs = [
1730 "test/prelu.cc",
1731 "test/prelu-operator-tester.h",
1732 ] + OPERATOR_TEST_PARAMS_HDRS,
1733 deps = OPERATOR_TEST_DEPS,
1734)
1735
1736xnnpack_unit_test(
1737 name = "sigmoid_test",
1738 srcs = [
1739 "test/sigmoid.cc",
1740 "test/sigmoid-operator-tester.h",
1741 ],
1742 deps = OPERATOR_TEST_DEPS,
1743)
1744
1745xnnpack_unit_test(
1746 name = "softargmax_test",
1747 srcs = [
1748 "test/softargmax.cc",
1749 "test/softargmax-operator-tester.h",
1750 ],
1751 deps = OPERATOR_TEST_DEPS,
1752)
1753
1754xnnpack_unit_test(
1755 name = "unpooling_test",
1756 srcs = [
1757 "test/unpooling.cc",
1758 "test/unpooling-operator-tester.h",
1759 ],
1760 deps = OPERATOR_TEST_DEPS,
1761)
1762
1763############################# Build configurations #############################
1764
Marat Dukhanb8642352019-10-30 15:43:02 -07001765# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07001766config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07001767 name = "xnn_enable_assembly_explicit_true",
1768 define_values = {"xnn_enable_assembly": "true"},
1769)
1770
1771# Disables usage of assembly kernels.
1772config_setting(
1773 name = "xnn_enable_assembly_explicit_false",
1774 define_values = {"xnn_enable_assembly": "false"},
1775)
1776
1777# Builds with -c dbg
1778config_setting(
1779 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001780 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07001781 "compilation_mode": "dbg",
1782 },
1783)
1784
1785# Builds with -c opt
1786config_setting(
1787 name = "optimized_build",
1788 values = {
1789 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001790 },
1791)
1792
1793config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07001794 name = "linux_k8",
1795 values = {"cpu": "k8"},
1796)
1797
1798config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07001799 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07001800 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07001801)
1802
1803config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001804 name = "android",
1805 values = {"crosstool_top": "//external:android/crosstool"},
1806)
1807
1808config_setting(
1809 name = "android_armv7",
1810 values = {
1811 "crosstool_top": "//external:android/crosstool",
1812 "cpu": "armeabi-v7a",
1813 },
1814)
1815
1816config_setting(
1817 name = "android_arm64",
1818 values = {
1819 "crosstool_top": "//external:android/crosstool",
1820 "cpu": "arm64-v8a",
1821 },
1822)
1823
1824config_setting(
1825 name = "android_x86",
1826 values = {
1827 "crosstool_top": "//external:android/crosstool",
1828 "cpu": "x86",
1829 },
1830)
1831
1832config_setting(
1833 name = "android_x86_64",
1834 values = {
1835 "crosstool_top": "//external:android/crosstool",
1836 "cpu": "x86_64",
1837 },
1838)
1839
1840config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07001841 name = "macos_x86_64",
1842 values = {
1843 "apple_platform_type": "macos",
1844 "cpu": "darwin",
1845 },
1846)
1847
1848config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001849 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001850 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07001851)
1852
1853config_setting(
1854 name = "emscripten_wasm",
1855 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001856 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001857 "cpu": "wasm",
1858 },
1859)
1860
1861config_setting(
1862 name = "emscripten_wasmsimd",
1863 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001864 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001865 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07001866 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001867 },
1868)
1869
1870config_setting(
1871 name = "emscripten_asmjs",
1872 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001873 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001874 "cpu": "asmjs",
1875 },
1876)