blob: b300e1dbaf358ada13b0c83031e25df883b36642 [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 Dukhan98ba4412019-10-23 02:14:28 -0700433 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700434]
435
436AARCH32_ASM_UKERNELS = [
437 "src/q8-dwconv/up8x9-aarch32-neon.S",
438]
439
440AARCH64_ASM_UKERNELS = [
441 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
442 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
443 "src/f32-gemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700444 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700445 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a57.S",
446 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a75.S",
447 "src/f32-gemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700448 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700449 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a57.S",
450 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a75.S",
451 "src/f32-gemm/4x8-aarch64-neonfma-ld128.S",
452 "src/f32-gemm/4x8-aarch64-neonfma-ld64.S",
453 "src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700454 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700455 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a57.S",
456 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S",
457 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S",
458 "src/f32-gemm/6x8-aarch64-neonfma-ld128.S",
459 "src/f32-gemm/6x8-aarch64-neonfma-ld64.S",
460 "src/f32-gemminc/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700461 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700462 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a57.S",
463 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a75.S",
464 "src/f32-gemminc/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700465 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700466 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a57.S",
467 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a75.S",
468 "src/f32-gemminc/4x8-aarch64-neonfma-ld128.S",
469 "src/f32-gemminc/4x8-aarch64-neonfma-ld64.S",
470 "src/f32-gemminc/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700471 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700472 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a57.S",
473 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a73.S",
474 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a75.S",
475 "src/f32-gemminc/6x8-aarch64-neonfma-ld128.S",
476 "src/f32-gemminc/6x8-aarch64-neonfma-ld64.S",
477 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700478 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700479 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S",
480 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S",
481 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700482 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700483 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
484 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700485 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700486 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
487 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
488 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
489]
490
491INTERNAL_MICROKERNEL_HDRS = [
492 "src/xnnpack/argmaxpool.h",
493 "src/xnnpack/avgpool.h",
494 "src/xnnpack/clamp.h",
495 "src/xnnpack/common.h",
496 "src/xnnpack/conv.h",
497 "src/xnnpack/dwconv.h",
498 "src/xnnpack/gavgpool.h",
499 "src/xnnpack/gemm.h",
500 "src/xnnpack/hswish.h",
501 "src/xnnpack/igemm.h",
502 "src/xnnpack/lut.h",
503 "src/xnnpack/math.h",
504 "src/xnnpack/maxpool.h",
505 "src/xnnpack/packx.h",
506 "src/xnnpack/pad.h",
507 "src/xnnpack/params.h",
508 "src/xnnpack/pavgpool.h",
509 "src/xnnpack/ppmm.h",
510 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700511 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700512 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700513 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700514 "src/xnnpack/rmax.h",
515 "src/xnnpack/scalar-utils.h",
516 "src/xnnpack/spmm.h",
517 "src/xnnpack/unpool.h",
518 "src/xnnpack/vadd.h",
519 "src/xnnpack/vmul.h",
520 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700521 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700522 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700523 "src/xnnpack/vscaleextexp.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700524 "src/xnnpack/vsub.h",
525 "src/xnnpack/zip.h",
526]
527
528INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
529 "include/xnnpack.h",
530 "src/xnnpack/allocator.h",
531 "src/xnnpack/compute.h",
532 "src/xnnpack/im2col.h",
533 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700534 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700535 "src/xnnpack/operator.h",
536 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700537 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700538 "src/xnnpack/requantization.h",
