blob: a70ddfb3ab42e65f8f1e9f4e81b25a8597588305 [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
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08009load(":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")
10
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan08c4a432019-10-03 09:29:21 -070015OPERATOR_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",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800109 "src/f32-prelu/scalar-2x1.c",
110 "src/f32-prelu/scalar-2x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700111 "src/f32-rmax/scalar.c",
112 "src/f32-spmm/1x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700113 "src/f32-spmm/1x1-scalar.c",
114 "src/f32-spmm/2x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700115 "src/f32-spmm/2x1-scalar.c",
116 "src/f32-spmm/4x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700117 "src/f32-spmm/4x1-scalar.c",
118 "src/f32-spmm/8x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700119 "src/f32-spmm/8x1-scalar.c",
Erich Elsenc6afd9b2019-10-24 16:10:53 -0700120 "src/f32-spmm/8x2-scalar.c",
121 "src/f32-spmm/8x4-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700122 "src/f32-vadd/scalar.c",
123 "src/f32-vmul/scalar.c",
124 "src/f32-vmulcaddc/c1-scalar-x2.c",
125 "src/f32-vsub/scalar.c",
126 "src/q8-avgpool/mp9p8q-scalar.c",
127 "src/q8-avgpool/up9-scalar.c",
128 "src/q8-igemm/2x2-scalar.c",
129 "src/q8-dwconv/up1x9-scalar.c",
130 "src/q8-gavgpool/mp7p7q-scalar.c",
131 "src/q8-gavgpool/up7-scalar.c",
132 "src/q8-gemm/2x2-scalar.c",
133 "src/q8-vadd/scalar.c",
134 "src/u8-clamp/scalar.c",
135 "src/u8-lut32norm/scalar.c",
136 "src/u8-maxpool/9p8q-scalar.c",
137 "src/u8-rmax/scalar.c",
138 "src/x32-packx/x2-scalar.c",
139 "src/x32-packx/x3-scalar.c",
140 "src/x32-packx/x4-scalar.c",
141 "src/x32-pad/x2-scalar.c",
142 "src/x32-unpool/scalar.c",
143 "src/x32-zip/x2-scalar.c",
144 "src/x32-zip/x3-scalar.c",
145 "src/x32-zip/x4-scalar.c",
146 "src/x32-zip/xm-scalar.c",
147 "src/x8-lut/scalar.c",
148 "src/x8-zip/x2-scalar.c",
149 "src/x8-zip/x3-scalar.c",
150 "src/x8-zip/x4-scalar.c",
151 "src/x8-zip/xm-scalar.c",
152]
153
154PSIMD_UKERNELS = [
155 "src/f32-argmaxpool/mp9p8q-psimd.c",
156 "src/f32-argmaxpool/up4-psimd.c",
157 "src/f32-argmaxpool/up9-psimd.c",
158 "src/f32-avgpool/mp9p8q-psimd.c",
159 "src/f32-avgpool/up9-psimd.c",
160 "src/f32-clamp/psimd.c",
161 "src/f32-igemm/1x8-psimd-loadsplat.c",
162 "src/f32-igemm/1x8-psimd-splat.c",
163 "src/f32-igemm/1x8s4-psimd.c",
164 "src/f32-igemm/4x2c4-psimd.c",
165 "src/f32-igemm/4x8-psimd-loadsplat.c",
166 "src/f32-igemm/4x8-psimd-splat.c",
167 "src/f32-igemm/4x8s4-psimd.c",
168 "src/f32-igemm/6x8-psimd-loadsplat.c",
169 "src/f32-igemm/6x8-psimd-splat.c",
170 "src/f32-igemm/6x8s4-psimd.c",
171 "src/f32-dwconv/up4x25-psimd.c",
172 "src/f32-dwconv/up4x4-psimd.c",
173 "src/f32-dwconv/up4x9-psimd.c",
174 "src/f32-gavgpool/mp7p7q-psimd.c",
175 "src/f32-gavgpool/up7-psimd.c",
176 "src/f32-gemm/1x8-psimd-loadsplat.c",
177 "src/f32-gemm/1x8-psimd-splat.c",
178 "src/f32-gemm/1x8s4-psimd.c",
179 "src/f32-gemm/4x8-psimd-loadsplat.c",
180 "src/f32-gemm/4x8-psimd-splat.c",
181 "src/f32-gemm/4x8s4-psimd.c",
182 "src/f32-gemm/6x8-psimd-loadsplat.c",
183 "src/f32-gemm/6x8-psimd-splat.c",
184 "src/f32-gemm/6x8s4-psimd.c",
185 "src/f32-gemminc/1x8-psimd-loadsplat.c",
186 "src/f32-gemminc/1x8-psimd-splat.c",
187 "src/f32-gemminc/1x8s4-psimd.c",
188 "src/f32-gemminc/4x8-psimd-loadsplat.c",
189 "src/f32-gemminc/4x8-psimd-splat.c",
190 "src/f32-gemminc/4x8s4-psimd.c",
191 "src/f32-gemminc/6x8-psimd-loadsplat.c",
192 "src/f32-gemminc/6x8-psimd-splat.c",
193 "src/f32-gemminc/6x8s4-psimd.c",
194 "src/f32-hswish/psimd.c",
195 "src/f32-maxpool/9p8q-psimd.c",
196 "src/f32-pavgpool/mp9p8q-psimd.c",
197 "src/f32-pavgpool/up9-psimd.c",
198 "src/f32-ppmm/4x8-psimd.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800199 "src/f32-prelu/psimd-2x4.c",
200 "src/f32-prelu/psimd-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700201 "src/f32-vadd/psimd.c",
202 "src/f32-vmul/psimd.c",
203 "src/f32-vmulcaddc/c4-psimd-x2.c",
204 "src/f32-vsub/psimd.c",
205 "src/x32-packx/x4-psimd.c",
206 "src/x32-pad/x2-psimd.c",
207 "src/x32-unpool/psimd.c",
208 "src/x32-zip/x2-psimd.c",
209 "src/x32-zip/x3-psimd.c",
210 "src/x32-zip/x4-psimd.c",
211 "src/x32-zip/xm-psimd.c",
212]
213
214# ISA-specific micro-kernels
215NEON_UKERNELS = [
216 "src/f32-avgpool/mp9p8q-neon.c",
217 "src/f32-avgpool/up9-neon.c",
218 "src/f32-clamp/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700219 "src/f32-dwconv/up4x9-neon.c",
220 "src/f32-gavgpool-spchw/neon-x4.c",
221 "src/f32-gavgpool/mp7p7q-neon.c",
222 "src/f32-gavgpool/up7-neon.c",
223 "src/f32-gemm/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700224 "src/f32-gemm/4x2-neon-ld64.c",
225 "src/f32-gemm/4x8-neon-ld128.c",
226 "src/f32-gemm/4x8-neon-ld64.c",
227 "src/f32-gemm/5x8-neon-ld64.c",
228 "src/f32-gemm/6x8-neon-ld64.c",
229 "src/f32-gemminc/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700230 "src/f32-gemminc/4x8-neon-ld128.c",
231 "src/f32-gemminc/4x8-neon-ld64.c",
232 "src/f32-gemminc/5x8-neon-ld64.c",
233 "src/f32-gemminc/6x8-neon-ld64.c",
234 "src/f32-hswish/neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800235 "src/f32-igemm/1x8-neon-ld64.c",
236 "src/f32-igemm/4x2-neon-ld64.c",
237 "src/f32-igemm/4x4-neon-ld64.c",
238 "src/f32-igemm/4x8-neon-ld128.c",
239 "src/f32-igemm/4x8-neon-ld64.c",
240 "src/f32-igemm/6x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700241 "src/f32-pavgpool/mp9p8q-neon.c",
242 "src/f32-pavgpool/up9-neon.c",
243 "src/f32-ppmm/4x8-neon.c",
244 "src/f32-ppmm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800245 "src/f32-prelu/neon-2x4.c",
246 "src/f32-prelu/neon-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700247 "src/f32-rmax/neon.c",
248 "src/f32-vmulcaddc/c4-neon-x2.c",
249 "src/q8-avgpool/mp9p8q-neon.c",
250 "src/q8-avgpool/up9-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700251 "src/q8-dwconv/up8x9-neon.c",
252 "src/q8-gavgpool/mp7p7q-neon.c",
253 "src/q8-gavgpool/up7-neon.c",
254 "src/q8-gemm/4x8-neon.c",
255 "src/q8-gemm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800256 "src/q8-igemm/4x8-neon.c",
257 "src/q8-igemm/8x8-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700258 "src/q8-vadd/neon.c",
259 "src/u8-clamp/neon.c",
260 "src/u8-maxpool/9p8q-neon.c",
261 "src/u8-rmax/neon.c",
262 "src/x32-packx/x4-neon-st4.c",
263 "src/x32-pad/x2-neon.c",
264 "src/x32-zip/x2-neon.c",
265 "src/x32-zip/x3-neon.c",
266 "src/x32-zip/x4-neon.c",
267 "src/x32-zip/xm-neon.c",
268 "src/x8-zip/x2-neon.c",
269 "src/x8-zip/x3-neon.c",
270 "src/x8-zip/x4-neon.c",
271 "src/x8-zip/xm-neon.c",
272]
273
274NEONFMA_UKERNELS = [
