blob: 749fc697306ddb422cf876b522842486e2062b46 [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",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800124 "src/f32-vmulcaddc/c1-scalar-2x.c",
125 "src/f32-vmulcaddc/c2-scalar-2x.c",
126 "src/f32-vmulcaddc/c4-scalar-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700127 "src/f32-vsub/scalar.c",
128 "src/q8-avgpool/mp9p8q-scalar.c",
129 "src/q8-avgpool/up9-scalar.c",
130 "src/q8-igemm/2x2-scalar.c",
131 "src/q8-dwconv/up1x9-scalar.c",
132 "src/q8-gavgpool/mp7p7q-scalar.c",
133 "src/q8-gavgpool/up7-scalar.c",
134 "src/q8-gemm/2x2-scalar.c",
135 "src/q8-vadd/scalar.c",
136 "src/u8-clamp/scalar.c",
137 "src/u8-lut32norm/scalar.c",
138 "src/u8-maxpool/9p8q-scalar.c",
139 "src/u8-rmax/scalar.c",
140 "src/x32-packx/x2-scalar.c",
141 "src/x32-packx/x3-scalar.c",
142 "src/x32-packx/x4-scalar.c",
143 "src/x32-pad/x2-scalar.c",
144 "src/x32-unpool/scalar.c",
145 "src/x32-zip/x2-scalar.c",
146 "src/x32-zip/x3-scalar.c",
147 "src/x32-zip/x4-scalar.c",
148 "src/x32-zip/xm-scalar.c",
149 "src/x8-lut/scalar.c",
150 "src/x8-zip/x2-scalar.c",
151 "src/x8-zip/x3-scalar.c",
152 "src/x8-zip/x4-scalar.c",
153 "src/x8-zip/xm-scalar.c",
154]
155
156PSIMD_UKERNELS = [
157 "src/f32-argmaxpool/mp9p8q-psimd.c",
158 "src/f32-argmaxpool/up4-psimd.c",
159 "src/f32-argmaxpool/up9-psimd.c",
160 "src/f32-avgpool/mp9p8q-psimd.c",
161 "src/f32-avgpool/up9-psimd.c",
162 "src/f32-clamp/psimd.c",
163 "src/f32-igemm/1x8-psimd-loadsplat.c",
164 "src/f32-igemm/1x8-psimd-splat.c",
165 "src/f32-igemm/1x8s4-psimd.c",
166 "src/f32-igemm/4x2c4-psimd.c",
167 "src/f32-igemm/4x8-psimd-loadsplat.c",
168 "src/f32-igemm/4x8-psimd-splat.c",
169 "src/f32-igemm/4x8s4-psimd.c",
170 "src/f32-igemm/6x8-psimd-loadsplat.c",
171 "src/f32-igemm/6x8-psimd-splat.c",
172 "src/f32-igemm/6x8s4-psimd.c",
173 "src/f32-dwconv/up4x25-psimd.c",
174 "src/f32-dwconv/up4x4-psimd.c",
175 "src/f32-dwconv/up4x9-psimd.c",
176 "src/f32-gavgpool/mp7p7q-psimd.c",
177 "src/f32-gavgpool/up7-psimd.c",
178 "src/f32-gemm/1x8-psimd-loadsplat.c",
179 "src/f32-gemm/1x8-psimd-splat.c",
180 "src/f32-gemm/1x8s4-psimd.c",
181 "src/f32-gemm/4x8-psimd-loadsplat.c",
182 "src/f32-gemm/4x8-psimd-splat.c",
183 "src/f32-gemm/4x8s4-psimd.c",
184 "src/f32-gemm/6x8-psimd-loadsplat.c",
185 "src/f32-gemm/6x8-psimd-splat.c",
186 "src/f32-gemm/6x8s4-psimd.c",
187 "src/f32-gemminc/1x8-psimd-loadsplat.c",
188 "src/f32-gemminc/1x8-psimd-splat.c",
189 "src/f32-gemminc/1x8s4-psimd.c",
190 "src/f32-gemminc/4x8-psimd-loadsplat.c",
191 "src/f32-gemminc/4x8-psimd-splat.c",
192 "src/f32-gemminc/4x8s4-psimd.c",
193 "src/f32-gemminc/6x8-psimd-loadsplat.c",
194 "src/f32-gemminc/6x8-psimd-splat.c",
195 "src/f32-gemminc/6x8s4-psimd.c",
196 "src/f32-hswish/psimd.c",
197 "src/f32-maxpool/9p8q-psimd.c",
