blob: b80d9db5aa4ae44c7c45eb5125d8d98dd01b7cb9 [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 Dukhanbad48fe2019-11-04 10:35:22 -0800863 deps = [
864 "@com_google_benchmark//:benchmark",
865 "@cpuinfo",
866 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700867)
868
Frank Barchard7e955972019-10-11 10:34:25 -0700869######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -0700870
871xnnpack_benchmark(
872 name = "q8_gemm_bench",
873 srcs = [
874 "bench/gemm.h",
875 "bench/q8-gemm.cc",
876 "src/xnnpack/AlignedAllocator.h",
877 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
878 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
879 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
880)
881
882xnnpack_benchmark(
883 name = "f16_gemm_bench",
884 srcs = [
885 "bench/f16-gemm.cc",
886 "bench/gemm.h",
887 "src/xnnpack/AlignedAllocator.h",
888 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
889 copts = ["-Wno-unused-function"],
890 deps = MICROKERNEL_BENCHMARK_DEPS,
891)
892
893xnnpack_benchmark(
894 name = "f32_igemm_bench",
895 srcs = [
896 "bench/f32-igemm.cc",
897 "bench/conv.h",
898 "src/xnnpack/AlignedAllocator.h",
899 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -0700900 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700901)
902
903xnnpack_benchmark(
904 name = "f32_conv_hwc_bench",
905 srcs = [
906 "bench/f32-conv-hwc.cc",
907 "bench/dconv.h",
908 "src/xnnpack/AlignedAllocator.h",
909 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
910 copts = ["-Wno-unused-function"],
911 deps = MICROKERNEL_BENCHMARK_DEPS,
912)
913
914xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -0700915 name = "f32_conv_hwc2spchw_bench",
916 srcs = [
917 "bench/f32-conv-hwc2spchw.cc",
918 "bench/dconv.h",
919 "src/xnnpack/AlignedAllocator.h",
920 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
921 copts = ["-Wno-unused-function"],
922 deps = MICROKERNEL_BENCHMARK_DEPS,
923)
924
925xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700926 name = "f32_dwconv_bench",
927 srcs = [
928 "bench/f32-dwconv.cc",
929 "bench/dwconv.h",
930 "src/xnnpack/AlignedAllocator.h",
931 ] + MICROKERNEL_BENCHMARK_HDRS,
932 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
933)
934
935xnnpack_benchmark(
936 name = "f32_dwconv_spchw_bench",
937 srcs = [
938 "bench/f32-dwconv-spchw.cc",
939 "bench/dwconv.h",
940 "src/xnnpack/AlignedAllocator.h",
941 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
942 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
943)
944
945xnnpack_benchmark(
946 name = "f32_gemm_bench",
947 srcs = [
948 "bench/f32-gemm.cc",
949 "bench/gemm.h",
950 "src/xnnpack/AlignedAllocator.h",
951 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
952 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -0700953 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700954)
955
956xnnpack_benchmark(
957 name = "f32_rmax_bench",
958 srcs = [
959 "bench/f32-rmax.cc",
960 "src/xnnpack/AlignedAllocator.h",
961 ] + MICROKERNEL_BENCHMARK_HDRS,
962 deps = MICROKERNEL_BENCHMARK_DEPS,
963)
964
965xnnpack_benchmark(
966 name = "f32_spmm_bench",
967 srcs = [
968 "bench/f32-spmm.cc",
969 "bench/gemm.h",
970 "src/xnnpack/AlignedAllocator.h",
971 ] + MICROKERNEL_BENCHMARK_HDRS,
972 copts = ["-Wno-unused-function"],
973 deps = MICROKERNEL_BENCHMARK_DEPS,
974)
975
976xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -0700977 name = "f32_softargmax_bench",