539]
540
541ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
542 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700543]
544
545MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
546 "src/xnnpack/requantization.h",
547 "include/xnnpack.h",
548]
549
550MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
551 "src/xnnpack/isa-checks.h",
552 "src/xnnpack/requantization.h",
553 "include/xnnpack.h",
554]
555
556OPERATOR_TEST_PARAMS_HDRS = [
557 "src/xnnpack/params.h",
558 "src/xnnpack/common.h",
559]
560
561WEIGHTS_PACK_HDRS = [
562 "src/xnnpack/pack.h",
563 "src/xnnpack/operator.h",
564 "src/xnnpack/compute.h",
565]
566
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700567LOGGING_COPTS = select({
568 # No logging in optimized mode
569 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
570 # Full logging in debug mode
571 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
572 # Error-only logging in default (fastbuild) mode
573 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
574})
575
576LOGGING_HDRS = [
577 "src/xnnpack/log.h",
578]
579
Marat Dukhan08c4a432019-10-03 09:29:21 -0700580xnnpack_cc_library(
581 name = "scalar_ukernels",
582 srcs = SCALAR_UKERNELS,
583 hdrs = INTERNAL_HDRS,
584 aarch32_copts = ["-marm"],
585 copts = xnnpack_std_copts(),
586 deps = [
587 "@FP16",
588 "@FXdiv",
589 ],
590)
591
592xnnpack_cc_library(
593 name = "psimd_ukernels",
594 srcs = PSIMD_UKERNELS,
595 hdrs = INTERNAL_HDRS,
596 aarch32_copts = [
597 "-marm",
598 "-mfpu=neon",
599 ],
600 copts = xnnpack_std_copts(),
601 optimized_copts = [
602 "-O3",
603 "-ffast-math",
604 ],
605 deps = [
606 "@FP16",
607 "@psimd",
608 ],
609)
610
611xnnpack_cc_library(
612 name = "neon_ukernels",
613 hdrs = INTERNAL_HDRS,
614 aarch32_copts = [
615 "-marm",
616 "-mfpu=neon",
617 ],
618 aarch32_srcs = NEON_UKERNELS,
619 aarch64_srcs = NEON_UKERNELS,
620 copts = xnnpack_std_copts(),
621 deps = ["@FP16"],
622)
623
624xnnpack_cc_library(
625 name = "neonfma_ukernels",
626 hdrs = INTERNAL_HDRS,
627 aarch32_copts = [
628 "-marm",
629 "-mfpu=neon-vfpv4",
630 ],
631 aarch32_srcs = NEONFMA_UKERNELS,
632 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
633 copts = xnnpack_std_copts(),
634 deps = ["@FP16"],
635)
636
637xnnpack_cc_library(
638 name = "neonfp16arith_ukernels",
639 hdrs = INTERNAL_HDRS,
640 aarch64_copts = ["-march=armv8.2-a+fp16"],
641 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
642 copts = xnnpack_std_copts(),
643 deps = ["@FP16"],
644)
645
646xnnpack_cc_library(
647 name = "sse2_ukernels",
648 hdrs = INTERNAL_HDRS,
649 copts = xnnpack_std_copts(),
650 x86_copts = ["-msse2"],
651 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
652 deps = ["@FP16"],
653)
654
655xnnpack_cc_library(
656 name = "avx_ukernels",
657 hdrs = INTERNAL_HDRS,
658 copts = xnnpack_std_copts(),
659 x86_copts = ["-mavx"],
660 x86_srcs = AVX_UKERNELS,
661 deps = ["@FP16"],
662)
663
664xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700665 name = "avx2_ukernels",
666 hdrs = INTERNAL_HDRS,
667 copts = xnnpack_std_copts(),
668 x86_copts = [
669 "-mfma",
670 "-mavx2",
671 ],
672 x86_srcs = AVX2_UKERNELS,
673 deps = ["@FP16"],
674)
675
676xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700677 name = "avx512f_ukernels",
678 hdrs = INTERNAL_HDRS,
679 copts = xnnpack_std_copts(),
680 x86_copts = ["-mavx512f"],
681 x86_srcs = AVX512F_UKERNELS,
682 deps = ["@FP16"],
683)
684
685xnnpack_cc_library(
686 name = "asm_ukernels",
687 hdrs = ["src/xnnpack/assembly.h"],
688 aarch32_srcs = AARCH32_ASM_UKERNELS,
689 aarch64_srcs = AARCH64_ASM_UKERNELS,
690)
691
692xnnpack_aggregate_library(
693 name = "ukernels",
694 aarch32_deps = [
695 ":psimd_ukernels",
696 ":neon_ukernels",
697 ":neonfma_ukernels",
698 ":asm_ukernels",
699 ],
700 aarch64_deps = [
701 ":psimd_ukernels",
702 ":neon_ukernels",
703 ":neonfma_ukernels",
704 ":neonfp16arith_ukernels",
705 ":asm_ukernels",
706 ],
707 generic_deps = [":scalar_ukernels"],
708 wasmsimd_deps = [
709 ":psimd_ukernels",
710 ],
711 x86_deps = [
712 ":psimd_ukernels",
713 ":sse2_ukernels",
714 ":avx_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700715 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700716 ":avx512f_ukernels",
717 ],
718)
719
720xnnpack_cc_library(
721 name = "im2col",
722 srcs = ["src/im2col.c"],
723 hdrs = [
724 "src/xnnpack/common.h",
725 "src/xnnpack/im2col.h",
726 ],
727 copts = xnnpack_std_copts(),
728)
729
730xnnpack_cc_library(
731 name = "indirection",
732 srcs = ["src/indirection.c"],
733 hdrs = INTERNAL_HDRS,
734 copts = xnnpack_std_copts(),
735 deps = [
736 "@FP16",
737 "@FXdiv",
738 "@pthreadpool",
739 ],
740)
741
742xnnpack_cc_library(
743 name = "operator_run",
744 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700745 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