Frank Barcharddb45b6a2019-10-09 16:42:45 -0700275 "src/f32-igemm/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700276 "src/f32-igemm/4x2-neonfma-ld64.c",
277 "src/f32-igemm/4x4-neonfma-ld64.c",
278 "src/f32-igemm/4x8-neonfma-ld128.c",
279 "src/f32-igemm/4x8-neonfma-ld64.c",
280 "src/f32-igemm/6x8-neonfma-ld64.c",
281 "src/f32-dwconv/up4x9-neonfma.c",
282 "src/f32-dwconv/up8x9-neonfma.c",
283 "src/f32-gemm/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700284 "src/f32-gemm/4x2-neonfma-ld64.c",
285 "src/f32-gemm/4x8-neonfma-ld128.c",
286 "src/f32-gemm/4x8-neonfma-ld64.c",
287 "src/f32-gemm/5x8-neonfma-ld64.c",
288 "src/f32-gemm/6x8-neonfma-ld64.c",
289 "src/f32-gemminc/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700290 "src/f32-gemminc/4x8-neonfma-ld128.c",
291 "src/f32-gemminc/4x8-neonfma-ld64.c",
292 "src/f32-gemminc/5x8-neonfma-ld64.c",
293 "src/f32-gemminc/6x8-neonfma-ld64.c",
294 "src/f32-hswish/neonfma.c",
295 "src/f32-ppmm/4x8-neonfma.c",
296 "src/f32-ppmm/8x8-neonfma.c",
297 "src/f32-vmulcaddc/c4-neonfma-x2.c",
298]
299
300AARCH64_NEONFMA_UKERNELS = [
301 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
302 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
303 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
304 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
305 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
306 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
307 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
308 "src/f32-spmm/12x1-neonfma.c",
309 "src/f32-spmm/12x2-neonfma.c",
310 "src/f32-spmm/12x4-neonfma.c",
311 "src/f32-spmm/16x1-neonfma-pipelined.c",
312 "src/f32-spmm/16x1-neonfma-unroll2.c",
313 "src/f32-spmm/16x1-neonfma.c",
314 "src/f32-spmm/16x2-neonfma.c",
315 "src/f32-spmm/16x4-neonfma.c",
316 "src/f32-spmm/4x1-neonfma-pipelined.c",
317 "src/f32-spmm/4x1-neonfma-unroll2.c",
318 "src/f32-spmm/4x1-neonfma.c",
319 "src/f32-spmm/4x2-neonfma.c",
320 "src/f32-spmm/4x4-neonfma.c",
321 "src/f32-spmm/8x1-neonfma-pipelined.c",
322 "src/f32-spmm/8x1-neonfma-unroll2.c",
323 "src/f32-spmm/8x1-neonfma.c",
324 "src/f32-spmm/8x2-neonfma.c",
325 "src/f32-spmm/8x4-neonfma.c",
326]
327
328AARCH64_NEONFP16ARITH_UKERNELS = [
329 "src/f16-gemm/4x8-neonfp16arith-ld64.c",
330 "src/f16-gemm/6x8-neonfp16arith-ld64.c",
331 "src/f16-gemm/8x8-neonfp16arith-ld64.c",
332]
333
334SSE_UKERNELS = [
335 "src/f32-avgpool/mp9p8q-sse.c",
336 "src/f32-avgpool/up9-sse.c",
337 "src/f32-clamp/sse.c",
338 "src/f32-igemm/1x8-sse-dup.c",
339 "src/f32-igemm/1x8-sse-load1.c",
340 "src/f32-igemm/1x8s4-sse.c",
341 "src/f32-igemm/4x2c4-sse.c",
342 "src/f32-igemm/4x8-sse-dup.c",
343 "src/f32-igemm/4x8-sse-load1.c",
344 "src/f32-igemm/4x8s4-sse.c",
345 "src/f32-dwconv/up4x25-sse.c",
346 "src/f32-dwconv/up4x4-sse.c",
347 "src/f32-dwconv/up4x9-sse.c",
348 "src/f32-gavgpool-spchw/sse-x4.c",
349 "src/f32-gavgpool/mp7p7q-sse.c",
350 "src/f32-gavgpool/up7-sse.c",
351 "src/f32-gemm/1x8-sse-dup.c",
352 "src/f32-gemm/1x8-sse-load1.c",
353 "src/f32-gemm/1x8s4-sse.c",
354 "src/f32-gemm/4x8-sse-dup.c",
355 "src/f32-gemm/4x8-sse-load1.c",
356 "src/f32-gemm/4x8s4-sse.c",
357 "src/f32-gemminc/1x8-sse-dup.c",
358 "src/f32-gemminc/1x8-sse-load1.c",
359 "src/f32-gemminc/1x8s4-sse.c",
360 "src/f32-gemminc/4x8-sse-dup.c",
361 "src/f32-gemminc/4x8-sse-load1.c",
362 "src/f32-gemminc/4x8s4-sse.c",
363 "src/f32-hswish/sse.c",
364 "src/f32-maxpool/9p8q-sse.c",
365 "src/f32-pavgpool/mp9p8q-sse.c",
366 "src/f32-pavgpool/up9-sse.c",
367 "src/f32-dwconv-spchw/3x3p1-sse.c",
368 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
369 "src/f32-ppmm/4x8-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700370 "src/f32-rmax/sse.c",
371 "src/f32-spmm/4x1-sse.c",
372 "src/f32-spmm/8x1-sse.c",
373 "src/f32-vadd/sse.c",
374 "src/f32-vmul/sse.c",
375 "src/f32-vmulcaddc/c4-sse-x2.c",
376 "src/f32-vsub/sse.c",
377 "src/x32-packx/x4-sse.c",
378]
379
380SSE2_UKERNELS = [
381 "src/f32-argmaxpool/mp9p8q-sse2.c",
382 "src/f32-argmaxpool/up4-sse2.c",
383 "src/f32-argmaxpool/up9-sse2.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800384 "src/f32-prelu/sse2-2x4.c",
385 "src/f32-prelu/sse2-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700386 "src/q8-avgpool/mp9p8q-sse2.c",
387 "src/q8-avgpool/up9-sse2.c",
388 "src/q8-igemm/4x4c2-sse2.c",
389 "src/q8-dwconv/up8x9-sse2.c",
390 "src/q8-gavgpool/mp7p7q-sse2.c",
391 "src/q8-gavgpool/up7-sse2.c",
392 "src/q8-gemm/2x4c8-sse2.c",
393 "src/q8-gemm/4x4c2-sse2.c",
394 "src/q8-vadd/sse2.c",
395 "src/u8-clamp/sse2.c",
396 "src/u8-maxpool/9p8q-sse2.c",
397 "src/u8-rmax/sse2.c",
398 "src/x32-pad/x2-sse2.c",
399 "src/x32-zip/x2-sse2.c",
400 "src/x32-zip/x3-sse2.c",
401 "src/x32-zip/x4-sse2.c",
402 "src/x32-zip/xm-sse2.c",
403 "src/x8-zip/x2-sse2.c",
404 "src/x8-zip/x3-sse2.c",
405 "src/x8-zip/x4-sse2.c",
406 "src/x8-zip/xm-sse2.c",
407]
408
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800409SSE41_UKERNELS = [
410 "src/f32-prelu/sse41-2x4.c",
411 "src/f32-prelu/sse41-2x8.c",
412]
413
Marat Dukhan08c4a432019-10-03 09:29:21 -0700414AVX_UKERNELS = [
415 "src/f32-rmax/avx.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700416 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700417]
418
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700419AVX2_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700420 "src/f32-raddexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700421 "src/f32-raddextexp/avx2-p5-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700422 "src/f32-raddstoreexpminusmax/avx2-p5-unroll64.c",
423 "src/f32-vscaleexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700424 "src/f32-vscaleextexp/avx2-p5-unroll64.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700425 "src/math/exp-avx2-p5.c",
426 "src/math/exp-avx2-perm-p3.c",
427 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -0700428 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700429 "src/math/extexp-avx2-p5.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700430]
431
Marat Dukhan08c4a432019-10-03 09:29:21 -0700432AVX512F_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700433 "src/f32-raddexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700434 "src/f32-raddextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700435 "src/f32-raddstoreexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700436 "src/f32-rmax/avx512f.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700437 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700438 "src/f32-vscaleexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700439 "src/f32-vscaleextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700440 "src/math/exp-avx512f-p5-scalef.c",
441 "src/math/exp-avx512f-p5.c",