978 srcs = [
979 "bench/f32-softargmax.cc",
980 ] + MICROKERNEL_BENCHMARK_HDRS,
981 copts = ["-Wno-unused-function"],
982 deps = MICROKERNEL_BENCHMARK_DEPS,
983)
984
985xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700986 name = "f32_im2col_gemm_bench",
987 srcs = [
988 "bench/f32-im2col-gemm.cc",
989 "bench/conv.h",
990 "src/xnnpack/AlignedAllocator.h",
991 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
992 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
993)
994
995########################### Benchmarks for operators ###########################
996
997xnnpack_benchmark(
998 name = "add_bench",
999 srcs = ["bench/add.cc"],
1000 deps = OPERATOR_BENCHMARK_DEPS,
1001)
1002
1003xnnpack_benchmark(
1004 name = "average_pooling_bench",
1005 srcs = ["bench/average-pooling.cc"],
1006 deps = OPERATOR_BENCHMARK_DEPS,
1007)
1008
1009xnnpack_benchmark(
1010 name = "channel_shuffle_bench",
1011 srcs = ["bench/channel-shuffle.cc"],
1012 deps = OPERATOR_BENCHMARK_DEPS,
1013)
1014
1015xnnpack_benchmark(
1016 name = "convolution_bench",
1017 srcs = ["bench/convolution.cc"],
1018 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1019 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1020)
1021
1022xnnpack_benchmark(
1023 name = "deconvolution_bench",
1024 srcs = ["bench/deconvolution.cc"],
1025 copts = xnnpack_optional_tflite_copts(),
1026 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1027)
1028
1029xnnpack_benchmark(
1030 name = "global_average_pooling_bench",
1031 srcs = ["bench/global-average-pooling.cc"],
1032 deps = OPERATOR_BENCHMARK_DEPS,
1033)
1034
1035xnnpack_benchmark(
1036 name = "max_pooling_bench",
1037 srcs = ["bench/max-pooling.cc"],
1038 deps = OPERATOR_BENCHMARK_DEPS,
1039)
1040
1041xnnpack_benchmark(
1042 name = "sigmoid_bench",
1043 srcs = ["bench/sigmoid.cc"],
1044 deps = OPERATOR_BENCHMARK_DEPS,
1045)
1046
1047xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07001048 name = "prelu_bench",
1049 srcs = ["bench/prelu.cc"],
1050 copts = xnnpack_optional_tflite_copts(),
1051 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1052)
1053
1054xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001055 name = "softargmax_bench",
1056 srcs = ["bench/softargmax.cc"],
1057 deps = OPERATOR_BENCHMARK_DEPS,
1058)
1059
Marat Dukhanc068bb62019-10-04 13:24:39 -07001060############################# End-to-end benchmarks ############################
1061
1062cc_library(
1063 name = "mobilenet_v1",
1064 srcs = ["models/mobilenet-v1.cc"],
1065 hdrs = ["models/models.h"],
1066 linkstatic = True,
1067 deps = [
1068 ":XNNPACK",
1069 "@pthreadpool",
1070 ],
1071)
1072
1073cc_library(
1074 name = "mobilenet_v2",
1075 srcs = ["models/mobilenet-v2.cc"],
1076 hdrs = ["models/models.h"],
1077 linkstatic = True,
1078 deps = [
1079 ":XNNPACK",
1080 "@pthreadpool",
1081 ],
1082)
1083
1084xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07001085 name = "f32_dwconv_e2e_bench",
1086 srcs = ["bench/f32-dwconv-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1087 copts = ["-Wno-unused-function"],
1088 deps = MICROKERNEL_BENCHMARK_DEPS + [
1089 ":XNNPACK",
1090 ":mobilenet_v1",
1091 ":mobilenet_v2",
1092 ],
1093)
1094
1095xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07001096 name = "f32_gemm_e2e_bench",
1097 srcs = ["bench/f32-gemm-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1098 copts = ["-Wno-unused-function"],