746 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700747 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
748 "-Wno-vla",
749 ],
750 deps = [
751 "@FP16",
752 "@FXdiv",
753 "@clog",
754 "@pthreadpool",
755 ],
756)
757
758cc_library(
759 name = "enable_assembly",
760 defines = select({
761 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
762 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -0700763 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700764 }),
765)
766
Marat Dukhancf056b22019-10-07 10:26:29 -0700767xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700768 name = "operators",
769 srcs = OPERATOR_SRCS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700770 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -0700771 ],
772 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700773 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700774 "-Isrc",
775 "-Iinclude",
776 ] + select({
777 ":debug_build": [],
778 "//conditions:default": xnnpack_min_size_copts(),
779 }),
Marat Dukhancf056b22019-10-07 10:26:29 -0700780 wasm_srcs = ["src/wasm-stubs.c"],
781 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700782 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700783 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700784 "@FP16",
785 "@FXdiv",
786 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700787 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700788 ],
789)
790
791cc_library(
792 name = "XNNPACK",
793 srcs = [
794 "src/init.c",
795 ],
796 copts = xnnpack_std_copts() + LOGGING_COPTS + [
797 "-Isrc",
798 "-Iinclude",
799 ] + select({
800 ":debug_build": [],
801 "//conditions:default": xnnpack_min_size_copts(),
802 }),
803 includes = ["include"],
804 linkstatic = True,
805 textual_hdrs = ["include/xnnpack.h"],
806 visibility = xnnpack_visibility(),
807 deps = [
808 ":enable_assembly",
809 ":ukernels",
810 ":operator_run",
811 ":operators",
812 "@clog",
813 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -0700814 ] + select({
815 ":emscripten": [],
816 "//conditions:default": ["@cpuinfo"],
817 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700818)
819
820cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700821 name = "xnnpack_operators_nhwc_f32",
822 srcs = [
823 "src/init.c",
824 ],
825 copts = xnnpack_std_copts() + LOGGING_COPTS + [
826 "-Isrc",
827 "-Iinclude",
828 ] + select({
829 ":debug_build": [],
830 "//conditions:default": xnnpack_min_size_copts(),
831 }),
832 defines = [
833 "XNN_NO_Q8_OPERATORS",
834 "XNN_NO_U8_OPERATORS",
835 "XNN_NO_X8_OPERATORS",
836 "XNN_NO_SPNCHW_OPERATORS",
837 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700838 includes = ["include"],
839 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700840 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700841 visibility = xnnpack_visibility(),
842 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700843 ":enable_assembly",
844 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700845 ":operator_run",
846 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700847 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700848 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700849 ] + select({
850 ":emscripten": [],
851 "//conditions:default": ["@cpuinfo"],
852 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700853)
854
Marat Dukhancf056b22019-10-07 10:26:29 -0700855xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700856 name = "bench_utils",
857 srcs = ["bench/utils.cc"],
858 hdrs = ["bench/utils.h"],
859 copts = ["-Wno-unused-result"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700860 deps = ["@cpuinfo"],
861)
862
Frank Barchard7e955972019-10-11 10:34:25 -0700863######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -0700864
865xnnpack_benchmark(
866 name = "q8_gemm_bench",
867 srcs = [
868 "bench/gemm.h",
869 "bench/q8-gemm.cc",
870 "src/xnnpack/AlignedAllocator.h",
871 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
872 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
873 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
874)
875
876xnnpack_benchmark(
877 name = "f16_gemm_bench",
878 srcs = [
879 "bench/f16-gemm.cc",
880 "bench/gemm.h",
881 "src/xnnpack/AlignedAllocator.h",
882 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
883 copts = ["-Wno-unused-function"],
884 deps = MICROKERNEL_BENCHMARK_DEPS,
885)
886
887xnnpack_benchmark(
888 name = "f32_igemm_bench",
889 srcs = [
890 "bench/f32-igemm.cc",
891 "bench/conv.h",
892 "src/xnnpack/AlignedAllocator.h",
893 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -0700894 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700895)
896
897xnnpack_benchmark(
898 name = "f32_conv_hwc_bench",
899 srcs = [
900 "bench/f32-conv-hwc.cc",
901 "bench/dconv.h",
902 "src/xnnpack/AlignedAllocator.h",
903 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
904 copts = ["-Wno-unused-function"],