442 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -0700443 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700444 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700445]
446
447AARCH32_ASM_UKERNELS = [
448 "src/q8-dwconv/up8x9-aarch32-neon.S",
449]
450
451AARCH64_ASM_UKERNELS = [
452 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
453 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
454 "src/f32-gemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700455 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700456 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a57.S",
457 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a75.S",
458 "src/f32-gemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700459 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700460 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a57.S",
461 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a75.S",
462 "src/f32-gemm/4x8-aarch64-neonfma-ld128.S",
463 "src/f32-gemm/4x8-aarch64-neonfma-ld64.S",
464 "src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700465 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700466 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a57.S",
467 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S",
468 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S",
469 "src/f32-gemm/6x8-aarch64-neonfma-ld128.S",
470 "src/f32-gemm/6x8-aarch64-neonfma-ld64.S",
471 "src/f32-gemminc/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700472 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700473 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a57.S",
474 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a75.S",
475 "src/f32-gemminc/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700476 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700477 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a57.S",
478 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a75.S",
479 "src/f32-gemminc/4x8-aarch64-neonfma-ld128.S",
480 "src/f32-gemminc/4x8-aarch64-neonfma-ld64.S",
481 "src/f32-gemminc/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700482 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700483 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a57.S",
484 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a73.S",
485 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a75.S",
486 "src/f32-gemminc/6x8-aarch64-neonfma-ld128.S",
487 "src/f32-gemminc/6x8-aarch64-neonfma-ld64.S",
488 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700489 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700490 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S",
491 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S",
492 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700493 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700494 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
495 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700496 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700497 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
498 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
499 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
500]
501
502INTERNAL_MICROKERNEL_HDRS = [
503 "src/xnnpack/argmaxpool.h",
504 "src/xnnpack/avgpool.h",
505 "src/xnnpack/clamp.h",
506 "src/xnnpack/common.h",
507 "src/xnnpack/conv.h",
508 "src/xnnpack/dwconv.h",
509 "src/xnnpack/gavgpool.h",
510 "src/xnnpack/gemm.h",
511 "src/xnnpack/hswish.h",
512 "src/xnnpack/igemm.h",
513 "src/xnnpack/lut.h",
514 "src/xnnpack/math.h",
515 "src/xnnpack/maxpool.h",
516 "src/xnnpack/packx.h",
517 "src/xnnpack/pad.h",
518 "src/xnnpack/params.h",
519 "src/xnnpack/pavgpool.h",
520 "src/xnnpack/ppmm.h",
521 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700522 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700523 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700524 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700525 "src/xnnpack/rmax.h",
526 "src/xnnpack/scalar-utils.h",
527 "src/xnnpack/spmm.h",
528 "src/xnnpack/unpool.h",
529 "src/xnnpack/vadd.h",
530 "src/xnnpack/vmul.h",
531 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700532 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700533 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700534 "src/xnnpack/vscaleextexp.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700535 "src/xnnpack/vsub.h",
536 "src/xnnpack/zip.h",
537]
538
539INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
540 "include/xnnpack.h",
541 "src/xnnpack/allocator.h",
542 "src/xnnpack/compute.h",
543 "src/xnnpack/im2col.h",
544 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700545 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700546 "src/xnnpack/operator.h",
547 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700548 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700549 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700550 "src/xnnpack/requantization.h",
551]
552
553ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
554 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700555]
556
557MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700558 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700559 "include/xnnpack.h",
560]
561
562MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
563 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700564 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700565 "src/xnnpack/requantization.h",
566 "include/xnnpack.h",
567]
568
569OPERATOR_TEST_PARAMS_HDRS = [
570 "src/xnnpack/params.h",
571 "src/xnnpack/common.h",
572]
573
574WEIGHTS_PACK_HDRS = [
575 "src/xnnpack/pack.h",
576 "src/xnnpack/operator.h",
577 "src/xnnpack/compute.h",
578]
579
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700580LOGGING_COPTS = select({
581 # No logging in optimized mode
582 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
583 # Full logging in debug mode
584 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
585 # Error-only logging in default (fastbuild) mode
586 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
587})
588
589LOGGING_HDRS = [
590 "src/xnnpack/log.h",
591]
592
Marat Dukhan08c4a432019-10-03 09:29:21 -0700593xnnpack_cc_library(
594 name = "scalar_ukernels",
595 srcs = SCALAR_UKERNELS,
596 hdrs = INTERNAL_HDRS,
597 aarch32_copts = ["-marm"],
598 copts = xnnpack_std_copts(),
599 deps = [
600 "@FP16",
601 "@FXdiv",
602 ],
603)
604
605xnnpack_cc_library(
606 name = "psimd_ukernels",
607 srcs = PSIMD_UKERNELS,
608 hdrs = INTERNAL_HDRS,
609 aarch32_copts = [
610 "-marm",
611 "-mfpu=neon",
612 ],
613 copts = xnnpack_std_copts(),
614 optimized_copts = [
615 "-O3",
616 "-ffast-math",
617 ],
618 deps = [
619 "@FP16",