198 "src/f32-pavgpool/mp9p8q-psimd.c",
199 "src/f32-pavgpool/up9-psimd.c",
200 "src/f32-ppmm/4x8-psimd.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800201 "src/f32-prelu/psimd-2x4.c",
202 "src/f32-prelu/psimd-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700203 "src/f32-vadd/psimd.c",
204 "src/f32-vmul/psimd.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800205 "src/f32-vmulcaddc/c4-psimd-2x.c",
206 "src/f32-vmulcaddc/c8-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700207 "src/f32-vsub/psimd.c",
208 "src/x32-packx/x4-psimd.c",
209 "src/x32-pad/x2-psimd.c",
210 "src/x32-unpool/psimd.c",
211 "src/x32-zip/x2-psimd.c",
212 "src/x32-zip/x3-psimd.c",
213 "src/x32-zip/x4-psimd.c",
214 "src/x32-zip/xm-psimd.c",
215]
216
217# ISA-specific micro-kernels
218NEON_UKERNELS = [
219 "src/f32-avgpool/mp9p8q-neon.c",
220 "src/f32-avgpool/up9-neon.c",
221 "src/f32-clamp/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700222 "src/f32-dwconv/up4x9-neon.c",
223 "src/f32-gavgpool-spchw/neon-x4.c",
224 "src/f32-gavgpool/mp7p7q-neon.c",
225 "src/f32-gavgpool/up7-neon.c",
226 "src/f32-gemm/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700227 "src/f32-gemm/4x2-neon-ld64.c",
228 "src/f32-gemm/4x8-neon-ld128.c",
229 "src/f32-gemm/4x8-neon-ld64.c",
230 "src/f32-gemm/5x8-neon-ld64.c",
231 "src/f32-gemm/6x8-neon-ld64.c",
232 "src/f32-gemminc/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700233 "src/f32-gemminc/4x8-neon-ld128.c",
234 "src/f32-gemminc/4x8-neon-ld64.c",
235 "src/f32-gemminc/5x8-neon-ld64.c",
236 "src/f32-gemminc/6x8-neon-ld64.c",
237 "src/f32-hswish/neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800238 "src/f32-igemm/1x8-neon-ld64.c",
239 "src/f32-igemm/4x2-neon-ld64.c",
240 "src/f32-igemm/4x4-neon-ld64.c",
241 "src/f32-igemm/4x8-neon-ld128.c",
242 "src/f32-igemm/4x8-neon-ld64.c",
243 "src/f32-igemm/6x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700244 "src/f32-pavgpool/mp9p8q-neon.c",
245 "src/f32-pavgpool/up9-neon.c",
246 "src/f32-ppmm/4x8-neon.c",
247 "src/f32-ppmm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800248 "src/f32-prelu/neon-2x4.c",
249 "src/f32-prelu/neon-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700250 "src/f32-rmax/neon.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800251 "src/f32-vmulcaddc/c4-neon-2x.c",
252 "src/f32-vmulcaddc/c8-neon-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700253 "src/q8-avgpool/mp9p8q-neon.c",
254 "src/q8-avgpool/up9-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700255 "src/q8-dwconv/up8x9-neon.c",
256 "src/q8-gavgpool/mp7p7q-neon.c",
257 "src/q8-gavgpool/up7-neon.c",
258 "src/q8-gemm/4x8-neon.c",
259 "src/q8-gemm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800260 "src/q8-igemm/4x8-neon.c",
