blob: 57d4a016f9dc9c8ac02317a0f8bbf0755d752c31 [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 Dukhaneeaa7bd2019-10-25 17:31:25 -0700537 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700538 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700539 "src/xnnpack/requantization.h",
540]
541
542ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
543 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700544]
545
546MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700547 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700548 "include/xnnpack.h",
549]
550
551MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
552 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700553 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700554 "src/xnnpack/requantization.h",
555 "include/xnnpack.h",
556]
557
558OPERATOR_TEST_PARAMS_HDRS = [
559 "src/xnnpack/params.h",
560 "src/xnnpack/common.h",
561]
562
563WEIGHTS_PACK_HDRS = [
564 "src/xnnpack/pack.h",
565 "src/xnnpack/operator.h",
566 "src/xnnpack/compute.h",
567]
568
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700569LOGGING_COPTS = select({
570 # No logging in optimized mode
571 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
572 # Full logging in debug mode
573 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
574 # Error-only logging in default (fastbuild) mode
575 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
576})
577
578LOGGING_HDRS = [
579 "src/xnnpack/log.h",
580]
581
Marat Dukhan08c4a432019-10-03 09:29:21 -0700582xnnpack_cc_library(
583 name = "scalar_ukernels",
584 srcs = SCALAR_UKERNELS,
585 hdrs = INTERNAL_HDRS,
586 aarch32_copts = ["-marm"],
587 copts = xnnpack_std_copts(),
588 deps = [
589 "@FP16",
590 "@FXdiv",
591 ],
592)
593
594xnnpack_cc_library(
595 name = "psimd_ukernels",
596 srcs = PSIMD_UKERNELS,
597 hdrs = INTERNAL_HDRS,
598 aarch32_copts = [
599 "-marm",
600 "-mfpu=neon",
601 ],
602 copts = xnnpack_std_copts(),
603 optimized_copts = [
604 "-O3",
605 "-ffast-math",
606 ],
607 deps = [
608 "@FP16",
609 "@psimd",
610 ],
611)
612
613xnnpack_cc_library(
614 name = "neon_ukernels",
615 hdrs = INTERNAL_HDRS,
616 aarch32_copts = [
617 "-marm",
618 "-mfpu=neon",
619 ],
620 aarch32_srcs = NEON_UKERNELS,
621 aarch64_srcs = NEON_UKERNELS,
622 copts = xnnpack_std_copts(),
623 deps = ["@FP16"],
624)
625
626xnnpack_cc_library(
627 name = "neonfma_ukernels",
628 hdrs = INTERNAL_HDRS,
629 aarch32_copts = [
630 "-marm",
631 "-mfpu=neon-vfpv4",
632 ],
633 aarch32_srcs = NEONFMA_UKERNELS,
634 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
635 copts = xnnpack_std_copts(),
636 deps = ["@FP16"],
637)
638
639xnnpack_cc_library(
640 name = "neonfp16arith_ukernels",
641 hdrs = INTERNAL_HDRS,
642 aarch64_copts = ["-march=armv8.2-a+fp16"],
643 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
644 copts = xnnpack_std_copts(),
645 deps = ["@FP16"],
646)
647
648xnnpack_cc_library(
649 name = "sse2_ukernels",
650 hdrs = INTERNAL_HDRS,
651 copts = xnnpack_std_copts(),
652 x86_copts = ["-msse2"],
653 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
654 deps = ["@FP16"],
655)
656
657xnnpack_cc_library(
658 name = "avx_ukernels",
659 hdrs = INTERNAL_HDRS,
660 copts = xnnpack_std_copts(),
661 x86_copts = ["-mavx"],
662 x86_srcs = AVX_UKERNELS,
663 deps = ["@FP16"],
664)
665
666xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700667 name = "avx2_ukernels",
668 hdrs = INTERNAL_HDRS,
669 copts = xnnpack_std_copts(),
670 x86_copts = [
671 "-mfma",
672 "-mavx2",
673 ],
674 x86_srcs = AVX2_UKERNELS,
675 deps = ["@FP16"],
676)
677
678xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700679 name = "avx512f_ukernels",
680 hdrs = INTERNAL_HDRS,
681 copts = xnnpack_std_copts(),
682 x86_copts = ["-mavx512f"],
683 x86_srcs = AVX512F_UKERNELS,
684 deps = ["@FP16"],
685)
686
687xnnpack_cc_library(
688 name = "asm_ukernels",
689 hdrs = ["src/xnnpack/assembly.h"],
690 aarch32_srcs = AARCH32_ASM_UKERNELS,
691 aarch64_srcs = AARCH64_ASM_UKERNELS,
692)
693
694xnnpack_aggregate_library(
695 name = "ukernels",
696 aarch32_deps = [
697 ":psimd_ukernels",
698 ":neon_ukernels",
699 ":neonfma_ukernels",
700 ":asm_ukernels",
701 ],
702 aarch64_deps = [
703 ":psimd_ukernels",
704 ":neon_ukernels",
705 ":neonfma_ukernels",
706 ":neonfp16arith_ukernels",
707 ":asm_ukernels",
708 ],