620 "@psimd",
621 ],
622)
623
624xnnpack_cc_library(
625 name = "neon_ukernels",
626 hdrs = INTERNAL_HDRS,
627 aarch32_copts = [
628 "-marm",
629 "-mfpu=neon",
630 ],
631 aarch32_srcs = NEON_UKERNELS,
632 aarch64_srcs = NEON_UKERNELS,
633 copts = xnnpack_std_copts(),
634 deps = ["@FP16"],
635)
636
637xnnpack_cc_library(
638 name = "neonfma_ukernels",
639 hdrs = INTERNAL_HDRS,
640 aarch32_copts = [
641 "-marm",
642 "-mfpu=neon-vfpv4",
643 ],
644 aarch32_srcs = NEONFMA_UKERNELS,
645 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
646 copts = xnnpack_std_copts(),
647 deps = ["@FP16"],
648)
649
650xnnpack_cc_library(
651 name = "neonfp16arith_ukernels",
652 hdrs = INTERNAL_HDRS,
653 aarch64_copts = ["-march=armv8.2-a+fp16"],
654 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
655 copts = xnnpack_std_copts(),
656 deps = ["@FP16"],
657)
658
659xnnpack_cc_library(
660 name = "sse2_ukernels",
661 hdrs = INTERNAL_HDRS,
662 copts = xnnpack_std_copts(),
663 x86_copts = ["-msse2"],
664 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
665 deps = ["@FP16"],
666)
667
668xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800669 name = "sse41_ukernels",
670 hdrs = INTERNAL_HDRS,
671 copts = xnnpack_std_copts(),
672 x86_copts = ["-msse4.1"],
673 x86_srcs = SSE41_UKERNELS,
674 deps = ["@FP16"],
675)
676
677xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700678 name = "avx_ukernels",
679 hdrs = INTERNAL_HDRS,
680 copts = xnnpack_std_copts(),
681 x86_copts = ["-mavx"],
682 x86_srcs = AVX_UKERNELS,
683 deps = ["@FP16"],
684)
685
686xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700687 name = "avx2_ukernels",
688 hdrs = INTERNAL_HDRS,
689 copts = xnnpack_std_copts(),
690 x86_copts = [
691 "-mfma",
692 "-mavx2",
693 ],
694 x86_srcs = AVX2_UKERNELS,
695 deps = ["@FP16"],
696)
697
698xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700699 name = "avx512f_ukernels",
700 hdrs = INTERNAL_HDRS,
701 copts = xnnpack_std_copts(),
702 x86_copts = ["-mavx512f"],
703 x86_srcs = AVX512F_UKERNELS,
704 deps = ["@FP16"],
705)
706
707xnnpack_cc_library(
708 name = "asm_ukernels",
709 hdrs = ["src/xnnpack/assembly.h"],
710 aarch32_srcs = AARCH32_ASM_UKERNELS,
711 aarch64_srcs = AARCH64_ASM_UKERNELS,
712)
713
714xnnpack_aggregate_library(
715 name = "ukernels",
716 aarch32_deps = [
717 ":psimd_ukernels",
718 ":neon_ukernels",
719 ":neonfma_ukernels",
720 ":asm_ukernels",
721 ],
722 aarch64_deps = [
723 ":psimd_ukernels",
724 ":neon_ukernels",
725 ":neonfma_ukernels",
726 ":neonfp16arith_ukernels",
727 ":asm_ukernels",
728 ],
729 generic_deps = [":scalar_ukernels"],
730 wasmsimd_deps = [
731 ":psimd_ukernels",
732 ],
733 x86_deps = [
734 ":psimd_ukernels",
735 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800736 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700737 ":avx_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700738 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700739 ":avx512f_ukernels",
740 ],
741)
742
743xnnpack_cc_library(
744 name = "im2col",
745 srcs = ["src/im2col.c"],
746 hdrs = [
747 "src/xnnpack/common.h",
748 "src/xnnpack/im2col.h",
749 ],
750 copts = xnnpack_std_copts(),
751)
752
753xnnpack_cc_library(
754 name = "indirection",
755 srcs = ["src/indirection.c"],
756 hdrs = INTERNAL_HDRS,
757 copts = xnnpack_std_copts(),
758 deps = [
759 "@FP16",
760 "@FXdiv",
761 "@pthreadpool",
762 ],
763)
764
765xnnpack_cc_library(
766 name = "operator_run",
767 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700768 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
769 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700770 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
771 "-Wno-vla",
772 ],
773 deps = [
774 "@FP16",
775 "@FXdiv",
776 "@clog",
777 "@pthreadpool",
778 ],
779)
780
781cc_library(
782 name = "enable_assembly",
783 defines = select({
784 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
785 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -0700786 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700787 }),
788)
789
Marat Dukhancf056b22019-10-07 10:26:29 -0700790xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700791 name = "operators",
792 srcs = OPERATOR_SRCS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700793 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -0700794 ],
795 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700796 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700797 "-Isrc",
798 "-Iinclude",
799 ] + select({
800 ":debug_build": [],
801 "//conditions:default": xnnpack_min_size_copts(),
802 }),
Marat Dukhancf056b22019-10-07 10:26:29 -0700803 wasm_srcs = ["src/wasm-stubs.c"],
804 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700805 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700806 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700807 "@FP16",
808 "@FXdiv",
809 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700810 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700811 ],
812)
813
814cc_library(
815 name = "XNNPACK",
816 srcs = [
817 "src/init.c",
818 ],
819 copts = xnnpack_std_copts() + LOGGING_COPTS + [
820 "-Isrc",
821 "-Iinclude",
822 ] + select({
823 ":debug_build": [],
824 "//conditions:default": xnnpack_min_size_copts(),
825 }),
826 includes = ["include"],
827 linkstatic = True,
828 textual_hdrs = ["include/xnnpack.h"],
829 visibility = xnnpack_visibility(),
830 deps = [
831 ":enable_assembly",
832 ":ukernels",
833 ":operator_run",
834 ":operators",
835 "@clog",
836 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -0700837 ] + select({
838 ":emscripten": [],
839 "//conditions:default": ["@cpuinfo"],
840 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700841)
842
843cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700844 name = "xnnpack_operators_nhwc_f32",
845 srcs = [
846 "src/init.c",
847 ],
848 copts = xnnpack_std_copts() + LOGGING_COPTS + [
849 "-Isrc",
850 "-Iinclude",
851 ] + select({
852 ":debug_build": [],
853 "//conditions:default": xnnpack_min_size_copts(),
854 }),
855 defines = [
856 "XNN_NO_Q8_OPERATORS",
857 "XNN_NO_U8_OPERATORS",
858 "XNN_NO_X8_OPERATORS",
859 "XNN_NO_SPNCHW_OPERATORS",
860 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700861 includes = ["include"],
862 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700863 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700864 visibility = xnnpack_visibility(),
865 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700866 ":enable_assembly",
867 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700868 ":operator_run",
869 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700870 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700871 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700872 ] + select({