905 deps = MICROKERNEL_BENCHMARK_DEPS,
906)
907
908xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -0700909 name = "f32_conv_hwc2spchw_bench",
910 srcs = [
911 "bench/f32-conv-hwc2spchw.cc",
912 "bench/dconv.h",
913 "src/xnnpack/AlignedAllocator.h",
914 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
915 copts = ["-Wno-unused-function"],
916 deps = MICROKERNEL_BENCHMARK_DEPS,
917)
918
919xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700920 name = "f32_dwconv_bench",
921 srcs = [
922 "bench/f32-dwconv.cc",
923 "bench/dwconv.h",
924 "src/xnnpack/AlignedAllocator.h",
925 ] + MICROKERNEL_BENCHMARK_HDRS,
926 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
927)
928
929xnnpack_benchmark(
930 name = "f32_dwconv_spchw_bench",
931 srcs = [
932 "bench/f32-dwconv-spchw.cc",
933 "bench/dwconv.h",
934 "src/xnnpack/AlignedAllocator.h",
935 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
936 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
937)
938
939xnnpack_benchmark(
940 name = "f32_gemm_bench",
941 srcs = [
942 "bench/f32-gemm.cc",
943 "bench/gemm.h",
944 "src/xnnpack/AlignedAllocator.h",
945 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
946 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -0700947 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700948)
949
950xnnpack_benchmark(
951 name = "f32_rmax_bench",
952 srcs = [
953 "bench/f32-rmax.cc",
954 "src/xnnpack/AlignedAllocator.h",
955 ] + MICROKERNEL_BENCHMARK_HDRS,
956 deps = MICROKERNEL_BENCHMARK_DEPS,
957)
958
959xnnpack_benchmark(
960 name = "f32_spmm_bench",
961 srcs = [
962 "bench/f32-spmm.cc",
963 "bench/gemm.h",
964 "src/xnnpack/AlignedAllocator.h",
965 ] + MICROKERNEL_BENCHMARK_HDRS,
966 copts = ["-Wno-unused-function"],
967 deps = MICROKERNEL_BENCHMARK_DEPS,
968)
969
970xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -0700971 name = "f32_softargmax_bench",
972 srcs = [
973 "bench/f32-softargmax.cc",
974 ] + MICROKERNEL_BENCHMARK_HDRS,
975 copts = ["-Wno-unused-function"],
976 deps = MICROKERNEL_BENCHMARK_DEPS,
977)
978
979xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700980 name = "f32_im2col_gemm_bench",
981 srcs = [
982 "bench/f32-im2col-gemm.cc",
983 "bench/conv.h",
984 "src/xnnpack/AlignedAllocator.h",
985 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
986 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
987)
988
989########################### Benchmarks for operators ###########################
990
991xnnpack_benchmark(
992 name = "add_bench",
993 srcs = ["bench/add.cc"],
994 deps = OPERATOR_BENCHMARK_DEPS,
995)
996
997xnnpack_benchmark(
998 name = "average_pooling_bench",
999 srcs = ["bench/average-pooling.cc"],
1000 deps = OPERATOR_BENCHMARK_DEPS,
1001)
1002
1003xnnpack_benchmark(
1004 name = "channel_shuffle_bench",
1005 srcs = ["bench/channel-shuffle.cc"],
1006 deps = OPERATOR_BENCHMARK_DEPS,
1007)
1008
1009xnnpack_benchmark(
1010 name = "convolution_bench",
1011 srcs = ["bench/convolution.cc"],
1012 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1013 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1014)
1015
1016xnnpack_benchmark(
1017 name = "deconvolution_bench",
1018 srcs = ["bench/deconvolution.cc"],
1019 copts = xnnpack_optional_tflite_copts(),
1020 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1021)
1022
1023xnnpack_benchmark(
1024 name = "global_average_pooling_bench",
1025 srcs = ["bench/global-average-pooling.cc"],
1026 deps = OPERATOR_BENCHMARK_DEPS,
1027)
1028
1029xnnpack_benchmark(
1030 name = "max_pooling_bench",
1031 srcs = ["bench/max-pooling.cc"],
1032 deps = OPERATOR_BENCHMARK_DEPS,
1033)
1034
1035xnnpack_benchmark(
1036 name = "sigmoid_bench",
1037 srcs = ["bench/sigmoid.cc"],
1038 deps = OPERATOR_BENCHMARK_DEPS,
1039)
1040
1041xnnpack_benchmark(
1042 name = "softargmax_bench",
1043 srcs = ["bench/softargmax.cc"],
1044 deps = OPERATOR_BENCHMARK_DEPS,
1045)
1046
Marat Dukhanc068bb62019-10-04 13:24:39 -07001047############################# End-to-end benchmarks ############################
1048
1049cc_library(
1050 name = "mobilenet_v1",
1051 srcs = ["models/mobilenet-v1.cc"],
1052 hdrs = ["models/models.h"],
1053 linkstatic = True,
1054 deps = [
1055 ":XNNPACK",
1056 "@pthreadpool",
1057 ],
1058)
1059
1060cc_library(
1061 name = "mobilenet_v2",
1062 srcs = ["models/mobilenet-v2.cc"],
1063 hdrs = ["models/models.h"],
1064 linkstatic = True,
1065 deps = [
1066 ":XNNPACK",
1067 "@pthreadpool",
1068 ],
1069)
1070
1071xnnpack_benchmark(
1072 name = "end2end_bench",
1073 srcs = ["bench/end2end.cc"],
1074 deps = [
1075 ":XNNPACK",
1076 ":mobilenet_v1",
1077 ":mobilenet_v2",
1078 "@pthreadpool",
1079 ],
1080)
1081
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001082#################### Accuracy evaluation for math functions ####################
1083
1084xnnpack_benchmark(
1085 name = "f32_exp_eval",
1086 srcs = [
1087 "eval/f32-exp.cc",