709 generic_deps = [":scalar_ukernels"],
710 wasmsimd_deps = [
711 ":psimd_ukernels",
712 ],
713 x86_deps = [
714 ":psimd_ukernels",
715 ":sse2_ukernels",
716 ":avx_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700717 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700718 ":avx512f_ukernels",
719 ],
720)
721
722xnnpack_cc_library(
723 name = "im2col",
724 srcs = ["src/im2col.c"],
725 hdrs = [
726 "src/xnnpack/common.h",
727 "src/xnnpack/im2col.h",
728 ],
729 copts = xnnpack_std_copts(),
730)
731
732xnnpack_cc_library(
733 name = "indirection",
734 srcs = ["src/indirection.c"],
735 hdrs = INTERNAL_HDRS,
736 copts = xnnpack_std_copts(),
737 deps = [
738 "@FP16",
739 "@FXdiv",
740 "@pthreadpool",
741 ],
742)
743
744xnnpack_cc_library(
745 name = "operator_run",
746 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700747 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
748 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700749 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
750 "-Wno-vla",
751 ],
752 deps = [
753 "@FP16",
754 "@FXdiv",
755 "@clog",
756 "@pthreadpool",
757 ],
758)
759
760cc_library(
761 name = "enable_assembly",
762 defines = select({
763 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
764 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -0700765 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700766 }),
767)
768
Marat Dukhancf056b22019-10-07 10:26:29 -0700769xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700770 name = "operators",
771 srcs = OPERATOR_SRCS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700772 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -0700773 ],
774 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700775 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700776 "-Isrc",
777 "-Iinclude",
778 ] + select({
779 ":debug_build": [],
780 "//conditions:default": xnnpack_min_size_copts(),
781 }),
Marat Dukhancf056b22019-10-07 10:26:29 -0700782 wasm_srcs = ["src/wasm-stubs.c"],
783 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700784 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700785 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700786 "@FP16",
787 "@FXdiv",
788 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700789 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700790 ],
791)
792
793cc_library(
794 name = "XNNPACK",
795 srcs = [
796 "src/init.c",
797 ],
798 copts = xnnpack_std_copts() + LOGGING_COPTS + [
799 "-Isrc",
800 "-Iinclude",
801 ] + select({
802 ":debug_build": [],
803 "//conditions:default": xnnpack_min_size_copts(),
804 }),
805 includes = ["include"],
806 linkstatic = True,
807 textual_hdrs = ["include/xnnpack.h"],
808 visibility = xnnpack_visibility(),
809 deps = [
810 ":enable_assembly",
811 ":ukernels",
812 ":operator_run",
813 ":operators",
814 "@clog",
815 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -0700816 ] + select({
817 ":emscripten": [],
818 "//conditions:default": ["@cpuinfo"],
819 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700820)
821
822cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700823 name = "xnnpack_operators_nhwc_f32",
824 srcs = [
825 "src/init.c",
826 ],
827 copts = xnnpack_std_copts() + LOGGING_COPTS + [
828 "-Isrc",
829 "-Iinclude",
830 ] + select({
831 ":debug_build": [],
832 "//conditions:default": xnnpack_min_size_copts(),
833 }),
834 defines = [
835 "XNN_NO_Q8_OPERATORS",
836 "XNN_NO_U8_OPERATORS",
837 "XNN_NO_X8_OPERATORS",
838 "XNN_NO_SPNCHW_OPERATORS",
839 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700840 includes = ["include"],
841 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700842 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700843 visibility = xnnpack_visibility(),
844 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700845 ":enable_assembly",
846 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700847 ":operator_run",
848 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700849 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700850 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700851 ] + select({
852 ":emscripten": [],
853 "//conditions:default": ["@cpuinfo"],
854 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700855)
856
Marat Dukhancf056b22019-10-07 10:26:29 -0700857xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700858 name = "bench_utils",