873 ":emscripten": [],
874 "//conditions:default": ["@cpuinfo"],
875 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700876)
877
Marat Dukhancf056b22019-10-07 10:26:29 -0700878xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700879 name = "bench_utils",
880 srcs = ["bench/utils.cc"],
881 hdrs = ["bench/utils.h"],
882 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -0800883 deps = [
884 "@com_google_benchmark//:benchmark",
885 "@cpuinfo",
886 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700887)
888
Frank Barchard7e955972019-10-11 10:34:25 -0700889######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -0700890
891xnnpack_benchmark(
892 name = "q8_gemm_bench",
893 srcs = [
894 "bench/gemm.h",
895 "bench/q8-gemm.cc",
896 "src/xnnpack/AlignedAllocator.h",
897 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
898 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
899 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
900)
901
902xnnpack_benchmark(
903 name = "f16_gemm_bench",
904 srcs = [
905 "bench/f16-gemm.cc",
906 "bench/gemm.h",
907 "src/xnnpack/AlignedAllocator.h",
908 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
909 copts = ["-Wno-unused-function"],
910 deps = MICROKERNEL_BENCHMARK_DEPS,
911)
912
913xnnpack_benchmark(
914 name = "f32_igemm_bench",
915 srcs = [
916 "bench/f32-igemm.cc",
917 "bench/conv.h",
918 "src/xnnpack/AlignedAllocator.h",
919 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -0700920 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700921)
922
923xnnpack_benchmark(
924 name = "f32_conv_hwc_bench",
925 srcs = [
926 "bench/f32-conv-hwc.cc",
927 "bench/dconv.h",
928 "src/xnnpack/AlignedAllocator.h",
929 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
930 copts = ["-Wno-unused-function"],
931 deps = MICROKERNEL_BENCHMARK_DEPS,
932)
933
934xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -0700935 name = "f32_conv_hwc2spchw_bench",
936 srcs = [
937 "bench/f32-conv-hwc2spchw.cc",
938 "bench/dconv.h",
939 "src/xnnpack/AlignedAllocator.h",
940 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
941 copts = ["-Wno-unused-function"],
942 deps = MICROKERNEL_BENCHMARK_DEPS,
943)
944
945xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700946 name = "f32_dwconv_bench",
947 srcs = [
948 "bench/f32-dwconv.cc",
949 "bench/dwconv.h",
950 "src/xnnpack/AlignedAllocator.h",
951 ] + MICROKERNEL_BENCHMARK_HDRS,
952 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
953)
954
955xnnpack_benchmark(
956 name = "f32_dwconv_spchw_bench",
957 srcs = [
958 "bench/f32-dwconv-spchw.cc",
959 "bench/dwconv.h",
960 "src/xnnpack/AlignedAllocator.h",
961 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
962 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
963)
964
965xnnpack_benchmark(
966 name = "f32_gemm_bench",
967 srcs = [
968 "bench/f32-gemm.cc",
969 "bench/gemm.h",
970 "src/xnnpack/AlignedAllocator.h",
971 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
972 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -0700973 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700974)
975
976xnnpack_benchmark(
977 name = "f32_rmax_bench",
978 srcs = [
979 "bench/f32-rmax.cc",
980 "src/xnnpack/AlignedAllocator.h",
981 ] + MICROKERNEL_BENCHMARK_HDRS,
982 deps = MICROKERNEL_BENCHMARK_DEPS,
983)
984
985xnnpack_benchmark(
986 name = "f32_spmm_bench",
987 srcs = [
988 "bench/f32-spmm.cc",
989 "bench/gemm.h",
990 "src/xnnpack/AlignedAllocator.h",
991 ] + MICROKERNEL_BENCHMARK_HDRS,
992 copts = ["-Wno-unused-function"],
993 deps = MICROKERNEL_BENCHMARK_DEPS,
994)
995
996xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -0700997 name = "f32_softargmax_bench",
998 srcs = [
999 "bench/f32-softargmax.cc",
1000 ] + MICROKERNEL_BENCHMARK_HDRS,
1001 copts = ["-Wno-unused-function"],
1002 deps = MICROKERNEL_BENCHMARK_DEPS,
1003)
1004
1005xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001006 name = "f32_im2col_gemm_bench",
1007 srcs = [
1008 "bench/f32-im2col-gemm.cc",
1009 "bench/conv.h",
1010 "src/xnnpack/AlignedAllocator.h",
1011 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1012 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
1013)
1014
1015########################### Benchmarks for operators ###########################
1016
1017xnnpack_benchmark(
1018 name = "add_bench",
1019 srcs = ["bench/add.cc"],
1020 deps = OPERATOR_BENCHMARK_DEPS,
1021)
1022
1023xnnpack_benchmark(
1024 name = "average_pooling_bench",
1025 srcs = ["bench/average-pooling.cc"],
1026 deps = OPERATOR_BENCHMARK_DEPS,
1027)
1028
1029xnnpack_benchmark(
1030 name = "channel_shuffle_bench",
1031 srcs = ["bench/channel-shuffle.cc"],
1032 deps = OPERATOR_BENCHMARK_DEPS,
1033)
1034
1035xnnpack_benchmark(
1036 name = "convolution_bench",
1037 srcs = ["bench/convolution.cc"],
1038 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1039 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1040)
1041
1042xnnpack_benchmark(
1043 name = "deconvolution_bench",
1044 srcs = ["bench/deconvolution.cc"],
1045 copts = xnnpack_optional_tflite_copts(),
1046 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1047)
1048
1049xnnpack_benchmark(
1050 name = "global_average_pooling_bench",
1051 srcs = ["bench/global-average-pooling.cc"],
1052 deps = OPERATOR_BENCHMARK_DEPS,
1053)
1054
1055xnnpack_benchmark(
1056 name = "max_pooling_bench",
1057 srcs = ["bench/max-pooling.cc"],
1058 deps = OPERATOR_BENCHMARK_DEPS,
1059)
1060
1061xnnpack_benchmark(
1062 name = "sigmoid_bench",
1063 srcs = ["bench/sigmoid.cc"],
1064 deps = OPERATOR_BENCHMARK_DEPS,
1065)
1066
1067xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07001068 name = "prelu_bench",
1069 srcs = ["bench/prelu.cc"],
1070 copts = xnnpack_optional_tflite_copts(),
1071 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1072)
1073
1074xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001075 name = "softargmax_bench",
1076 srcs = ["bench/softargmax.cc"],
1077 deps = OPERATOR_BENCHMARK_DEPS,
1078)
1079
Marat Dukhanc068bb62019-10-04 13:24:39 -07001080############################# End-to-end benchmarks ############################
1081
1082cc_library(
1083 name = "mobilenet_v1",
1084 srcs = ["models/mobilenet-v1.cc"],
1085 hdrs = ["models/models.h"],
1086 linkstatic = True,
1087 deps = [
1088 ":XNNPACK",
1089 "@pthreadpool",
1090 ],
1091)
1092
1093cc_library(
1094 name = "mobilenet_v2",
1095 srcs = ["models/mobilenet-v2.cc"],
1096 hdrs = ["models/models.h"],
1097 linkstatic = True,
1098 deps = [
1099 ":XNNPACK",
1100 "@pthreadpool",
1101 ],
1102)
1103
1104xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07001105 name = "f32_dwconv_e2e_bench",
1106 srcs = ["bench/f32-dwconv-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1107 copts = ["-Wno-unused-function"],