1088 "src/xnnpack/AlignedAllocator.h",
1089 ] + ACCURACY_EVAL_HDRS,
1090 deps = ACCURACY_EVAL_DEPS,
1091)
1092
Marat Dukhan515c9772019-10-17 18:07:57 -07001093xnnpack_benchmark(
1094 name = "f32_expminus_eval",
1095 srcs = [
1096 "eval/f32-expminus.cc",
1097 "src/xnnpack/AlignedAllocator.h",
1098 ] + ACCURACY_EVAL_HDRS,
1099 deps = ACCURACY_EVAL_DEPS,
1100)
1101
Marat Dukhan98ba4412019-10-23 02:14:28 -07001102xnnpack_benchmark(
1103 name = "f32_extexp_eval",
1104 srcs = [
1105 "eval/f32-extexp.cc",
1106 "src/xnnpack/AlignedAllocator.h",
1107 ] + ACCURACY_EVAL_HDRS,
1108 deps = ACCURACY_EVAL_DEPS,
1109)
1110
Marat Dukhan08c4a432019-10-03 09:29:21 -07001111######################### Unit tests for micro-kernels #########################
1112
1113xnnpack_unit_test(
1114 name = "f16_gemm_test",
1115 srcs = [
1116 "test/f16-gemm.cc",
1117 "test/gemm-microkernel-tester.h",
1118 "src/xnnpack/AlignedAllocator.h",
1119 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1120 deps = MICROKERNEL_TEST_DEPS,
1121)
1122
1123xnnpack_unit_test(
1124 name = "f32_argmaxpool_test",
1125 srcs = [
1126 "test/f32-argmaxpool.cc",
1127 "test/argmaxpool-microkernel-tester.h",
1128 "src/xnnpack/AlignedAllocator.h",
1129 ] + MICROKERNEL_TEST_HDRS,
1130 deps = MICROKERNEL_TEST_DEPS,
1131)
1132
1133xnnpack_unit_test(
1134 name = "f32_avgpool_test",
1135 srcs = [
1136 "test/f32-avgpool.cc",
1137 "test/avgpool-microkernel-tester.h",
1138 "src/xnnpack/AlignedAllocator.h",
1139 ] + MICROKERNEL_TEST_HDRS,
1140 deps = MICROKERNEL_TEST_DEPS,
1141)
1142
1143xnnpack_unit_test(
1144 name = "f32_clamp_test",
1145 srcs = [
1146 "test/f32-clamp.cc",
1147 "test/clamp-microkernel-tester.h",
1148 ] + MICROKERNEL_TEST_HDRS,
1149 deps = MICROKERNEL_TEST_DEPS,
1150)
1151
1152xnnpack_unit_test(
1153 name = "f32_igemm_test",
1154 srcs = [
1155 "test/f32-igemm.cc",
1156 "test/gemm-microkernel-tester.h",
1157 "src/xnnpack/AlignedAllocator.h",
1158 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1159 deps = MICROKERNEL_TEST_DEPS,
1160)
1161
1162xnnpack_unit_test(
1163 name = "f32_conv_hwc_test",
1164 srcs = [
1165 "test/f32-conv-hwc.cc",
1166 "test/conv-hwc-microkernel-tester.h",
1167 "src/xnnpack/AlignedAllocator.h",
1168 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1169 deps = MICROKERNEL_TEST_DEPS,
1170)
1171
1172xnnpack_unit_test(
1173 name = "f32_conv_hwc2spchw_test",
1174 srcs = [
1175 "test/f32-conv-hwc2spchw.cc",
1176 "test/conv-hwc2spchw-microkernel-tester.h",
1177 "src/xnnpack/AlignedAllocator.h",
1178 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1179 deps = MICROKERNEL_TEST_DEPS,
1180)
1181
1182xnnpack_unit_test(
1183 name = "f32_dwconv_test",
1184 srcs = [
1185 "test/f32-dwconv.cc",
1186 "test/dwconv-microkernel-tester.h",
1187 "src/xnnpack/AlignedAllocator.h",
1188 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1189 deps = MICROKERNEL_TEST_DEPS,
1190)
1191
1192xnnpack_unit_test(
1193 name = "f32_dwconv_spchw_test",
1194 srcs = [
1195 "test/f32-dwconv-spchw.cc",
1196 "test/dwconv-spchw-microkernel-tester.h",
1197 "src/xnnpack/AlignedAllocator.h",
1198 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1199 deps = MICROKERNEL_TEST_DEPS,
1200)
1201
1202xnnpack_unit_test(
1203 name = "f32_gavgpool_test",
1204 srcs = [
1205 "test/f32-gavgpool.cc",
1206 "test/gavgpool-microkernel-tester.h",
1207 "src/xnnpack/AlignedAllocator.h",
1208 ] + MICROKERNEL_TEST_HDRS,
1209 deps = MICROKERNEL_TEST_DEPS,
1210)
1211
1212xnnpack_unit_test(
1213 name = "f32_gavgpool_spchw_test",
1214 srcs = [
1215 "test/f32-gavgpool-spchw.cc",
1216 "test/gavgpool-spchw-microkernel-tester.h",
1217 "src/xnnpack/AlignedAllocator.h",
1218 ] + MICROKERNEL_TEST_HDRS,
1219 deps = MICROKERNEL_TEST_DEPS,
1220)
1221
1222xnnpack_unit_test(
1223 name = "f32_gemm_test",
1224 srcs = [
1225 "test/f32-gemm.cc",
1226 "test/gemm-microkernel-tester.h",
1227 "src/xnnpack/AlignedAllocator.h",
1228 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1229 deps = MICROKERNEL_TEST_DEPS,
1230)
1231
1232xnnpack_unit_test(
1233 name = "f32_gemminc_test",
1234 srcs = [
1235 "test/f32-gemminc.cc",
1236 "test/gemm-microkernel-tester.h",
1237 "src/xnnpack/AlignedAllocator.h",
1238 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1239 deps = MICROKERNEL_TEST_DEPS,
1240)
1241
1242xnnpack_unit_test(
1243 name = "f32_hswish_test",
1244 srcs = [
1245 "test/f32-hswish.cc",
1246 "test/hswish-microkernel-tester.h",
1247 ] + MICROKERNEL_TEST_HDRS,
1248 deps = MICROKERNEL_TEST_DEPS,
1249)
1250
1251xnnpack_unit_test(
1252 name = "f32_maxpool_test",
1253 srcs = [
1254 "test/f32-maxpool.cc",
1255 "test/maxpool-microkernel-tester.h",
1256 ] + MICROKERNEL_TEST_HDRS,
1257 deps = MICROKERNEL_TEST_DEPS,
1258)
1259
1260xnnpack_unit_test(
1261 name = "f32_pavgpool_test",
1262 srcs = [
1263 "test/f32-pavgpool.cc",
1264 "test/avgpool-microkernel-tester.h",
1265 "src/xnnpack/AlignedAllocator.h",
1266 ] + MICROKERNEL_TEST_HDRS,