1099 deps = MICROKERNEL_BENCHMARK_DEPS + [
1100 ":XNNPACK",
1101 ":mobilenet_v1",
1102 ":mobilenet_v2",
1103 ],
1104)
1105
1106xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07001107 name = "end2end_bench",
1108 srcs = ["bench/end2end.cc"],
1109 deps = [
1110 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07001111 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07001112 ":mobilenet_v1",
1113 ":mobilenet_v2",
1114 "@pthreadpool",
1115 ],
1116)
1117
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001118#################### Accuracy evaluation for math functions ####################
1119
1120xnnpack_benchmark(
1121 name = "f32_exp_eval",
1122 srcs = [
1123 "eval/f32-exp.cc",
1124 "src/xnnpack/AlignedAllocator.h",
1125 ] + ACCURACY_EVAL_HDRS,
1126 deps = ACCURACY_EVAL_DEPS,
1127)
1128
Marat Dukhan515c9772019-10-17 18:07:57 -07001129xnnpack_benchmark(
1130 name = "f32_expminus_eval",
1131 srcs = [
1132 "eval/f32-expminus.cc",
1133 "src/xnnpack/AlignedAllocator.h",
1134 ] + ACCURACY_EVAL_HDRS,
1135 deps = ACCURACY_EVAL_DEPS,
1136)
1137
Marat Dukhan98ba4412019-10-23 02:14:28 -07001138xnnpack_benchmark(
1139 name = "f32_extexp_eval",
1140 srcs = [
1141 "eval/f32-extexp.cc",
1142 "src/xnnpack/AlignedAllocator.h",
1143 ] + ACCURACY_EVAL_HDRS,
1144 deps = ACCURACY_EVAL_DEPS,
1145)
1146
Marat Dukhan08c4a432019-10-03 09:29:21 -07001147######################### Unit tests for micro-kernels #########################
1148
1149xnnpack_unit_test(
1150 name = "f16_gemm_test",
1151 srcs = [
1152 "test/f16-gemm.cc",
1153 "test/gemm-microkernel-tester.h",
1154 "src/xnnpack/AlignedAllocator.h",
1155 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1156 deps = MICROKERNEL_TEST_DEPS,
1157)
1158
1159xnnpack_unit_test(
1160 name = "f32_argmaxpool_test",
1161 srcs = [
1162 "test/f32-argmaxpool.cc",
1163 "test/argmaxpool-microkernel-tester.h",
1164 "src/xnnpack/AlignedAllocator.h",
1165 ] + MICROKERNEL_TEST_HDRS,
1166 deps = MICROKERNEL_TEST_DEPS,
1167)
1168
1169xnnpack_unit_test(
1170 name = "f32_avgpool_test",
1171 srcs = [
1172 "test/f32-avgpool.cc",
1173 "test/avgpool-microkernel-tester.h",
1174 "src/xnnpack/AlignedAllocator.h",
1175 ] + MICROKERNEL_TEST_HDRS,
1176 deps = MICROKERNEL_TEST_DEPS,
1177)
1178
1179xnnpack_unit_test(
1180 name = "f32_clamp_test",
1181 srcs = [
1182 "test/f32-clamp.cc",
1183 "test/clamp-microkernel-tester.h",
1184 ] + MICROKERNEL_TEST_HDRS,
1185 deps = MICROKERNEL_TEST_DEPS,
1186)
1187
1188xnnpack_unit_test(
1189 name = "f32_igemm_test",
1190 srcs = [
1191 "test/f32-igemm.cc",
1192 "test/gemm-microkernel-tester.h",
1193 "src/xnnpack/AlignedAllocator.h",
1194 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1195 deps = MICROKERNEL_TEST_DEPS,
1196)
1197
1198xnnpack_unit_test(
1199 name = "f32_conv_hwc_test",
1200 srcs = [
1201 "test/f32-conv-hwc.cc",
1202 "test/conv-hwc-microkernel-tester.h",
1203 "src/xnnpack/AlignedAllocator.h",
1204 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1205 deps = MICROKERNEL_TEST_DEPS,
1206)
1207
1208xnnpack_unit_test(
1209 name = "f32_conv_hwc2spchw_test",
1210 srcs = [
1211 "test/f32-conv-hwc2spchw.cc",
1212 "test/conv-hwc2spchw-microkernel-tester.h",
1213 "src/xnnpack/AlignedAllocator.h",
1214 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1215 deps = MICROKERNEL_TEST_DEPS,
1216)
1217
1218xnnpack_unit_test(
1219 name = "f32_dwconv_test",
1220 srcs = [