859 srcs = ["bench/utils.cc"],
860 hdrs = ["bench/utils.h"],
861 copts = ["-Wno-unused-result"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700862 deps = ["@cpuinfo"],
863)
864
Frank Barchard7e955972019-10-11 10:34:25 -0700865######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -0700866
867xnnpack_benchmark(
868 name = "q8_gemm_bench",
869 srcs = [
870 "bench/gemm.h",
871 "bench/q8-gemm.cc",
872 "src/xnnpack/AlignedAllocator.h",
873 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
874 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
875 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
876)
877
878xnnpack_benchmark(
879 name = "f16_gemm_bench",
880 srcs = [
881 "bench/f16-gemm.cc",
882 "bench/gemm.h",
883 "src/xnnpack/AlignedAllocator.h",
884 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
885 copts = ["-Wno-unused-function"],
886 deps = MICROKERNEL_BENCHMARK_DEPS,
887)
888
889xnnpack_benchmark(
890 name = "f32_igemm_bench",
891 srcs = [
892 "bench/f32-igemm.cc",
893 "bench/conv.h",
894 "src/xnnpack/AlignedAllocator.h",
895 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -0700896 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700897)
898
899xnnpack_benchmark(
900 name = "f32_conv_hwc_bench",
901 srcs = [
902 "bench/f32-conv-hwc.cc",
903 "bench/dconv.h",
904 "src/xnnpack/AlignedAllocator.h",
905 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
906 copts = ["-Wno-unused-function"],
907 deps = MICROKERNEL_BENCHMARK_DEPS,
908)
909
910xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -0700911 name = "f32_conv_hwc2spchw_bench",
912 srcs = [
913 "bench/f32-conv-hwc2spchw.cc",
914 "bench/dconv.h",
915 "src/xnnpack/AlignedAllocator.h",
916 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
917 copts = ["-Wno-unused-function"],
918 deps = MICROKERNEL_BENCHMARK_DEPS,
919)
920
921xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700922 name = "f32_dwconv_bench",
923 srcs = [
924 "bench/f32-dwconv.cc",
925 "bench/dwconv.h",
926 "src/xnnpack/AlignedAllocator.h",
927 ] + MICROKERNEL_BENCHMARK_HDRS,
928 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
929)
930
931xnnpack_benchmark(
932 name = "f32_dwconv_spchw_bench",
933 srcs = [
934 "bench/f32-dwconv-spchw.cc",
935 "bench/dwconv.h",
936 "src/xnnpack/AlignedAllocator.h",
937 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
938 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
939)
940
941xnnpack_benchmark(
942 name = "f32_gemm_bench",
943 srcs = [
944 "bench/f32-gemm.cc",
945 "bench/gemm.h",
946 "src/xnnpack/AlignedAllocator.h",
947 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
948 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -0700949 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700950)
951
952xnnpack_benchmark(
953 name = "f32_rmax_bench",
954 srcs = [
955 "bench/f32-rmax.cc",
956 "src/xnnpack/AlignedAllocator.h",
957 ] + MICROKERNEL_BENCHMARK_HDRS,
958 deps = MICROKERNEL_BENCHMARK_DEPS,
959)
960
961xnnpack_benchmark(
962 name = "f32_spmm_bench",
963 srcs = [
964 "bench/f32-spmm.cc",
965 "bench/gemm.h",
966 "src/xnnpack/AlignedAllocator.h",
967 ] + MICROKERNEL_BENCHMARK_HDRS,
968 copts = ["-Wno-unused-function"],
969 deps = MICROKERNEL_BENCHMARK_DEPS,
970)
971
972xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -0700973 name = "f32_softargmax_bench",
974 srcs = [
975 "bench/f32-softargmax.cc",
976 ] + MICROKERNEL_BENCHMARK_HDRS,
977 copts = ["-Wno-unused-function"],
978 deps = MICROKERNEL_BENCHMARK_DEPS,
979)
980
981xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700982 name = "f32_im2col_gemm_bench",
983 srcs = [
984 "bench/f32-im2col-gemm.cc",
985 "bench/conv.h",
986 "src/xnnpack/AlignedAllocator.h",
987 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
988 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
989)
990
991########################### Benchmarks for operators ###########################
992
993xnnpack_benchmark(
994 name = "add_bench",
995 srcs = ["bench/add.cc"],
996 deps = OPERATOR_BENCHMARK_DEPS,
997)
998
999xnnpack_benchmark(
1000 name = "average_pooling_bench",
1001 srcs = ["bench/average-pooling.cc"],
1002 deps = OPERATOR_BENCHMARK_DEPS,
1003)
1004
1005xnnpack_benchmark(
1006 name = "channel_shuffle_bench",
1007 srcs = ["bench/channel-shuffle.cc"],