1267 deps = MICROKERNEL_TEST_DEPS,
1268)
1269
1270xnnpack_unit_test(
1271 name = "f32_ppmm_test",
1272 srcs = [
1273 "test/f32-ppmm.cc",
1274 "test/gemm-microkernel-tester.h",
1275 "src/xnnpack/AlignedAllocator.h",
1276 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1277 deps = MICROKERNEL_TEST_DEPS,
1278)
1279
1280xnnpack_unit_test(
1281 name = "f32_prelu_test",
1282 srcs = [
1283 "test/f32-prelu.cc",
1284 "test/prelu-microkernel-tester.h",
1285 "src/xnnpack/AlignedAllocator.h",
1286 ] + MICROKERNEL_TEST_HDRS,
1287 deps = MICROKERNEL_TEST_DEPS,
1288)
1289
1290xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001291 name = "f32_raddexpminusmax_test",
1292 srcs = [
1293 "test/f32-raddexpminusmax.cc",
1294 "test/raddexpminusmax-microkernel-tester.h",
1295 ] + MICROKERNEL_TEST_HDRS,
1296 deps = MICROKERNEL_TEST_DEPS,
1297)
1298
1299xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001300 name = "f32_raddextexp_test",
1301 srcs = [
1302 "test/f32-raddextexp.cc",
1303 "test/raddextexp-microkernel-tester.h",
1304 ] + MICROKERNEL_TEST_HDRS,
1305 deps = MICROKERNEL_TEST_DEPS,
1306)
1307
1308xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001309 name = "f32_raddstoreexpminusmax_test",
1310 srcs = [
1311 "test/f32-raddstoreexpminusmax.cc",
1312 "test/raddstoreexpminusmax-microkernel-tester.h",
1313 ] + MICROKERNEL_TEST_HDRS,
1314 deps = MICROKERNEL_TEST_DEPS,
1315)
1316
1317xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001318 name = "f32_rmax_test",
1319 srcs = [
1320 "test/f32-rmax.cc",
1321 "test/rmax-microkernel-tester.h",
1322 ] + MICROKERNEL_TEST_HDRS,
1323 deps = MICROKERNEL_TEST_DEPS,
1324)
1325
1326xnnpack_unit_test(
1327 name = "f32_spmm_test",
1328 srcs = [
1329 "test/f32-spmm.cc",
1330 "test/spmm-microkernel-tester.h",
1331 "src/xnnpack/AlignedAllocator.h",
1332 ] + MICROKERNEL_TEST_HDRS,
1333 deps = MICROKERNEL_TEST_DEPS,
1334)
1335
1336xnnpack_unit_test(
1337 name = "f32_vadd_test",
1338 srcs = [
1339 "test/f32-vadd.cc",
1340 "test/vadd-microkernel-tester.h",
1341 ] + MICROKERNEL_TEST_HDRS,
1342 deps = MICROKERNEL_TEST_DEPS,
1343)
1344
1345xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001346 name = "f32_vmul_test",
1347 srcs = [
1348 "test/f32-vmul.cc",
1349 "test/vmul-microkernel-tester.h",
1350 ] + MICROKERNEL_TEST_HDRS,
1351 deps = MICROKERNEL_TEST_DEPS,
1352)
1353
1354xnnpack_unit_test(
1355 name = "f32_vmulcaddc_test",
1356 srcs = [
1357 "test/f32-vmulcaddc.cc",
1358 "test/vmulcaddc-microkernel-tester.h",
1359 "src/xnnpack/AlignedAllocator.h",
1360 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1361 deps = MICROKERNEL_TEST_DEPS,
1362)
1363
1364xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001365 name = "f32_vscale_test",
1366 srcs = [
1367 "test/f32-vscale.cc",
1368 "test/vscale-microkernel-tester.h",
1369 ] + MICROKERNEL_TEST_HDRS,
1370 deps = MICROKERNEL_TEST_DEPS,
1371)
1372
1373xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001374 name = "f32_vscaleexpminusmax_test",
1375 srcs = [
1376 "test/f32-vscaleexpminusmax.cc",
1377 "test/vscaleexpminusmax-microkernel-tester.h",
1378 ] + MICROKERNEL_TEST_HDRS,
1379 deps = MICROKERNEL_TEST_DEPS,
1380)
1381
1382xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001383 name = "f32_vscaleextexp_test",
1384 srcs = [
1385 "test/f32-vscaleextexp.cc",
1386 "test/vscaleextexp-microkernel-tester.h",
1387 ] + MICROKERNEL_TEST_HDRS,
1388 deps = MICROKERNEL_TEST_DEPS,
1389)
1390
1391xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001392 name = "f32_vsub_test",
1393 srcs = [
1394 "test/f32-vsub.cc",
1395 "test/vsub-microkernel-tester.h",
1396 ] + MICROKERNEL_TEST_HDRS,
1397 deps = MICROKERNEL_TEST_DEPS,
1398)
1399
1400xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001401 name = "q8_avgpool_test",
1402 srcs = [
1403 "test/q8-avgpool.cc",
1404 "test/avgpool-microkernel-tester.h",
1405 "src/xnnpack/AlignedAllocator.h",
1406 ] + MICROKERNEL_TEST_HDRS,
1407 deps = MICROKERNEL_TEST_DEPS,
1408)
1409
1410xnnpack_unit_test(
1411 name = "q8_igemm_test",
1412 srcs = [
1413 "test/q8-igemm.cc",
1414 "test/gemm-microkernel-tester.h",
1415 "src/xnnpack/AlignedAllocator.h",
1416 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1417 deps = MICROKERNEL_TEST_DEPS,
1418)
1419
1420xnnpack_unit_test(
1421 name = "q8_dwconv_test",
1422 srcs = [
1423 "test/q8-dwconv.cc",
1424 "test/dwconv-microkernel-tester.h",
1425 "src/xnnpack/AlignedAllocator.h",
1426 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1427 deps = MICROKERNEL_TEST_DEPS,
1428)
1429
1430xnnpack_unit_test(
1431 name = "q8_gavgpool_test",
1432 srcs = [
1433 "test/q8-gavgpool.cc",
1434 "test/gavgpool-microkernel-tester.h",
1435 "src/xnnpack/AlignedAllocator.h",
1436 ] + MICROKERNEL_TEST_HDRS,
1437 deps = MICROKERNEL_TEST_DEPS,
1438)
1439
1440xnnpack_unit_test(
1441 name = "q8_gemm_test",
1442 srcs = [
1443 "test/q8-gemm.cc",
1444 "test/gemm-microkernel-tester.h",