1221 "test/f32-dwconv.cc",
1222 "test/dwconv-microkernel-tester.h",
1223 "src/xnnpack/AlignedAllocator.h",
1224 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1225 deps = MICROKERNEL_TEST_DEPS,
1226)
1227
1228xnnpack_unit_test(
1229 name = "f32_dwconv_spchw_test",
1230 srcs = [
1231 "test/f32-dwconv-spchw.cc",
1232 "test/dwconv-spchw-microkernel-tester.h",
1233 "src/xnnpack/AlignedAllocator.h",
1234 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1235 deps = MICROKERNEL_TEST_DEPS,
1236)
1237
1238xnnpack_unit_test(
1239 name = "f32_gavgpool_test",
1240 srcs = [
1241 "test/f32-gavgpool.cc",
1242 "test/gavgpool-microkernel-tester.h",
1243 "src/xnnpack/AlignedAllocator.h",
1244 ] + MICROKERNEL_TEST_HDRS,
1245 deps = MICROKERNEL_TEST_DEPS,
1246)
1247
1248xnnpack_unit_test(
1249 name = "f32_gavgpool_spchw_test",
1250 srcs = [
1251 "test/f32-gavgpool-spchw.cc",
1252 "test/gavgpool-spchw-microkernel-tester.h",
1253 "src/xnnpack/AlignedAllocator.h",
1254 ] + MICROKERNEL_TEST_HDRS,
1255 deps = MICROKERNEL_TEST_DEPS,
1256)
1257
1258xnnpack_unit_test(
1259 name = "f32_gemm_test",
1260 srcs = [
1261 "test/f32-gemm.cc",
1262 "test/gemm-microkernel-tester.h",
1263 "src/xnnpack/AlignedAllocator.h",
1264 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1265 deps = MICROKERNEL_TEST_DEPS,
1266)
1267
1268xnnpack_unit_test(
1269 name = "f32_gemminc_test",
1270 srcs = [
1271 "test/f32-gemminc.cc",
1272 "test/gemm-microkernel-tester.h",
1273 "src/xnnpack/AlignedAllocator.h",
1274 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1275 deps = MICROKERNEL_TEST_DEPS,
1276)
1277
1278xnnpack_unit_test(
1279 name = "f32_hswish_test",
1280 srcs = [
1281 "test/f32-hswish.cc",
1282 "test/hswish-microkernel-tester.h",
1283 ] + MICROKERNEL_TEST_HDRS,
1284 deps = MICROKERNEL_TEST_DEPS,
1285)
1286
1287xnnpack_unit_test(
1288 name = "f32_maxpool_test",
1289 srcs = [
1290 "test/f32-maxpool.cc",
1291 "test/maxpool-microkernel-tester.h",
1292 ] + MICROKERNEL_TEST_HDRS,
1293 deps = MICROKERNEL_TEST_DEPS,
1294)
1295
1296xnnpack_unit_test(
1297 name = "f32_pavgpool_test",
1298 srcs = [
1299 "test/f32-pavgpool.cc",
1300 "test/avgpool-microkernel-tester.h",
1301 "src/xnnpack/AlignedAllocator.h",
1302 ] + MICROKERNEL_TEST_HDRS,
1303 deps = MICROKERNEL_TEST_DEPS,
1304)
1305
1306xnnpack_unit_test(
1307 name = "f32_ppmm_test",
1308 srcs = [
1309 "test/f32-ppmm.cc",
1310 "test/gemm-microkernel-tester.h",
1311 "src/xnnpack/AlignedAllocator.h",
1312 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1313 deps = MICROKERNEL_TEST_DEPS,
1314)
1315
1316xnnpack_unit_test(
1317 name = "f32_prelu_test",
1318 srcs = [
1319 "test/f32-prelu.cc",
1320 "test/prelu-microkernel-tester.h",
1321 "src/xnnpack/AlignedAllocator.h",
1322 ] + MICROKERNEL_TEST_HDRS,
1323 deps = MICROKERNEL_TEST_DEPS,
1324)
1325
1326xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001327 name = "f32_raddexpminusmax_test",
1328 srcs = [
1329 "test/f32-raddexpminusmax.cc",
1330 "test/raddexpminusmax-microkernel-tester.h",
1331 ] + MICROKERNEL_TEST_HDRS,
1332 deps = MICROKERNEL_TEST_DEPS,
1333)
1334
1335xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001336 name = "f32_raddextexp_test",
1337 srcs = [
1338 "test/f32-raddextexp.cc",
1339 "test/raddextexp-microkernel-tester.h",
1340 ] + MICROKERNEL_TEST_HDRS,
1341 deps = MICROKERNEL_TEST_DEPS,
1342)
1343
1344xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001345 name = "f32_raddstoreexpminusmax_test",