261 "src/q8-igemm/8x8-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700262 "src/q8-vadd/neon.c",
263 "src/u8-clamp/neon.c",
264 "src/u8-maxpool/9p8q-neon.c",
265 "src/u8-rmax/neon.c",
266 "src/x32-packx/x4-neon-st4.c",
267 "src/x32-pad/x2-neon.c",
268 "src/x32-zip/x2-neon.c",
269 "src/x32-zip/x3-neon.c",
270 "src/x32-zip/x4-neon.c",
271 "src/x32-zip/xm-neon.c",
272 "src/x8-zip/x2-neon.c",
273 "src/x8-zip/x3-neon.c",
274 "src/x8-zip/x4-neon.c",
275 "src/x8-zip/xm-neon.c",
276]
277
278NEONFMA_UKERNELS = [
Frank Barcharddb45b6a2019-10-09 16:42:45 -0700279 "src/f32-igemm/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700280 "src/f32-igemm/4x2-neonfma-ld64.c",
281 "src/f32-igemm/4x4-neonfma-ld64.c",
282 "src/f32-igemm/4x8-neonfma-ld128.c",
283 "src/f32-igemm/4x8-neonfma-ld64.c",
284 "src/f32-igemm/6x8-neonfma-ld64.c",
285 "src/f32-dwconv/up4x9-neonfma.c",
286 "src/f32-dwconv/up8x9-neonfma.c",
287 "src/f32-gemm/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700288 "src/f32-gemm/4x2-neonfma-ld64.c",
289 "src/f32-gemm/4x8-neonfma-ld128.c",
290 "src/f32-gemm/4x8-neonfma-ld64.c",
291 "src/f32-gemm/5x8-neonfma-ld64.c",
292 "src/f32-gemm/6x8-neonfma-ld64.c",
293 "src/f32-gemminc/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700294 "src/f32-gemminc/4x8-neonfma-ld128.c",
295 "src/f32-gemminc/4x8-neonfma-ld64.c",
296 "src/f32-gemminc/5x8-neonfma-ld64.c",
297 "src/f32-gemminc/6x8-neonfma-ld64.c",
298 "src/f32-hswish/neonfma.c",
299 "src/f32-ppmm/4x8-neonfma.c",
300 "src/f32-ppmm/8x8-neonfma.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800301 "src/f32-vmulcaddc/c4-neonfma-2x.c",
302 "src/f32-vmulcaddc/c8-neonfma-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700303]
304
305AARCH64_NEONFMA_UKERNELS = [
306 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
307 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
308 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
309 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
310 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
311 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
312 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
313 "src/f32-spmm/12x1-neonfma.c",
314 "src/f32-spmm/12x2-neonfma.c",
315 "src/f32-spmm/12x4-neonfma.c",
316 "src/f32-spmm/16x1-neonfma-pipelined.c",
317 "src/f32-spmm/16x1-neonfma-unroll2.c",
318 "src/f32-spmm/16x1-neonfma.c",
319 "src/f32-spmm/16x2-neonfma.c",
320 "src/f32-spmm/16x4-neonfma.c",
321 "src/f32-spmm/4x1-neonfma-pipelined.c",
322 "src/f32-spmm/4x1-neonfma-unroll2.c",
323 "src/f32-spmm/4x1-neonfma.c",
324 "src/f32-spmm/4x2-neonfma.c",
325 "src/f32-spmm/4x4-neonfma.c",
326 "src/f32-spmm/8x1-neonfma-pipelined.c",
327 "src/f32-spmm/8x1-neonfma-unroll2.c",
328 "src/f32-spmm/8x1-neonfma.c",
329 "src/f32-spmm/8x2-neonfma.c",