1445 "src/xnnpack/AlignedAllocator.h",
1446 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1447 deps = MICROKERNEL_TEST_DEPS,
1448)
1449
1450xnnpack_unit_test(
1451 name = "q8_vadd_test",
1452 srcs = [
1453 "test/q8-vadd.cc",
1454 "test/vadd-microkernel-tester.h",
1455 ] + MICROKERNEL_TEST_HDRS,
1456 deps = MICROKERNEL_TEST_DEPS,
1457)
1458
1459xnnpack_unit_test(
1460 name = "u8_clamp_test",
1461 srcs = [
1462 "test/u8-clamp.cc",
1463 "test/clamp-microkernel-tester.h",
1464 ] + MICROKERNEL_TEST_HDRS,
1465 deps = MICROKERNEL_TEST_DEPS,
1466)
1467
1468xnnpack_unit_test(
1469 name = "u8_lut32norm_test",
1470 srcs = [
1471 "test/u8-lut32norm.cc",
1472 "test/lut-norm-microkernel-tester.h",
1473 ] + MICROKERNEL_TEST_HDRS,
1474 deps = MICROKERNEL_TEST_DEPS,
1475)
1476
1477xnnpack_unit_test(
1478 name = "u8_maxpool_test",
1479 srcs = [
1480 "test/u8-maxpool.cc",
1481 "test/maxpool-microkernel-tester.h",
1482 ] + MICROKERNEL_TEST_HDRS,
1483 deps = MICROKERNEL_TEST_DEPS,
1484)
1485
1486xnnpack_unit_test(
1487 name = "u8_rmax_test",
1488 srcs = [
1489 "test/u8-rmax.cc",
1490 "test/rmax-microkernel-tester.h",
1491 ] + MICROKERNEL_TEST_HDRS,
1492 deps = MICROKERNEL_TEST_DEPS,
1493)
1494
1495xnnpack_unit_test(
1496 name = "x32_packx_test",
1497 srcs = [
1498 "test/x32-packx.cc",
1499 "test/pack-microkernel-tester.h",
1500 "src/xnnpack/AlignedAllocator.h",
1501 ] + MICROKERNEL_TEST_HDRS,
1502 deps = MICROKERNEL_TEST_DEPS,
1503)
1504
1505xnnpack_unit_test(
1506 name = "x32_pad_test",
1507 srcs = [
1508 "test/x32-pad.cc",
1509 "test/pad-microkernel-tester.h",
1510 ] + MICROKERNEL_TEST_HDRS,
1511 deps = MICROKERNEL_TEST_DEPS,
1512)
1513
1514xnnpack_unit_test(
1515 name = "x32_unpool_test",
1516 srcs = [
1517 "test/x32-unpool.cc",
1518 "test/unpool-microkernel-tester.h",
1519 ] + MICROKERNEL_TEST_HDRS,
1520 deps = MICROKERNEL_TEST_DEPS,
1521)
1522
1523xnnpack_unit_test(
1524 name = "x32_zip_test",
1525 srcs = [
1526 "test/x32-zip.cc",
1527 "test/zip-microkernel-tester.h",
1528 ] + MICROKERNEL_TEST_HDRS,
1529 deps = MICROKERNEL_TEST_DEPS,
1530)
1531
1532xnnpack_unit_test(
1533 name = "x8_lut_test",
1534 srcs = [
1535 "test/x8-lut.cc",
1536 "test/lut-microkernel-tester.h",
1537 ] + MICROKERNEL_TEST_HDRS,
1538 deps = MICROKERNEL_TEST_DEPS,
1539)
1540
1541xnnpack_unit_test(
1542 name = "x8_zip_test",
1543 srcs = [
1544 "test/x8-zip.cc",
1545 "test/zip-microkernel-tester.h",
1546 ] + MICROKERNEL_TEST_HDRS,
1547 deps = MICROKERNEL_TEST_DEPS,
1548)
1549
1550########################### Size test for the library ##########################
1551
1552xnnpack_binary(
1553 name = "size_test",
1554 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001555 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001556)
1557
1558########################### Unit tests for operators ###########################
1559
1560xnnpack_unit_test(
1561 name = "add_test",
1562 srcs = [
1563 "test/add.cc",
1564 "test/add-operator-tester.h",
1565 ],
1566 deps = OPERATOR_TEST_DEPS,
1567)
1568
1569xnnpack_unit_test(
1570 name = "argmax_pooling_test",
1571 srcs = [
1572 "test/argmax-pooling.cc",
1573 "test/argmax-pooling-operator-tester.h",
1574 ] + OPERATOR_TEST_PARAMS_HDRS,
1575 deps = OPERATOR_TEST_DEPS,
1576)
1577
1578xnnpack_unit_test(
1579 name = "average_pooling_test",
1580 srcs = [
1581 "test/average-pooling.cc",
1582 "test/average-pooling-operator-tester.h",
1583 ] + OPERATOR_TEST_PARAMS_HDRS,
1584 deps = OPERATOR_TEST_DEPS,
1585)
1586
1587xnnpack_unit_test(
1588 name = "channel_pad_test",
1589 srcs = [
1590 "test/channel-pad.cc",
1591 "test/channel-pad-operator-tester.h",
1592 ] + OPERATOR_TEST_PARAMS_HDRS,
1593 deps = OPERATOR_TEST_DEPS,
1594)
1595
1596xnnpack_unit_test(
1597 name = "channel_shuffle_test",
1598 srcs = [
1599 "test/channel-shuffle.cc",
1600 "test/channel-shuffle-operator-tester.h",
1601 ],
1602 deps = OPERATOR_TEST_DEPS,
1603)
1604
1605xnnpack_unit_test(
1606 name = "clamp_test",
1607 srcs = [
1608 "test/clamp.cc",
1609 "test/clamp-operator-tester.h",
1610 ],
1611 deps = OPERATOR_TEST_DEPS,
1612)
1613
1614xnnpack_unit_test(
1615 name = "convolution_test",
1616 srcs = [
1617 "test/convolution.cc",
1618 "test/convolution-operator-tester.h",
1619 ],
1620 deps = OPERATOR_TEST_DEPS,
1621)
1622
1623xnnpack_unit_test(
1624 name = "convolution_spnchw_test",
1625 srcs = [
1626 "test/convolution-spnchw.cc",
1627 "test/convolution-spnchw-operator-tester.h",
1628 ],
1629 deps = OPERATOR_TEST_DEPS,
1630)
1631
1632xnnpack_unit_test(
1633 name = "deconvolution_test",
1634 srcs = [
1635 "test/deconvolution.cc",
1636 "test/deconvolution-operator-tester.h",
1637 ] + OPERATOR_TEST_PARAMS_HDRS,
1638 deps = OPERATOR_TEST_DEPS,
1639)
1640
1641xnnpack_unit_test(
1642 name = "fully_connected_test",
1643 srcs = [
1644 "test/fully-connected.cc",
1645 "test/fully-connected-operator-tester.h",
1646 ],
1647 deps = OPERATOR_TEST_DEPS,
1648)
1649
1650xnnpack_unit_test(