1008 deps = OPERATOR_BENCHMARK_DEPS,
1009)
1010
1011xnnpack_benchmark(
1012 name = "convolution_bench",
1013 srcs = ["bench/convolution.cc"],
1014 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1015 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1016)
1017
1018xnnpack_benchmark(
1019 name = "deconvolution_bench",
1020 srcs = ["bench/deconvolution.cc"],
1021 copts = xnnpack_optional_tflite_copts(),
1022 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1023)
1024
1025xnnpack_benchmark(
1026 name = "global_average_pooling_bench",
1027 srcs = ["bench/global-average-pooling.cc"],
1028 deps = OPERATOR_BENCHMARK_DEPS,
1029)
1030
1031xnnpack_benchmark(
1032 name = "max_pooling_bench",
1033 srcs = ["bench/max-pooling.cc"],
1034 deps = OPERATOR_BENCHMARK_DEPS,
1035)
1036
1037xnnpack_benchmark(
1038 name = "sigmoid_bench",
1039 srcs = ["bench/sigmoid.cc"],
1040 deps = OPERATOR_BENCHMARK_DEPS,
1041)
1042
1043xnnpack_benchmark(
1044 name = "softargmax_bench",
1045 srcs = ["bench/softargmax.cc"],
1046 deps = OPERATOR_BENCHMARK_DEPS,
1047)
1048
Marat Dukhanc068bb62019-10-04 13:24:39 -07001049############################# End-to-end benchmarks ############################
1050
1051cc_library(
1052 name = "mobilenet_v1",
1053 srcs = ["models/mobilenet-v1.cc"],
1054 hdrs = ["models/models.h"],
1055 linkstatic = True,
1056 deps = [
1057 ":XNNPACK",
1058 "@pthreadpool",
1059 ],
1060)
1061
1062cc_library(
1063 name = "mobilenet_v2",
1064 srcs = ["models/mobilenet-v2.cc"],
1065 hdrs = ["models/models.h"],
1066 linkstatic = True,
1067 deps = [
1068 ":XNNPACK",
1069 "@pthreadpool",
1070 ],
1071)
1072
1073xnnpack_benchmark(
1074 name = "end2end_bench",
1075 srcs = ["bench/end2end.cc"],
1076 deps = [
1077 ":XNNPACK",
1078 ":mobilenet_v1",
1079 ":mobilenet_v2",
1080 "@pthreadpool",
1081 ],
1082)
1083
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001084#################### Accuracy evaluation for math functions ####################
1085
1086xnnpack_benchmark(
1087 name = "f32_exp_eval",
1088 srcs = [
1089 "eval/f32-exp.cc",
1090 "src/xnnpack/AlignedAllocator.h",
1091 ] + ACCURACY_EVAL_HDRS,
1092 deps = ACCURACY_EVAL_DEPS,
1093)
1094
Marat Dukhan515c9772019-10-17 18:07:57 -07001095xnnpack_benchmark(
1096 name = "f32_expminus_eval",
1097 srcs = [
1098 "eval/f32-expminus.cc",
1099 "src/xnnpack/AlignedAllocator.h",
1100 ] + ACCURACY_EVAL_HDRS,
1101 deps = ACCURACY_EVAL_DEPS,
1102)
1103
Marat Dukhan98ba4412019-10-23 02:14:28 -07001104xnnpack_benchmark(
1105 name = "f32_extexp_eval",
1106 srcs = [
1107 "eval/f32-extexp.cc",
1108 "src/xnnpack/AlignedAllocator.h",
1109 ] + ACCURACY_EVAL_HDRS,
1110 deps = ACCURACY_EVAL_DEPS,
1111)
1112
Marat Dukhan08c4a432019-10-03 09:29:21 -07001113######################### Unit tests for micro-kernels #########################
1114
1115xnnpack_unit_test(
1116 name = "f16_gemm_test",
1117 srcs = [
1118 "test/f16-gemm.cc",
1119 "test/gemm-microkernel-tester.h",
1120 "src/xnnpack/AlignedAllocator.h",
1121 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1122 deps = MICROKERNEL_TEST_DEPS,
1123)
1124
1125xnnpack_unit_test(
1126 name = "f32_argmaxpool_test",
1127 srcs = [
1128 "test/f32-argmaxpool.cc",
1129 "test/argmaxpool-microkernel-tester.h",
1130 "src/xnnpack/AlignedAllocator.h",
1131 ] + MICROKERNEL_TEST_HDRS,
1132 deps = MICROKERNEL_TEST_DEPS,
1133)
1134
1135xnnpack_unit_test(
1136 name = "f32_avgpool_test",
1137 srcs = [
1138 "test/f32-avgpool.cc",
1139 "test/avgpool-microkernel-tester.h",
1140 "src/xnnpack/AlignedAllocator.h",
1141 ] + MICROKERNEL_TEST_HDRS,
1142 deps = MICROKERNEL_TEST_DEPS,
1143)
1144
1145xnnpack_unit_test(
1146 name = "f32_clamp_test",
1147 srcs = [
1148 "test/f32-clamp.cc",
1149 "test/clamp-microkernel-tester.h",
1150 ] + MICROKERNEL_TEST_HDRS,
1151 deps = MICROKERNEL_TEST_DEPS,
1152)
1153
1154xnnpack_unit_test(
1155 name = "f32_igemm_test",
1156 srcs = [
1157 "test/f32-igemm.cc",
1158 "test/gemm-microkernel-tester.h",
1159 "src/xnnpack/AlignedAllocator.h",
1160 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1161 deps = MICROKERNEL_TEST_DEPS,
1162)
1163
1164xnnpack_unit_test(
1165 name = "f32_conv_hwc_test",
1166 srcs = [
1167 "test/f32-conv-hwc.cc",
1168 "test/conv-hwc-microkernel-tester.h",
1169 "src/xnnpack/AlignedAllocator.h",
1170 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1171 deps = MICROKERNEL_TEST_DEPS,