1108 deps = MICROKERNEL_BENCHMARK_DEPS + [
1109 ":XNNPACK",
1110 ":mobilenet_v1",
1111 ":mobilenet_v2",
1112 ],
1113)
1114
1115xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07001116 name = "f32_gemm_e2e_bench",
1117 srcs = ["bench/f32-gemm-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1118 copts = ["-Wno-unused-function"],
1119 deps = MICROKERNEL_BENCHMARK_DEPS + [
1120 ":XNNPACK",
1121 ":mobilenet_v1",
1122 ":mobilenet_v2",
1123 ],
1124)
1125
1126xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07001127 name = "end2end_bench",
1128 srcs = ["bench/end2end.cc"],
1129 deps = [
1130 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07001131 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07001132 ":mobilenet_v1",
1133 ":mobilenet_v2",
1134 "@pthreadpool",
1135 ],
1136)
1137
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001138#################### Accuracy evaluation for math functions ####################
1139
1140xnnpack_benchmark(
1141 name = "f32_exp_eval",
1142 srcs = [
1143 "eval/f32-exp.cc",
1144 "src/xnnpack/AlignedAllocator.h",
1145 ] + ACCURACY_EVAL_HDRS,
1146 deps = ACCURACY_EVAL_DEPS,
1147)
1148
Marat Dukhan515c9772019-10-17 18:07:57 -07001149xnnpack_benchmark(
1150 name = "f32_expminus_eval",
1151 srcs = [
1152 "eval/f32-expminus.cc",
1153 "src/xnnpack/AlignedAllocator.h",
1154 ] + ACCURACY_EVAL_HDRS,
1155 deps = ACCURACY_EVAL_DEPS,
1156)
1157
Marat Dukhan98ba4412019-10-23 02:14:28 -07001158xnnpack_benchmark(
1159 name = "f32_extexp_eval",
1160 srcs = [
1161 "eval/f32-extexp.cc",
1162 "src/xnnpack/AlignedAllocator.h",
1163 ] + ACCURACY_EVAL_HDRS,
1164 deps = ACCURACY_EVAL_DEPS,
1165)
1166
Marat Dukhan08c4a432019-10-03 09:29:21 -07001167######################### Unit tests for micro-kernels #########################
1168
1169xnnpack_unit_test(
1170 name = "f16_gemm_test",
1171 srcs = [
1172 "test/f16-gemm.cc",
1173 "test/gemm-microkernel-tester.h",
1174 "src/xnnpack/AlignedAllocator.h",
1175 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1176 deps = MICROKERNEL_TEST_DEPS,
1177)
1178
1179xnnpack_unit_test(
1180 name = "f32_argmaxpool_test",
1181 srcs = [
1182 "test/f32-argmaxpool.cc",
1183 "test/argmaxpool-microkernel-tester.h",
1184 "src/xnnpack/AlignedAllocator.h",
1185 ] + MICROKERNEL_TEST_HDRS,
1186 deps = MICROKERNEL_TEST_DEPS,
1187)
1188
1189xnnpack_unit_test(
1190 name = "f32_avgpool_test",
1191 srcs = [
1192 "test/f32-avgpool.cc",
1193 "test/avgpool-microkernel-tester.h",
1194 "src/xnnpack/AlignedAllocator.h",
1195 ] + MICROKERNEL_TEST_HDRS,
1196 deps = MICROKERNEL_TEST_DEPS,
1197)
1198
1199xnnpack_unit_test(
1200 name = "f32_clamp_test",
1201 srcs = [
1202 "test/f32-clamp.cc",
1203 "test/clamp-microkernel-tester.h",
1204 ] + MICROKERNEL_TEST_HDRS,
1205 deps = MICROKERNEL_TEST_DEPS,
1206)
1207
1208xnnpack_unit_test(
1209 name = "f32_igemm_test",
1210 srcs = [
1211 "test/f32-igemm.cc",
1212 "test/gemm-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_conv_hwc_test",
1220 srcs = [
1221 "test/f32-conv-hwc.cc",
1222 "test/conv-hwc-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_conv_hwc2spchw_test",
1230 srcs = [
1231 "test/f32-conv-hwc2spchw.cc",
1232 "test/conv-hwc2spchw-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_dwconv_test",
1240 srcs = [
1241 "test/f32-dwconv.cc",
1242 "test/dwconv-microkernel-tester.h",
1243 "src/xnnpack/AlignedAllocator.h",
1244 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1245 deps = MICROKERNEL_TEST_DEPS,
1246)
1247
1248xnnpack_unit_test(
1249 name = "f32_dwconv_spchw_test",
1250 srcs = [
1251 "test/f32-dwconv-spchw.cc",
1252 "test/dwconv-spchw-microkernel-tester.h",
1253 "src/xnnpack/AlignedAllocator.h",
1254 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1255 deps = MICROKERNEL_TEST_DEPS,
1256)
1257
1258xnnpack_unit_test(
1259 name = "f32_gavgpool_test",
1260 srcs = [
1261 "test/f32-gavgpool.cc",
1262 "test/gavgpool-microkernel-tester.h",
1263 "src/xnnpack/AlignedAllocator.h",
1264 ] + MICROKERNEL_TEST_HDRS,
1265 deps = MICROKERNEL_TEST_DEPS,
1266)
1267
1268xnnpack_unit_test(
1269 name = "f32_gavgpool_spchw_test",
1270 srcs = [
1271 "test/f32-gavgpool-spchw.cc",
1272 "test/gavgpool-spchw-microkernel-tester.h",
1273 "src/xnnpack/AlignedAllocator.h",
1274 ] + MICROKERNEL_TEST_HDRS,
1275 deps = MICROKERNEL_TEST_DEPS,
1276)
1277
1278xnnpack_unit_test(
1279 name = "f32_gemm_test",
1280 srcs = [
1281 "test/f32-gemm.cc",
1282 "test/gemm-microkernel-tester.h",
1283 "src/xnnpack/AlignedAllocator.h",
1284 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1285 deps = MICROKERNEL_TEST_DEPS,
1286)
1287
1288xnnpack_unit_test(
1289 name = "f32_gemminc_test",
1290 srcs = [
1291 "test/f32-gemminc.cc",
1292 "test/gemm-microkernel-tester.h",
1293 "src/xnnpack/AlignedAllocator.h",
1294 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1295 deps = MICROKERNEL_TEST_DEPS,
1296)
1297
1298xnnpack_unit_test(
1299 name = "f32_hswish_test",
1300 srcs = [
1301 "test/f32-hswish.cc",
1302 "test/hswish-microkernel-tester.h",
1303 ] + MICROKERNEL_TEST_HDRS,
1304 deps = MICROKERNEL_TEST_DEPS,
1305)
1306
1307xnnpack_unit_test(
1308 name = "f32_maxpool_test",
1309 srcs = [
1310 "test/f32-maxpool.cc",
1311 "test/maxpool-microkernel-tester.h",
1312 ] + MICROKERNEL_TEST_HDRS,
1313 deps = MICROKERNEL_TEST_DEPS,
1314)
1315
1316xnnpack_unit_test(
1317 name = "f32_pavgpool_test",
1318 srcs = [
1319 "test/f32-pavgpool.cc",
1320 "test/avgpool-microkernel-tester.h",
1321 "src/xnnpack/AlignedAllocator.h",
1322 ] + MICROKERNEL_TEST_HDRS,
1323 deps = MICROKERNEL_TEST_DEPS,
1324)
1325
1326xnnpack_unit_test(
1327 name = "f32_ppmm_test",
1328 srcs = [
1329 "test/f32-ppmm.cc",
1330 "test/gemm-microkernel-tester.h",
1331 "src/xnnpack/AlignedAllocator.h",
1332 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1333 deps = MICROKERNEL_TEST_DEPS,
1334)
1335
1336xnnpack_unit_test(
1337 name = "f32_prelu_test",
1338 srcs = [
1339 "test/f32-prelu.cc",
1340 "test/prelu-microkernel-tester.h",
1341 "src/xnnpack/AlignedAllocator.h",
1342 ] + MICROKERNEL_TEST_HDRS,
1343 deps = MICROKERNEL_TEST_DEPS,
1344)
1345
1346xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001347 name = "f32_raddexpminusmax_test",
1348 srcs = [
1349 "test/f32-raddexpminusmax.cc",
1350 "test/raddexpminusmax-microkernel-tester.h",
1351 ] + MICROKERNEL_TEST_HDRS,
1352 deps = MICROKERNEL_TEST_DEPS,
1353)
1354
1355xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001356 name = "f32_raddextexp_test",
1357 srcs = [
1358 "test/f32-raddextexp.cc",
1359 "test/raddextexp-microkernel-tester.h",
1360 ] + MICROKERNEL_TEST_HDRS,
1361 deps = MICROKERNEL_TEST_DEPS,
1362)
1363
1364xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001365 name = "f32_raddstoreexpminusmax_test",