1651 name = "global_average_pooling_test",
1652 srcs = [
1653 "test/global-average-pooling.cc",
1654 "test/global-average-pooling-operator-tester.h",
1655 ] + OPERATOR_TEST_PARAMS_HDRS,
1656 deps = OPERATOR_TEST_DEPS,
1657)
1658
1659xnnpack_unit_test(
1660 name = "global_average_pooling_spnchw_test",
1661 srcs = [
1662 "test/global-average-pooling-spnchw.cc",
1663 "test/global-average-pooling-spnchw-operator-tester.h",
1664 ],
1665 deps = OPERATOR_TEST_DEPS,
1666)
1667
1668xnnpack_unit_test(
1669 name = "hardswish_test",
1670 srcs = [
1671 "test/hardswish.cc",
1672 "test/hardswish-operator-tester.h",
1673 ],
1674 deps = OPERATOR_TEST_DEPS,
1675)
1676
1677xnnpack_unit_test(
1678 name = "leaky_relu_test",
1679 srcs = [
1680 "test/leaky-relu.cc",
1681 "test/leaky-relu-operator-tester.h",
1682 ],
1683 deps = OPERATOR_TEST_DEPS,
1684)
1685
1686xnnpack_unit_test(
1687 name = "max_pooling_test",
1688 srcs = [
1689 "test/max-pooling.cc",
1690 "test/max-pooling-operator-tester.h",
1691 ] + OPERATOR_TEST_PARAMS_HDRS,
1692 deps = OPERATOR_TEST_DEPS,
1693)
1694
1695xnnpack_unit_test(
1696 name = "prelu_test",
1697 srcs = [
1698 "test/prelu.cc",
1699 "test/prelu-operator-tester.h",
1700 ] + OPERATOR_TEST_PARAMS_HDRS,
1701 deps = OPERATOR_TEST_DEPS,
1702)
1703
1704xnnpack_unit_test(
1705 name = "sigmoid_test",
1706 srcs = [
1707 "test/sigmoid.cc",
1708 "test/sigmoid-operator-tester.h",
1709 ],
1710 deps = OPERATOR_TEST_DEPS,
1711)
1712
1713xnnpack_unit_test(
1714 name = "softargmax_test",
1715 srcs = [
1716 "test/softargmax.cc",
1717 "test/softargmax-operator-tester.h",
1718 ],
1719 deps = OPERATOR_TEST_DEPS,
1720)
1721
1722xnnpack_unit_test(
1723 name = "unpooling_test",
1724 srcs = [
1725 "test/unpooling.cc",
1726 "test/unpooling-operator-tester.h",
1727 ],
1728 deps = OPERATOR_TEST_DEPS,
1729)
1730
1731############################# Build configurations #############################
1732
1733config_setting(
1734 name = "linux_k8",
1735 values = {
1736 "cpu": "k8",
1737 },
1738)
1739
1740config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07001741 name = "linux_aarch64",
1742 values = {
1743 "cpu": "aarch64",
1744 },
1745)
1746
1747config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001748 name = "android",
1749 values = {"crosstool_top": "//external:android/crosstool"},
1750)
1751
1752config_setting(
1753 name = "android_armv7",
1754 values = {
1755 "crosstool_top": "//external:android/crosstool",
1756 "cpu": "armeabi-v7a",
1757 },
1758)
1759
1760config_setting(
1761 name = "android_arm64",
1762 values = {
1763 "crosstool_top": "//external:android/crosstool",
1764 "cpu": "arm64-v8a",
1765 },
1766)
1767
1768config_setting(
1769 name = "android_x86",
1770 values = {
1771 "crosstool_top": "//external:android/crosstool",
1772 "cpu": "x86",
1773 },
1774)
1775
1776config_setting(
1777 name = "android_x86_64",
1778 values = {
1779 "crosstool_top": "//external:android/crosstool",
1780 "cpu": "x86_64",
1781 },
1782)
1783
1784config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07001785 name = "macos_x86_64",
1786 values = {
1787 "apple_platform_type": "macos",
1788 "cpu": "darwin",
1789 },
1790)
1791
1792config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001793 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001794 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07001795)
1796
1797config_setting(
1798 name = "emscripten_wasm",
1799 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001800 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001801 "cpu": "wasm",
1802 },
1803)
1804
1805config_setting(
1806 name = "emscripten_wasmsimd",
1807 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001808 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001809 "cpu": "wasm",
1810 "features": "wasmsimd",
1811 },
1812)
1813
1814config_setting(
1815 name = "emscripten_asmjs",
1816 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001817 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001818 "cpu": "asmjs",
1819 },
1820)
1821
Marat Dukhan174706e2019-10-09 16:33:53 -07001822# Builds with -c dbg
Marat Dukhan08c4a432019-10-03 09:29:21 -07001823config_setting(
1824 name = "debug_build",
1825 values = {
1826 "compilation_mode": "dbg",
1827 },
Marat Dukhan08c4a432019-10-03 09:29:21 -07001828)
1829
Marat Dukhan174706e2019-10-09 16:33:53 -07001830# Builds with -c opt
Marat Dukhan08c4a432019-10-03 09:29:21 -07001831config_setting(
1832 name = "optimized_build",
1833 values = {
1834 "compilation_mode": "opt",
1835 },
Marat Dukhan08c4a432019-10-03 09:29:21 -07001836)
1837
1838# Enables usage of assembly kernels.
1839config_setting(
1840 name = "xnn_enable_assembly_explicit_true",
1841 define_values = {"xnn_enable_assembly": "true"},
1842)
1843
1844# Disables usage of assembly kernels.
1845config_setting(
1846 name = "xnn_enable_assembly_explicit_false",
1847 define_values = {"xnn_enable_assembly": "false"},
1848)