1172)
1173
1174xnnpack_unit_test(
1175 name = "f32_conv_hwc2spchw_test",
1176 srcs = [
1177 "test/f32-conv-hwc2spchw.cc",
1178 "test/conv-hwc2spchw-microkernel-tester.h",
1179 "src/xnnpack/AlignedAllocator.h",
1180 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1181 deps = MICROKERNEL_TEST_DEPS,
1182)
1183
1184xnnpack_unit_test(
1185 name = "f32_dwconv_test",
1186 srcs = [
1187 "test/f32-dwconv.cc",
1188 "test/dwconv-microkernel-tester.h",
1189 "src/xnnpack/AlignedAllocator.h",
1190 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1191 deps = MICROKERNEL_TEST_DEPS,
1192)
1193
1194xnnpack_unit_test(
1195 name = "f32_dwconv_spchw_test",
1196 srcs = [
1197 "test/f32-dwconv-spchw.cc",
1198 "test/dwconv-spchw-microkernel-tester.h",
1199 "src/xnnpack/AlignedAllocator.h",
1200 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1201 deps = MICROKERNEL_TEST_DEPS,
1202)
1203
1204xnnpack_unit_test(
1205 name = "f32_gavgpool_test",
1206 srcs = [
1207 "test/f32-gavgpool.cc",
1208 "test/gavgpool-microkernel-tester.h",
1209 "src/xnnpack/AlignedAllocator.h",
1210 ] + MICROKERNEL_TEST_HDRS,
1211 deps = MICROKERNEL_TEST_DEPS,
1212)
1213
1214xnnpack_unit_test(
1215 name = "f32_gavgpool_spchw_test",
1216 srcs = [
1217 "test/f32-gavgpool-spchw.cc",
1218 "test/gavgpool-spchw-microkernel-tester.h",
1219 "src/xnnpack/AlignedAllocator.h",
1220 ] + MICROKERNEL_TEST_HDRS,
1221 deps = MICROKERNEL_TEST_DEPS,
1222)
1223
1224xnnpack_unit_test(
1225 name = "f32_gemm_test",
1226 srcs = [
1227 "test/f32-gemm.cc",
1228 "test/gemm-microkernel-tester.h",
1229 "src/xnnpack/AlignedAllocator.h",
1230 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1231 deps = MICROKERNEL_TEST_DEPS,
1232)
1233
1234xnnpack_unit_test(
1235 name = "f32_gemminc_test",
1236 srcs = [
1237 "test/f32-gemminc.cc",
1238 "test/gemm-microkernel-tester.h",
1239 "src/xnnpack/AlignedAllocator.h",
1240 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1241 deps = MICROKERNEL_TEST_DEPS,
1242)
1243
1244xnnpack_unit_test(
1245 name = "f32_hswish_test",
1246 srcs = [
1247 "test/f32-hswish.cc",
1248 "test/hswish-microkernel-tester.h",
1249 ] + MICROKERNEL_TEST_HDRS,
1250 deps = MICROKERNEL_TEST_DEPS,
1251)
1252
1253xnnpack_unit_test(
1254 name = "f32_maxpool_test",
1255 srcs = [
1256 "test/f32-maxpool.cc",
1257 "test/maxpool-microkernel-tester.h",
1258 ] + MICROKERNEL_TEST_HDRS,
1259 deps = MICROKERNEL_TEST_DEPS,
1260)
1261
1262xnnpack_unit_test(
1263 name = "f32_pavgpool_test",
1264 srcs = [
1265 "test/f32-pavgpool.cc",
1266 "test/avgpool-microkernel-tester.h",
1267 "src/xnnpack/AlignedAllocator.h",
1268 ] + MICROKERNEL_TEST_HDRS,
1269 deps = MICROKERNEL_TEST_DEPS,
1270)
1271
1272xnnpack_unit_test(
1273 name = "f32_ppmm_test",
1274 srcs = [
1275 "test/f32-ppmm.cc",
1276 "test/gemm-microkernel-tester.h",
1277 "src/xnnpack/AlignedAllocator.h",
1278 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1279 deps = MICROKERNEL_TEST_DEPS,
1280)
1281
1282xnnpack_unit_test(
1283 name = "f32_prelu_test",
1284 srcs = [
1285 "test/f32-prelu.cc",
1286 "test/prelu-microkernel-tester.h",
1287 "src/xnnpack/AlignedAllocator.h",
1288 ] + MICROKERNEL_TEST_HDRS,
1289 deps = MICROKERNEL_TEST_DEPS,
1290)
1291
1292xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001293 name = "f32_raddexpminusmax_test",
1294 srcs = [
1295 "test/f32-raddexpminusmax.cc",
1296 "test/raddexpminusmax-microkernel-tester.h",
1297 ] + MICROKERNEL_TEST_HDRS,
1298 deps = MICROKERNEL_TEST_DEPS,
1299)
1300
1301xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001302 name = "f32_raddextexp_test",
1303 srcs = [
1304 "test/f32-raddextexp.cc",
1305 "test/raddextexp-microkernel-tester.h",
1306 ] + MICROKERNEL_TEST_HDRS,
1307 deps = MICROKERNEL_TEST_DEPS,
1308)
1309
1310xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001311 name = "f32_raddstoreexpminusmax_test",
1312 srcs = [
1313 "test/f32-raddstoreexpminusmax.cc",
1314 "test/raddstoreexpminusmax-microkernel-tester.h",
1315 ] + MICROKERNEL_TEST_HDRS,
1316 deps = MICROKERNEL_TEST_DEPS,
1317)
1318
1319xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001320 name = "f32_rmax_test",
1321 srcs = [
1322 "test/f32-rmax.cc",
1323 "test/rmax-microkernel-tester.h",
1324 ] + MICROKERNEL_TEST_HDRS,
1325 deps = MICROKERNEL_TEST_DEPS,
1326)
1327
1328xnnpack_unit_test(
1329 name = "f32_spmm_test",
1330 srcs = [