1346 srcs = [
1347 "test/f32-raddstoreexpminusmax.cc",
1348 "test/raddstoreexpminusmax-microkernel-tester.h",
1349 ] + MICROKERNEL_TEST_HDRS,
1350 deps = MICROKERNEL_TEST_DEPS,
1351)
1352
1353xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001354 name = "f32_rmax_test",
1355 srcs = [
1356 "test/f32-rmax.cc",
1357 "test/rmax-microkernel-tester.h",
1358 ] + MICROKERNEL_TEST_HDRS,
1359 deps = MICROKERNEL_TEST_DEPS,
1360)
1361
1362xnnpack_unit_test(
1363 name = "f32_spmm_test",
1364 srcs = [
1365 "test/f32-spmm.cc",
1366 "test/spmm-microkernel-tester.h",
1367 "src/xnnpack/AlignedAllocator.h",
1368 ] + MICROKERNEL_TEST_HDRS,
1369 deps = MICROKERNEL_TEST_DEPS,
1370)
1371
1372xnnpack_unit_test(
1373 name = "f32_vadd_test",
1374 srcs = [
1375 "test/f32-vadd.cc",
1376 "test/vadd-microkernel-tester.h",
1377 ] + MICROKERNEL_TEST_HDRS,
1378 deps = MICROKERNEL_TEST_DEPS,
1379)
1380
1381xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001382 name = "f32_vmul_test",
1383 srcs = [
1384 "test/f32-vmul.cc",
1385 "test/vmul-microkernel-tester.h",
1386 ] + MICROKERNEL_TEST_HDRS,
1387 deps = MICROKERNEL_TEST_DEPS,
1388)
1389
1390xnnpack_unit_test(
1391 name = "f32_vmulcaddc_test",
1392 srcs = [
1393 "test/f32-vmulcaddc.cc",
1394 "test/vmulcaddc-microkernel-tester.h",
1395 "src/xnnpack/AlignedAllocator.h",
1396 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1397 deps = MICROKERNEL_TEST_DEPS,
1398)
1399
1400xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001401 name = "f32_vscale_test",
1402 srcs = [
1403 "test/f32-vscale.cc",
1404 "test/vscale-microkernel-tester.h",
1405 ] + MICROKERNEL_TEST_HDRS,
1406 deps = MICROKERNEL_TEST_DEPS,
1407)
1408
1409xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001410 name = "f32_vscaleexpminusmax_test",
1411 srcs = [
1412 "test/f32-vscaleexpminusmax.cc",
1413 "test/vscaleexpminusmax-microkernel-tester.h",
1414 ] + MICROKERNEL_TEST_HDRS,
1415 deps = MICROKERNEL_TEST_DEPS,
1416)
1417
1418xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001419 name = "f32_vscaleextexp_test",
1420 srcs = [
1421 "test/f32-vscaleextexp.cc",
1422 "test/vscaleextexp-microkernel-tester.h",
1423 ] + MICROKERNEL_TEST_HDRS,
1424 deps = MICROKERNEL_TEST_DEPS,
1425)
1426
1427xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001428 name = "f32_vsub_test",
1429 srcs = [
1430 "test/f32-vsub.cc",
1431 "test/vsub-microkernel-tester.h",
1432 ] + MICROKERNEL_TEST_HDRS,
1433 deps = MICROKERNEL_TEST_DEPS,
1434)
1435
1436xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001437 name = "q8_avgpool_test",
1438 srcs = [
1439 "test/q8-avgpool.cc",
1440 "test/avgpool-microkernel-tester.h",
1441 "src/xnnpack/AlignedAllocator.h",
1442 ] + MICROKERNEL_TEST_HDRS,
1443 deps = MICROKERNEL_TEST_DEPS,
1444)
1445
1446xnnpack_unit_test(
1447 name = "q8_igemm_test",
1448 srcs = [
1449 "test/q8-igemm.cc",
1450 "test/gemm-microkernel-tester.h",
1451 "src/xnnpack/AlignedAllocator.h",
1452 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1453 deps = MICROKERNEL_TEST_DEPS,
1454)
1455
1456xnnpack_unit_test(
1457 name = "q8_dwconv_test",
1458 srcs = [
1459 "test/q8-dwconv.cc",
1460 "test/dwconv-microkernel-tester.h",
1461 "src/xnnpack/AlignedAllocator.h",
1462 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1463 deps = MICROKERNEL_TEST_DEPS,
1464)
1465
1466xnnpack_unit_test(
1467 name = "q8_gavgpool_test",
1468 srcs = [