1366 srcs = [
1367 "test/f32-raddstoreexpminusmax.cc",
1368 "test/raddstoreexpminusmax-microkernel-tester.h",
1369 ] + MICROKERNEL_TEST_HDRS,
1370 deps = MICROKERNEL_TEST_DEPS,
1371)
1372
1373xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001374 name = "f32_rmax_test",
1375 srcs = [
1376 "test/f32-rmax.cc",
1377 "test/rmax-microkernel-tester.h",
1378 ] + MICROKERNEL_TEST_HDRS,
1379 deps = MICROKERNEL_TEST_DEPS,
1380)
1381
1382xnnpack_unit_test(
1383 name = "f32_spmm_test",
1384 srcs = [
1385 "test/f32-spmm.cc",
1386 "test/spmm-microkernel-tester.h",
1387 "src/xnnpack/AlignedAllocator.h",
1388 ] + MICROKERNEL_TEST_HDRS,
1389 deps = MICROKERNEL_TEST_DEPS,
1390)
1391
1392xnnpack_unit_test(
1393 name = "f32_vadd_test",
1394 srcs = [
1395 "test/f32-vadd.cc",
1396 "test/vadd-microkernel-tester.h",
1397 ] + MICROKERNEL_TEST_HDRS,
1398 deps = MICROKERNEL_TEST_DEPS,
1399)
1400
1401xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001402 name = "f32_vmul_test",
1403 srcs = [
1404 "test/f32-vmul.cc",
1405 "test/vmul-microkernel-tester.h",
1406 ] + MICROKERNEL_TEST_HDRS,
1407 deps = MICROKERNEL_TEST_DEPS,
1408)
1409
1410xnnpack_unit_test(
1411 name = "f32_vmulcaddc_test",
1412 srcs = [
1413 "test/f32-vmulcaddc.cc",
1414 "test/vmulcaddc-microkernel-tester.h",
1415 "src/xnnpack/AlignedAllocator.h",
1416 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1417 deps = MICROKERNEL_TEST_DEPS,
1418)
1419
1420xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001421 name = "f32_vscale_test",
1422 srcs = [
1423 "test/f32-vscale.cc",
1424 "test/vscale-microkernel-tester.h",
1425 ] + MICROKERNEL_TEST_HDRS,
1426 deps = MICROKERNEL_TEST_DEPS,
1427)
1428
1429xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001430 name = "f32_vscaleexpminusmax_test",
1431 srcs = [
1432 "test/f32-vscaleexpminusmax.cc",
1433 "test/vscaleexpminusmax-microkernel-tester.h",
1434 ] + MICROKERNEL_TEST_HDRS,
1435 deps = MICROKERNEL_TEST_DEPS,
1436)
1437
1438xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001439 name = "f32_vscaleextexp_test",
1440 srcs = [
1441 "test/f32-vscaleextexp.cc",
1442 "test/vscaleextexp-microkernel-tester.h",
1443 ] + MICROKERNEL_TEST_HDRS,
1444 deps = MICROKERNEL_TEST_DEPS,
1445)
1446
1447xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001448 name = "f32_vsub_test",
1449 srcs = [
1450 "test/f32-vsub.cc",
1451 "test/vsub-microkernel-tester.h",
1452 ] + MICROKERNEL_TEST_HDRS,
1453 deps = MICROKERNEL_TEST_DEPS,
1454)
1455
1456xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001457 name = "q8_avgpool_test",
1458 srcs = [
1459 "test/q8-avgpool.cc",
1460 "test/avgpool-microkernel-tester.h",
1461 "src/xnnpack/AlignedAllocator.h",
1462 ] + MICROKERNEL_TEST_HDRS,
1463 deps = MICROKERNEL_TEST_DEPS,
1464)
1465
1466xnnpack_unit_test(
1467 name = "q8_igemm_test",
1468 srcs = [
1469 "test/q8-igemm.cc",
1470 "test/gemm-microkernel-tester.h",
1471 "src/xnnpack/AlignedAllocator.h",
1472 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1473 deps = MICROKERNEL_TEST_DEPS,
1474)
1475
1476xnnpack_unit_test(
1477 name = "q8_dwconv_test",
1478 srcs = [
1479 "test/q8-dwconv.cc",
1480 "test/dwconv-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_gavgpool_test",
1488 srcs = [
1489 "test/q8-gavgpool.cc",
1490 "test/gavgpool-microkernel-tester.h",
1491 "src/xnnpack/AlignedAllocator.h",
1492 ] + MICROKERNEL_TEST_HDRS,
1493 deps = MICROKERNEL_TEST_DEPS,
1494)
1495
1496xnnpack_unit_test(
1497 name = "q8_gemm_test",
1498 srcs = [
1499 "test/q8-gemm.cc",
1500 "test/gemm-microkernel-tester.h",
1501 "src/xnnpack/AlignedAllocator.h",
1502 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1503 deps = MICROKERNEL_TEST_DEPS,
1504)
1505
1506xnnpack_unit_test(
1507 name = "q8_vadd_test",
1508 srcs = [
1509 "test/q8-vadd.cc",
1510 "test/vadd-microkernel-tester.h",
1511 ] + MICROKERNEL_TEST_HDRS,
1512 deps = MICROKERNEL_TEST_DEPS,
1513)
1514
1515xnnpack_unit_test(
1516 name = "u8_clamp_test",
1517 srcs = [
1518 "test/u8-clamp.cc",
1519 "test/clamp-microkernel-tester.h",
1520 ] + MICROKERNEL_TEST_HDRS,
1521 deps = MICROKERNEL_TEST_DEPS,
1522)
1523
1524xnnpack_unit_test(
1525 name = "u8_lut32norm_test",
1526 srcs = [
1527 "test/u8-lut32norm.cc",
1528 "test/lut-norm-microkernel-tester.h",
1529 ] + MICROKERNEL_TEST_HDRS,
1530 deps = MICROKERNEL_TEST_DEPS,
1531)
1532
1533xnnpack_unit_test(
1534 name = "u8_maxpool_test",
1535 srcs = [
1536 "test/u8-maxpool.cc",
1537 "test/maxpool-microkernel-tester.h",
1538 ] + MICROKERNEL_TEST_HDRS,
1539 deps = MICROKERNEL_TEST_DEPS,
1540)
1541
1542xnnpack_unit_test(
1543 name = "u8_rmax_test",
1544 srcs = [
1545 "test/u8-rmax.cc",
1546 "test/rmax-microkernel-tester.h",
1547 ] + MICROKERNEL_TEST_HDRS,
1548 deps = MICROKERNEL_TEST_DEPS,
1549)
1550
1551xnnpack_unit_test(
1552 name = "x32_packx_test",
1553 srcs = [
1554 "test/x32-packx.cc",
1555 "test/pack-microkernel-tester.h",
1556 "src/xnnpack/AlignedAllocator.h",
1557 ] + MICROKERNEL_TEST_HDRS,
1558 deps = MICROKERNEL_TEST_DEPS,
1559)
1560
1561xnnpack_unit_test(
1562 name = "x32_pad_test",
1563 srcs = [
1564 "test/x32-pad.cc",
1565 "test/pad-microkernel-tester.h",
1566 ] + MICROKERNEL_TEST_HDRS,
1567 deps = MICROKERNEL_TEST_DEPS,
1568)
1569
1570xnnpack_unit_test(
1571 name = "x32_unpool_test",
1572 srcs = [
1573 "test/x32-unpool.cc",
1574 "test/unpool-microkernel-tester.h",
1575 ] + MICROKERNEL_TEST_HDRS,
1576 deps = MICROKERNEL_TEST_DEPS,
1577)
1578
1579xnnpack_unit_test(
1580 name = "x32_zip_test",
1581 srcs = [
1582 "test/x32-zip.cc",
1583 "test/zip-microkernel-tester.h",
1584 ] + MICROKERNEL_TEST_HDRS,
1585 deps = MICROKERNEL_TEST_DEPS,
1586)
1587
1588xnnpack_unit_test(
1589 name = "x8_lut_test",
1590 srcs = [
1591 "test/x8-lut.cc",
1592 "test/lut-microkernel-tester.h",
1593 ] + MICROKERNEL_TEST_HDRS,
1594 deps = MICROKERNEL_TEST_DEPS,
1595)
1596
1597xnnpack_unit_test(
1598 name = "x8_zip_test",
1599 srcs = [
1600 "test/x8-zip.cc",
1601 "test/zip-microkernel-tester.h",
1602 ] + MICROKERNEL_TEST_HDRS,
1603 deps = MICROKERNEL_TEST_DEPS,
1604)
1605
1606########################### Size test for the library ##########################
1607
1608xnnpack_binary(
1609 name = "size_test",
1610 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001611 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001612)
1613
1614########################### Unit tests for operators ###########################
1615
1616xnnpack_unit_test(
1617 name = "add_test",
1618 srcs = [
1619 "test/add.cc",
1620 "test/add-operator-tester.h",
1621 ],
1622 deps = OPERATOR_TEST_DEPS,
1623)
1624
1625xnnpack_unit_test(
1626 name = "argmax_pooling_test",
1627 srcs = [
1628 "test/argmax-pooling.cc",
1629 "test/argmax-pooling-operator-tester.h",
1630 ] + OPERATOR_TEST_PARAMS_HDRS,