1331 "test/f32-spmm.cc",
1332 "test/spmm-microkernel-tester.h",
1333 "src/xnnpack/AlignedAllocator.h",
1334 ] + MICROKERNEL_TEST_HDRS,
1335 deps = MICROKERNEL_TEST_DEPS,
1336)
1337
1338xnnpack_unit_test(
1339 name = "f32_vadd_test",
1340 srcs = [
1341 "test/f32-vadd.cc",
1342 "test/vadd-microkernel-tester.h",
1343 ] + MICROKERNEL_TEST_HDRS,
1344 deps = MICROKERNEL_TEST_DEPS,
1345)
1346
1347xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001348 name = "f32_vmul_test",
1349 srcs = [
1350 "test/f32-vmul.cc",
1351 "test/vmul-microkernel-tester.h",
1352 ] + MICROKERNEL_TEST_HDRS,
1353 deps = MICROKERNEL_TEST_DEPS,
1354)
1355
1356xnnpack_unit_test(
1357 name = "f32_vmulcaddc_test",
1358 srcs = [
1359 "test/f32-vmulcaddc.cc",
1360 "test/vmulcaddc-microkernel-tester.h",
1361 "src/xnnpack/AlignedAllocator.h",
1362 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1363 deps = MICROKERNEL_TEST_DEPS,
1364)
1365
1366xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001367 name = "f32_vscale_test",
1368 srcs = [
1369 "test/f32-vscale.cc",
1370 "test/vscale-microkernel-tester.h",
1371 ] + MICROKERNEL_TEST_HDRS,
1372 deps = MICROKERNEL_TEST_DEPS,
1373)
1374
1375xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001376 name = "f32_vscaleexpminusmax_test",
1377 srcs = [
1378 "test/f32-vscaleexpminusmax.cc",
1379 "test/vscaleexpminusmax-microkernel-tester.h",
1380 ] + MICROKERNEL_TEST_HDRS,
1381 deps = MICROKERNEL_TEST_DEPS,
1382)
1383
1384xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001385 name = "f32_vscaleextexp_test",
1386 srcs = [
1387 "test/f32-vscaleextexp.cc",
1388 "test/vscaleextexp-microkernel-tester.h",
1389 ] + MICROKERNEL_TEST_HDRS,
1390 deps = MICROKERNEL_TEST_DEPS,
1391)
1392
1393xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001394 name = "f32_vsub_test",
1395 srcs = [
1396 "test/f32-vsub.cc",
1397 "test/vsub-microkernel-tester.h",
1398 ] + MICROKERNEL_TEST_HDRS,
1399 deps = MICROKERNEL_TEST_DEPS,
1400)
1401
1402xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001403 name = "q8_avgpool_test",
1404 srcs = [
1405 "test/q8-avgpool.cc",
1406 "test/avgpool-microkernel-tester.h",
1407 "src/xnnpack/AlignedAllocator.h",
1408 ] + MICROKERNEL_TEST_HDRS,
1409 deps = MICROKERNEL_TEST_DEPS,
1410)
1411
1412xnnpack_unit_test(
1413 name = "q8_igemm_test",
1414 srcs = [
1415 "test/q8-igemm.cc",
1416 "test/gemm-microkernel-tester.h",
1417 "src/xnnpack/AlignedAllocator.h",
1418 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1419 deps = MICROKERNEL_TEST_DEPS,
1420)
1421
1422xnnpack_unit_test(
1423 name = "q8_dwconv_test",
1424 srcs = [
1425 "test/q8-dwconv.cc",
1426 "test/dwconv-microkernel-tester.h",
1427 "src/xnnpack/AlignedAllocator.h",
1428 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1429 deps = MICROKERNEL_TEST_DEPS,
1430)
1431
1432xnnpack_unit_test(
1433 name = "q8_gavgpool_test",
1434 srcs = [
1435 "test/q8-gavgpool.cc",
1436 "test/gavgpool-microkernel-tester.h",
1437 "src/xnnpack/AlignedAllocator.h",
1438 ] + MICROKERNEL_TEST_HDRS,
1439 deps = MICROKERNEL_TEST_DEPS,
1440)
1441
1442xnnpack_unit_test(
1443 name = "q8_gemm_test",
1444 srcs = [
1445 "test/q8-gemm.cc",
1446 "test/gemm-microkernel-tester.h",
1447 "src/xnnpack/AlignedAllocator.h",
1448 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1449 deps = MICROKERNEL_TEST_DEPS,
1450)
1451
1452xnnpack_unit_test(
1453 name = "q8_vadd_test",
1454 srcs = [
1455 "test/q8-vadd.cc",
1456 "test/vadd-microkernel-tester.h",
1457 ] + MICROKERNEL_TEST_HDRS,
1458 deps = MICROKERNEL_TEST_DEPS,
1459)
1460
1461xnnpack_unit_test(
1462 name = "u8_clamp_test",
1463 srcs = [
1464 "test/u8-clamp.cc",
1465 "test/clamp-microkernel-tester.h",
1466 ] + MICROKERNEL_TEST_HDRS,
1467 deps = MICROKERNEL_TEST_DEPS,
1468)
1469
1470xnnpack_unit_test(
1471 name = "u8_lut32norm_test",
1472 srcs = [
1473 "test/u8-lut32norm.cc",
1474 "test/lut-norm-microkernel-tester.h",
1475 ] + MICROKERNEL_TEST_HDRS,
1476 deps = MICROKERNEL_TEST_DEPS,
1477)
1478
1479xnnpack_unit_test(
1480 name = "u8_maxpool_test",
1481 srcs = [
1482 "test/u8-maxpool.cc",
1483 "test/maxpool-microkernel-tester.h",
1484 ] + MICROKERNEL_TEST_HDRS,
1485 deps = MICROKERNEL_TEST_DEPS,
1486)
1487
1488xnnpack_unit_test(
1489 name = "u8_rmax_test",
1490 srcs = [
1491 "test/u8-rmax.cc",
1492 "test/rmax-microkernel-tester.h",
1493 ] + MICROKERNEL_TEST_HDRS,