1469 "test/q8-gavgpool.cc",
1470 "test/gavgpool-microkernel-tester.h",
1471 "src/xnnpack/AlignedAllocator.h",
1472 ] + MICROKERNEL_TEST_HDRS,
1473 deps = MICROKERNEL_TEST_DEPS,
1474)
1475
1476xnnpack_unit_test(
1477 name = "q8_gemm_test",
1478 srcs = [
1479 "test/q8-gemm.cc",
1480 "test/gemm-microkernel-tester.h",
1481 "src/xnnpack/AlignedAllocator.h",
1482 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1483 deps = MICROKERNEL_TEST_DEPS,
1484)
1485
1486xnnpack_unit_test(
1487 name = "q8_vadd_test",
1488 srcs = [
1489 "test/q8-vadd.cc",
1490 "test/vadd-microkernel-tester.h",
1491 ] + MICROKERNEL_TEST_HDRS,
1492 deps = MICROKERNEL_TEST_DEPS,
1493)
1494
1495xnnpack_unit_test(
1496 name = "u8_clamp_test",
1497 srcs = [
1498 "test/u8-clamp.cc",
1499 "test/clamp-microkernel-tester.h",
1500 ] + MICROKERNEL_TEST_HDRS,
1501 deps = MICROKERNEL_TEST_DEPS,
1502)
1503
1504xnnpack_unit_test(
1505 name = "u8_lut32norm_test",
1506 srcs = [
1507 "test/u8-lut32norm.cc",
1508 "test/lut-norm-microkernel-tester.h",
1509 ] + MICROKERNEL_TEST_HDRS,
1510 deps = MICROKERNEL_TEST_DEPS,
1511)
1512
1513xnnpack_unit_test(
1514 name = "u8_maxpool_test",
1515 srcs = [
1516 "test/u8-maxpool.cc",
1517 "test/maxpool-microkernel-tester.h",
1518 ] + MICROKERNEL_TEST_HDRS,
1519 deps = MICROKERNEL_TEST_DEPS,
1520)
1521
1522xnnpack_unit_test(
1523 name = "u8_rmax_test",
1524 srcs = [
1525 "test/u8-rmax.cc",
1526 "test/rmax-microkernel-tester.h",
1527 ] + MICROKERNEL_TEST_HDRS,
1528 deps = MICROKERNEL_TEST_DEPS,
1529)
1530
1531xnnpack_unit_test(
1532 name = "x32_packx_test",
1533 srcs = [
1534 "test/x32-packx.cc",
1535 "test/pack-microkernel-tester.h",
1536 "src/xnnpack/AlignedAllocator.h",
1537 ] + MICROKERNEL_TEST_HDRS,
1538 deps = MICROKERNEL_TEST_DEPS,
1539)
1540
1541xnnpack_unit_test(
1542 name = "x32_pad_test",
1543 srcs = [
1544 "test/x32-pad.cc",
1545 "test/pad-microkernel-tester.h",
1546 ] + MICROKERNEL_TEST_HDRS,
1547 deps = MICROKERNEL_TEST_DEPS,
1548)
1549
1550xnnpack_unit_test(
1551 name = "x32_unpool_test",
1552 srcs = [
1553 "test/x32-unpool.cc",
1554 "test/unpool-microkernel-tester.h",
1555 ] + MICROKERNEL_TEST_HDRS,
1556 deps = MICROKERNEL_TEST_DEPS,
1557)
1558
1559xnnpack_unit_test(
1560 name = "x32_zip_test",
1561 srcs = [
1562 "test/x32-zip.cc",
1563 "test/zip-microkernel-tester.h",
1564 ] + MICROKERNEL_TEST_HDRS,
1565 deps = MICROKERNEL_TEST_DEPS,
1566)
1567
1568xnnpack_unit_test(
1569 name = "x8_lut_test",
1570 srcs = [
1571 "test/x8-lut.cc",
1572 "test/lut-microkernel-tester.h",
1573 ] + MICROKERNEL_TEST_HDRS,
1574 deps = MICROKERNEL_TEST_DEPS,
1575)
1576
1577xnnpack_unit_test(
1578 name = "x8_zip_test",
1579 srcs = [
1580 "test/x8-zip.cc",
1581 "test/zip-microkernel-tester.h",
1582 ] + MICROKERNEL_TEST_HDRS,
1583 deps = MICROKERNEL_TEST_DEPS,
1584)
1585
1586########################### Size test for the library ##########################
1587
1588xnnpack_binary(
1589 name = "size_test",
1590 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001591 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001592)
1593
1594########################### Unit tests for operators ###########################
1595
1596xnnpack_unit_test(
1597 name = "add_test",
1598 srcs = [
1599 "test/add.cc",
1600 "test/add-operator-tester.h",
1601 ],
1602 deps = OPERATOR_TEST_DEPS,
1603)
1604
1605xnnpack_unit_test(
1606 name = "argmax_pooling_test",
1607 srcs = [
1608 "test/argmax-pooling.cc",