1631 deps = OPERATOR_TEST_DEPS,
1632)
1633
1634xnnpack_unit_test(
1635 name = "average_pooling_test",
1636 srcs = [
1637 "test/average-pooling.cc",
1638 "test/average-pooling-operator-tester.h",
1639 ] + OPERATOR_TEST_PARAMS_HDRS,
1640 deps = OPERATOR_TEST_DEPS,
1641)
1642
1643xnnpack_unit_test(
1644 name = "channel_pad_test",
1645 srcs = [
1646 "test/channel-pad.cc",
1647 "test/channel-pad-operator-tester.h",
1648 ] + OPERATOR_TEST_PARAMS_HDRS,
1649 deps = OPERATOR_TEST_DEPS,
1650)
1651
1652xnnpack_unit_test(
1653 name = "channel_shuffle_test",
1654 srcs = [
1655 "test/channel-shuffle.cc",
1656 "test/channel-shuffle-operator-tester.h",
1657 ],
1658 deps = OPERATOR_TEST_DEPS,
1659)
1660
1661xnnpack_unit_test(
1662 name = "clamp_test",
1663 srcs = [
1664 "test/clamp.cc",
1665 "test/clamp-operator-tester.h",
1666 ],
1667 deps = OPERATOR_TEST_DEPS,
1668)
1669
1670xnnpack_unit_test(
1671 name = "convolution_test",
1672 srcs = [
1673 "test/convolution.cc",
1674 "test/convolution-operator-tester.h",
1675 ],
1676 deps = OPERATOR_TEST_DEPS,
1677)
1678
1679xnnpack_unit_test(
1680 name = "convolution_spnchw_test",
1681 srcs = [
1682 "test/convolution-spnchw.cc",
1683 "test/convolution-spnchw-operator-tester.h",
1684 ],
1685 deps = OPERATOR_TEST_DEPS,
1686)
1687
1688xnnpack_unit_test(
1689 name = "deconvolution_test",
1690 srcs = [
1691 "test/deconvolution.cc",
1692 "test/deconvolution-operator-tester.h",
1693 ] + OPERATOR_TEST_PARAMS_HDRS,
1694 deps = OPERATOR_TEST_DEPS,
1695)
1696
1697xnnpack_unit_test(
1698 name = "fully_connected_test",
1699 srcs = [
1700 "test/fully-connected.cc",
1701 "test/fully-connected-operator-tester.h",
1702 ],
1703 deps = OPERATOR_TEST_DEPS,
1704)
1705
1706xnnpack_unit_test(
1707 name = "global_average_pooling_test",
1708 srcs = [
1709 "test/global-average-pooling.cc",
1710 "test/global-average-pooling-operator-tester.h",
1711 ] + OPERATOR_TEST_PARAMS_HDRS,
1712 deps = OPERATOR_TEST_DEPS,
1713)
1714
1715xnnpack_unit_test(
1716 name = "global_average_pooling_spnchw_test",
1717 srcs = [
1718 "test/global-average-pooling-spnchw.cc",
1719 "test/global-average-pooling-spnchw-operator-tester.h",
1720 ],
1721 deps = OPERATOR_TEST_DEPS,
1722)
1723
1724xnnpack_unit_test(
1725 name = "hardswish_test",
1726 srcs = [
1727 "test/hardswish.cc",
1728 "test/hardswish-operator-tester.h",
1729 ],
1730 deps = OPERATOR_TEST_DEPS,
1731)
1732
1733xnnpack_unit_test(
1734 name = "leaky_relu_test",
1735 srcs = [
1736 "test/leaky-relu.cc",
1737 "test/leaky-relu-operator-tester.h",
1738 ],
1739 deps = OPERATOR_TEST_DEPS,
1740)
1741
1742xnnpack_unit_test(
1743 name = "max_pooling_test",
1744 srcs = [
1745 "test/max-pooling.cc",
1746 "test/max-pooling-operator-tester.h",
1747 ] + OPERATOR_TEST_PARAMS_HDRS,
1748 deps = OPERATOR_TEST_DEPS,
1749)
1750
1751xnnpack_unit_test(
1752 name = "prelu_test",
1753 srcs = [
1754 "test/prelu.cc",
1755 "test/prelu-operator-tester.h",
1756 ] + OPERATOR_TEST_PARAMS_HDRS,
1757 deps = OPERATOR_TEST_DEPS,
1758)
1759
1760xnnpack_unit_test(
1761 name = "sigmoid_test",
1762 srcs = [
1763 "test/sigmoid.cc",
1764 "test/sigmoid-operator-tester.h",
1765 ],
1766 deps = OPERATOR_TEST_DEPS,
1767)
1768
1769xnnpack_unit_test(
1770 name = "softargmax_test",
1771 srcs = [
1772 "test/softargmax.cc",
1773 "test/softargmax-operator-tester.h",
1774 ],
1775 deps = OPERATOR_TEST_DEPS,
1776)
1777
1778xnnpack_unit_test(
1779 name = "unpooling_test",
1780 srcs = [
1781 "test/unpooling.cc",
1782 "test/unpooling-operator-tester.h",
1783 ],
1784 deps = OPERATOR_TEST_DEPS,
1785)
1786
1787############################# Build configurations #############################
1788
Marat Dukhanb8642352019-10-30 15:43:02 -07001789# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07001790config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07001791 name = "xnn_enable_assembly_explicit_true",
1792 define_values = {"xnn_enable_assembly": "true"},
1793)
1794
1795# Disables usage of assembly kernels.
1796config_setting(
1797 name = "xnn_enable_assembly_explicit_false",
1798 define_values = {"xnn_enable_assembly": "false"},
1799)
1800
1801# Builds with -c dbg
1802config_setting(
1803 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001804 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07001805 "compilation_mode": "dbg",
1806 },
1807)
1808
1809# Builds with -c opt
1810config_setting(
1811 name = "optimized_build",
1812 values = {
1813 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001814 },
1815)
1816
1817config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07001818 name = "linux_k8",
1819 values = {"cpu": "k8"},
1820)
1821
1822config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07001823 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07001824 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07001825)
1826
1827config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001828 name = "android",
1829 values = {"crosstool_top": "//external:android/crosstool"},
1830)
1831
1832config_setting(
1833 name = "android_armv7",
1834 values = {
1835 "crosstool_top": "//external:android/crosstool",
1836 "cpu": "armeabi-v7a",
1837 },
1838)
1839
1840config_setting(
1841 name = "android_arm64",
1842 values = {
1843 "crosstool_top": "//external:android/crosstool",
1844 "cpu": "arm64-v8a",
1845 },
1846)
1847
1848config_setting(
1849 name = "android_x86",
1850 values = {
1851 "crosstool_top": "//external:android/crosstool",
1852 "cpu": "x86",
1853 },
1854)
1855
1856config_setting(
1857 name = "android_x86_64",
1858 values = {
1859 "crosstool_top": "//external:android/crosstool",
1860 "cpu": "x86_64",
1861 },
1862)
1863
1864config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07001865 name = "macos_x86_64",
1866 values = {
1867 "apple_platform_type": "macos",
1868 "cpu": "darwin",
1869 },
1870)
1871
1872config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001873 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001874 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07001875)
1876
1877config_setting(
1878 name = "emscripten_wasm",
1879 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001880 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001881 "cpu": "wasm",
1882 },
1883)
1884
1885config_setting(
1886 name = "emscripten_wasmsimd",
1887 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001888 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001889 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07001890 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001891 },
1892)
1893
1894config_setting(
1895 name = "emscripten_asmjs",
1896 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001897 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001898 "cpu": "asmjs",
1899 },
1900)