1494 deps = MICROKERNEL_TEST_DEPS,
1495)
1496
1497xnnpack_unit_test(
1498 name = "x32_packx_test",
1499 srcs = [
1500 "test/x32-packx.cc",
1501 "test/pack-microkernel-tester.h",
1502 "src/xnnpack/AlignedAllocator.h",
1503 ] + MICROKERNEL_TEST_HDRS,
1504 deps = MICROKERNEL_TEST_DEPS,
1505)
1506
1507xnnpack_unit_test(
1508 name = "x32_pad_test",
1509 srcs = [
1510 "test/x32-pad.cc",
1511 "test/pad-microkernel-tester.h",
1512 ] + MICROKERNEL_TEST_HDRS,
1513 deps = MICROKERNEL_TEST_DEPS,
1514)
1515
1516xnnpack_unit_test(
1517 name = "x32_unpool_test",
1518 srcs = [
1519 "test/x32-unpool.cc",
1520 "test/unpool-microkernel-tester.h",
1521 ] + MICROKERNEL_TEST_HDRS,
1522 deps = MICROKERNEL_TEST_DEPS,
1523)
1524
1525xnnpack_unit_test(
1526 name = "x32_zip_test",
1527 srcs = [
1528 "test/x32-zip.cc",
1529 "test/zip-microkernel-tester.h",
1530 ] + MICROKERNEL_TEST_HDRS,
1531 deps = MICROKERNEL_TEST_DEPS,
1532)
1533
1534xnnpack_unit_test(
1535 name = "x8_lut_test",
1536 srcs = [
1537 "test/x8-lut.cc",
1538 "test/lut-microkernel-tester.h",
1539 ] + MICROKERNEL_TEST_HDRS,
1540 deps = MICROKERNEL_TEST_DEPS,
1541)
1542
1543xnnpack_unit_test(
1544 name = "x8_zip_test",
1545 srcs = [
1546 "test/x8-zip.cc",
1547 "test/zip-microkernel-tester.h",
1548 ] + MICROKERNEL_TEST_HDRS,
1549 deps = MICROKERNEL_TEST_DEPS,
1550)
1551
1552########################### Size test for the library ##########################
1553
1554xnnpack_binary(
1555 name = "size_test",
1556 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001557 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001558)
1559
1560########################### Unit tests for operators ###########################
1561
1562xnnpack_unit_test(
1563 name = "add_test",
1564 srcs = [
1565 "test/add.cc",
1566 "test/add-operator-tester.h",
1567 ],
1568 deps = OPERATOR_TEST_DEPS,
1569)
1570
1571xnnpack_unit_test(
1572 name = "argmax_pooling_test",
1573 srcs = [
1574 "test/argmax-pooling.cc",
1575 "test/argmax-pooling-operator-tester.h",
1576 ] + OPERATOR_TEST_PARAMS_HDRS,
1577 deps = OPERATOR_TEST_DEPS,
1578)
1579
1580xnnpack_unit_test(
1581 name = "average_pooling_test",
1582 srcs = [
1583 "test/average-pooling.cc",
1584 "test/average-pooling-operator-tester.h",
1585 ] + OPERATOR_TEST_PARAMS_HDRS,
1586 deps = OPERATOR_TEST_DEPS,
1587)
1588
1589xnnpack_unit_test(
1590 name = "channel_pad_test",
1591 srcs = [
1592 "test/channel-pad.cc",
1593 "test/channel-pad-operator-tester.h",
1594 ] + OPERATOR_TEST_PARAMS_HDRS,
1595 deps = OPERATOR_TEST_DEPS,
1596)
1597
1598xnnpack_unit_test(
1599 name = "channel_shuffle_test",
1600 srcs = [
1601 "test/channel-shuffle.cc",
1602 "test/channel-shuffle-operator-tester.h",
1603 ],
1604 deps = OPERATOR_TEST_DEPS,
1605)
1606
1607xnnpack_unit_test(
1608 name = "clamp_test",
1609 srcs = [
1610 "test/clamp.cc",
1611 "test/clamp-operator-tester.h",
1612 ],
1613 deps = OPERATOR_TEST_DEPS,
1614)
1615
1616xnnpack_unit_test(
1617 name = "convolution_test",
1618 srcs = [
1619 "test/convolution.cc",
1620 "test/convolution-operator-tester.h",
1621 ],
1622 deps = OPERATOR_TEST_DEPS,
1623)
1624
1625xnnpack_unit_test(
1626 name = "convolution_spnchw_test",
1627 srcs = [
1628 "test/convolution-spnchw.cc",
1629 "test/convolution-spnchw-operator-tester.h",
1630 ],
1631 deps = OPERATOR_TEST_DEPS,
1632)
1633
1634xnnpack_unit_test(
1635 name = "deconvolution_test",
1636 srcs = [
1637 "test/deconvolution.cc",
1638 "test/deconvolution-operator-tester.h",
1639 ] + OPERATOR_TEST_PARAMS_HDRS,
1640 deps = OPERATOR_TEST_DEPS,
1641)
1642
1643xnnpack_unit_test(
1644 name = "fully_connected_test",
1645 srcs = [
1646 "test/fully-connected.cc",
1647 "test/fully-connected-operator-tester.h",
1648 ],
1649 deps = OPERATOR_TEST_DEPS,
1650)
1651
1652xnnpack_unit_test(
1653 name = "global_average_pooling_test",
1654 srcs = [
1655 "test/global-average-pooling.cc",
1656 "test/global-average-pooling-operator-tester.h",
1657 ] + OPERATOR_TEST_PARAMS_HDRS,
1658 deps = OPERATOR_TEST_DEPS,
1659)
1660
1661xnnpack_unit_test(
1662 name = "global_average_pooling_spnchw_test",
1663 srcs = [
1664 "test/global-average-pooling-spnchw.cc",
1665 "test/global-average-pooling-spnchw-operator-tester.h",
1666 ],
1667 deps = OPERATOR_TEST_DEPS,
1668)
1669
1670xnnpack_unit_test(
1671 name = "hardswish_test",
1672 srcs = [
1673 "test/hardswish.cc",