1609 "test/argmax-pooling-operator-tester.h",
1610 ] + OPERATOR_TEST_PARAMS_HDRS,
1611 deps = OPERATOR_TEST_DEPS,
1612)
1613
1614xnnpack_unit_test(
1615 name = "average_pooling_test",
1616 srcs = [
1617 "test/average-pooling.cc",
1618 "test/average-pooling-operator-tester.h",
1619 ] + OPERATOR_TEST_PARAMS_HDRS,
1620 deps = OPERATOR_TEST_DEPS,
1621)
1622
1623xnnpack_unit_test(
1624 name = "channel_pad_test",
1625 srcs = [
1626 "test/channel-pad.cc",
1627 "test/channel-pad-operator-tester.h",
1628 ] + OPERATOR_TEST_PARAMS_HDRS,
1629 deps = OPERATOR_TEST_DEPS,
1630)
1631
1632xnnpack_unit_test(
1633 name = "channel_shuffle_test",
1634 srcs = [
1635 "test/channel-shuffle.cc",
1636 "test/channel-shuffle-operator-tester.h",
1637 ],
1638 deps = OPERATOR_TEST_DEPS,
1639)
1640
1641xnnpack_unit_test(
1642 name = "clamp_test",
1643 srcs = [
1644 "test/clamp.cc",
1645 "test/clamp-operator-tester.h",
1646 ],
1647 deps = OPERATOR_TEST_DEPS,
1648)
1649
1650xnnpack_unit_test(
1651 name = "convolution_test",
1652 srcs = [
1653 "test/convolution.cc",
1654 "test/convolution-operator-tester.h",
1655 ],
1656 deps = OPERATOR_TEST_DEPS,
1657)
1658
1659xnnpack_unit_test(
1660 name = "convolution_spnchw_test",
1661 srcs = [
1662 "test/convolution-spnchw.cc",
1663 "test/convolution-spnchw-operator-tester.h",
1664 ],
1665 deps = OPERATOR_TEST_DEPS,
1666)
1667
1668xnnpack_unit_test(
1669 name = "deconvolution_test",
1670 srcs = [
1671 "test/deconvolution.cc",
1672 "test/deconvolution-operator-tester.h",
1673 ] + OPERATOR_TEST_PARAMS_HDRS,
1674 deps = OPERATOR_TEST_DEPS,
1675)
1676
1677xnnpack_unit_test(
1678 name = "fully_connected_test",
1679 srcs = [
1680 "test/fully-connected.cc",
1681 "test/fully-connected-operator-tester.h",
1682 ],
1683 deps = OPERATOR_TEST_DEPS,
1684)
1685
1686xnnpack_unit_test(
1687 name = "global_average_pooling_test",
1688 srcs = [
1689 "test/global-average-pooling.cc",
1690 "test/global-average-pooling-operator-tester.h",
1691 ] + OPERATOR_TEST_PARAMS_HDRS,
1692 deps = OPERATOR_TEST_DEPS,
1693)
1694
1695xnnpack_unit_test(
1696 name = "global_average_pooling_spnchw_test",
1697 srcs = [
1698 "test/global-average-pooling-spnchw.cc",
1699 "test/global-average-pooling-spnchw-operator-tester.h",
1700 ],
1701 deps = OPERATOR_TEST_DEPS,
1702)
1703
1704xnnpack_unit_test(
1705 name = "hardswish_test",
1706 srcs = [
1707 "test/hardswish.cc",
1708 "test/hardswish-operator-tester.h",
1709 ],
1710 deps = OPERATOR_TEST_DEPS,
1711)
1712
1713xnnpack_unit_test(
1714 name = "leaky_relu_test",
1715 srcs = [
1716 "test/leaky-relu.cc",
1717 "test/leaky-relu-operator-tester.h",
1718 ],
1719 deps = OPERATOR_TEST_DEPS,
1720)
1721
1722xnnpack_unit_test(
1723 name = "max_pooling_test",
1724 srcs = [
1725 "test/max-pooling.cc",
1726 "test/max-pooling-operator-tester.h",
1727 ] + OPERATOR_TEST_PARAMS_HDRS,
1728 deps = OPERATOR_TEST_DEPS,
1729)
1730
1731xnnpack_unit_test(
1732 name = "prelu_test",
1733 srcs = [
1734 "test/prelu.cc",
1735 "test/prelu-operator-tester.h",
1736 ] + OPERATOR_TEST_PARAMS_HDRS,
1737 deps = OPERATOR_TEST_DEPS,
1738)
1739
1740xnnpack_unit_test(
1741 name = "sigmoid_test",
1742 srcs = [
1743 "test/sigmoid.cc",
1744 "test/sigmoid-operator-tester.h",
1745 ],
1746 deps = OPERATOR_TEST_DEPS,
1747)
1748
1749xnnpack_unit_test(
1750 name = "softargmax_test",
1751 srcs = [
1752 "test/softargmax.cc",
1753 "test/softargmax-operator-tester.h",
1754 ],
1755 deps = OPERATOR_TEST_DEPS,
1756)
1757
1758xnnpack_unit_test(
1759 name = "unpooling_test",
1760 srcs = [
1761 "test/unpooling.cc",
1762 "test/unpooling-operator-tester.h",
1763 ],
1764 deps = OPERATOR_TEST_DEPS,
1765)
1766
1767############################# Build configurations #############################
1768
Marat Dukhanb8642352019-10-30 15:43:02 -07001769# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07001770config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07001771 name = "xnn_enable_assembly_explicit_true",
1772 define_values = {"xnn_enable_assembly": "true"},
1773)
1774
1775# Disables usage of assembly kernels.
1776config_setting(
1777 name = "xnn_enable_assembly_explicit_false",
1778 define_values = {"xnn_enable_assembly": "false"},
1779)
1780
1781# Builds with -c dbg
1782config_setting(
1783 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001784 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07001785 "compilation_mode": "dbg",
1786 },
1787)
1788
1789# Builds with -c opt
1790config_setting(
1791 name = "optimized_build",
1792 values = {
1793 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001794 },
1795)
1796
1797config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07001798 name = "linux_k8",
1799 values = {"cpu": "k8"},
1800)
1801
1802config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07001803 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07001804 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07001805)
1806
1807config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001808 name = "android",
1809 values = {"crosstool_top": "//external:android/crosstool"},
1810)
1811
1812config_setting(
1813 name = "android_armv7",
1814 values = {
1815 "crosstool_top": "//external:android/crosstool",
1816 "cpu": "armeabi-v7a",
1817 },
1818)
1819
1820config_setting(
1821 name = "android_arm64",
1822 values = {
1823 "crosstool_top": "//external:android/crosstool",
1824 "cpu": "arm64-v8a",
1825 },
1826)
1827
1828config_setting(
1829 name = "android_x86",
1830 values = {
1831 "crosstool_top": "//external:android/crosstool",
1832 "cpu": "x86",
1833 },
1834)
1835
1836config_setting(
1837 name = "android_x86_64",
1838 values = {
1839 "crosstool_top": "//external:android/crosstool",
1840 "cpu": "x86_64",
1841 },
1842)
1843
1844config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07001845 name = "macos_x86_64",
1846 values = {
1847 "apple_platform_type": "macos",
1848 "cpu": "darwin",
1849 },
1850)
1851
1852config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001853 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001854 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07001855)
1856
1857config_setting(
1858 name = "emscripten_wasm",
1859 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001860 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001861 "cpu": "wasm",
1862 },
1863)
1864
1865config_setting(
1866 name = "emscripten_wasmsimd",
1867 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001868 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001869 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07001870 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001871 },
1872)
1873
1874config_setting(
1875 name = "emscripten_asmjs",
1876 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001877 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001878 "cpu": "asmjs",
1879 },
1880)