1674 "test/hardswish-operator-tester.h",
1675 ],
1676 deps = OPERATOR_TEST_DEPS,
1677)
1678
1679xnnpack_unit_test(
1680 name = "leaky_relu_test",
1681 srcs = [
1682 "test/leaky-relu.cc",
1683 "test/leaky-relu-operator-tester.h",
1684 ],
1685 deps = OPERATOR_TEST_DEPS,
1686)
1687
1688xnnpack_unit_test(
1689 name = "max_pooling_test",
1690 srcs = [
1691 "test/max-pooling.cc",
1692 "test/max-pooling-operator-tester.h",
1693 ] + OPERATOR_TEST_PARAMS_HDRS,
1694 deps = OPERATOR_TEST_DEPS,
1695)
1696
1697xnnpack_unit_test(
1698 name = "prelu_test",
1699 srcs = [
1700 "test/prelu.cc",
1701 "test/prelu-operator-tester.h",
1702 ] + OPERATOR_TEST_PARAMS_HDRS,
1703 deps = OPERATOR_TEST_DEPS,
1704)
1705
1706xnnpack_unit_test(
1707 name = "sigmoid_test",
1708 srcs = [
1709 "test/sigmoid.cc",
1710 "test/sigmoid-operator-tester.h",
1711 ],
1712 deps = OPERATOR_TEST_DEPS,
1713)
1714
1715xnnpack_unit_test(
1716 name = "softargmax_test",
1717 srcs = [
1718 "test/softargmax.cc",
1719 "test/softargmax-operator-tester.h",
1720 ],
1721 deps = OPERATOR_TEST_DEPS,
1722)
1723
1724xnnpack_unit_test(
1725 name = "unpooling_test",
1726 srcs = [
1727 "test/unpooling.cc",
1728 "test/unpooling-operator-tester.h",
1729 ],
1730 deps = OPERATOR_TEST_DEPS,
1731)
1732
1733############################# Build configurations #############################
1734
1735config_setting(
1736 name = "linux_k8",
1737 values = {
1738 "cpu": "k8",
1739 },
1740)
1741
1742config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07001743 name = "linux_aarch64",
1744 values = {
1745 "cpu": "aarch64",
1746 },
1747)
1748
1749config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001750 name = "android",
1751 values = {"crosstool_top": "//external:android/crosstool"},
1752)
1753
1754config_setting(
1755 name = "android_armv7",
1756 values = {
1757 "crosstool_top": "//external:android/crosstool",
1758 "cpu": "armeabi-v7a",
1759 },
1760)
1761
1762config_setting(
1763 name = "android_arm64",
1764 values = {
1765 "crosstool_top": "//external:android/crosstool",
1766 "cpu": "arm64-v8a",
1767 },
1768)
1769
1770config_setting(
1771 name = "android_x86",
1772 values = {
1773 "crosstool_top": "//external:android/crosstool",
1774 "cpu": "x86",
1775 },
1776)
1777
1778config_setting(
1779 name = "android_x86_64",
1780 values = {
1781 "crosstool_top": "//external:android/crosstool",
1782 "cpu": "x86_64",
1783 },
1784)
1785
1786config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07001787 name = "macos_x86_64",
1788 values = {
1789 "apple_platform_type": "macos",
1790 "cpu": "darwin",
1791 },
1792)
1793
1794config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001795 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001796 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07001797)
1798
1799config_setting(
1800 name = "emscripten_wasm",
1801 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001802 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001803 "cpu": "wasm",
1804 },
1805)
1806
1807config_setting(
1808 name = "emscripten_wasmsimd",
1809 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001810 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001811 "cpu": "wasm",
1812 "features": "wasmsimd",
1813 },
1814)
1815
1816config_setting(
1817 name = "emscripten_asmjs",
1818 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001819 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001820 "cpu": "asmjs",
1821 },
1822)
1823
Marat Dukhan174706e2019-10-09 16:33:53 -07001824# Builds with -c dbg
Marat Dukhan08c4a432019-10-03 09:29:21 -07001825config_setting(
1826 name = "debug_build",
1827 values = {
1828 "compilation_mode": "dbg",
1829 },
Marat Dukhan08c4a432019-10-03 09:29:21 -07001830)
1831
Marat Dukhan174706e2019-10-09 16:33:53 -07001832# Builds with -c opt
Marat Dukhan08c4a432019-10-03 09:29:21 -07001833config_setting(
1834 name = "optimized_build",
1835 values = {
1836 "compilation_mode": "opt",
1837 },
Marat Dukhan08c4a432019-10-03 09:29:21 -07001838)
1839
1840# Enables usage of assembly kernels.
1841config_setting(
1842 name = "xnn_enable_assembly_explicit_true",
1843 define_values = {"xnn_enable_assembly": "true"},
1844)
1845
1846# Disables usage of assembly kernels.
1847config_setting(
1848 name = "xnn_enable_assembly_explicit_false",
1849 define_values = {"xnn_enable_assembly": "false"},
1850)