330 "src/f32-spmm/8x4-neonfma.c",
331]
332
333AARCH64_NEONFP16ARITH_UKERNELS = [
334 "src/f16-gemm/4x8-neonfp16arith-ld64.c",
335 "src/f16-gemm/6x8-neonfp16arith-ld64.c",
336 "src/f16-gemm/8x8-neonfp16arith-ld64.c",
337]
338
339SSE_UKERNELS = [
340 "src/f32-avgpool/mp9p8q-sse.c",
341 "src/f32-avgpool/up9-sse.c",
342 "src/f32-clamp/sse.c",
343 "src/f32-igemm/1x8-sse-dup.c",
344 "src/f32-igemm/1x8-sse-load1.c",
345 "src/f32-igemm/1x8s4-sse.c",
346 "src/f32-igemm/4x2c4-sse.c",
347 "src/f32-igemm/4x8-sse-dup.c",
348 "src/f32-igemm/4x8-sse-load1.c",
349 "src/f32-igemm/4x8s4-sse.c",
350 "src/f32-dwconv/up4x25-sse.c",
351 "src/f32-dwconv/up4x4-sse.c",
352 "src/f32-dwconv/up4x9-sse.c",
353 "src/f32-gavgpool-spchw/sse-x4.c",
354 "src/f32-gavgpool/mp7p7q-sse.c",
355 "src/f32-gavgpool/up7-sse.c",
356 "src/f32-gemm/1x8-sse-dup.c",
357 "src/f32-gemm/1x8-sse-load1.c",
358 "src/f32-gemm/1x8s4-sse.c",
359 "src/f32-gemm/4x8-sse-dup.c",
360 "src/f32-gemm/4x8-sse-load1.c",
361 "src/f32-gemm/4x8s4-sse.c",
362 "src/f32-gemminc/1x8-sse-dup.c",
363 "src/f32-gemminc/1x8-sse-load1.c",
364 "src/f32-gemminc/1x8s4-sse.c",
365 "src/f32-gemminc/4x8-sse-dup.c",
366 "src/f32-gemminc/4x8-sse-load1.c",
367 "src/f32-gemminc/4x8s4-sse.c",
368 "src/f32-hswish/sse.c",
369 "src/f32-maxpool/9p8q-sse.c",
370 "src/f32-pavgpool/mp9p8q-sse.c",
371 "src/f32-pavgpool/up9-sse.c",
372 "src/f32-dwconv-spchw/3x3p1-sse.c",
373 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
374 "src/f32-ppmm/4x8-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700375 "src/f32-rmax/sse.c",
376 "src/f32-spmm/4x1-sse.c",
377 "src/f32-spmm/8x1-sse.c",
378 "src/f32-vadd/sse.c",
379 "src/f32-vmul/sse.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800380 "src/f32-vmulcaddc/c4-sse-2x.c",
381 "src/f32-vmulcaddc/c8-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700382 "src/f32-vsub/sse.c",
383 "src/x32-packx/x4-sse.c",
384]
385
386SSE2_UKERNELS = [
387 "src/f32-argmaxpool/mp9p8q-sse2.c",
388 "src/f32-argmaxpool/up4-sse2.c",
389 "src/f32-argmaxpool/up9-sse2.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800390 "src/f32-prelu/sse2-2x4.c",
391 "src/f32-prelu/sse2-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700392 "src/q8-avgpool/mp9p8q-sse2.c",
393 "src/q8-avgpool/up9-sse2.c",
394 "src/q8-igemm/4x4c2-sse2.c",
395 "src/q8-dwconv/up8x9-sse2.c",
396 "src/q8-gavgpool/mp7p7q-sse2.c",
397 "src/q8-gavgpool/up7-sse2.c",
398 "src/q8-gemm/2x4c8-sse2.c",
399 "src/q8-gemm/4x4c2-sse2.c",
400 "src/q8-vadd/sse2.c",
401 "src/u8-clamp/sse2.c",
402 "src/u8-maxpool/9p8q-sse2.c",
403 "src/u8-rmax/sse2.c",
404 "src/x32-pad/x2-sse2.c",
405 "src/x32-zip/x2-sse2.c",
406 "src/x32-zip/x3-sse2.c",
407 "src/x32-zip/x4-sse2.c",
408 "src/x32-zip/xm-sse2.c",
409 "src/x8-zip/x2-sse2.c",
410 "src/x8-zip/x3-sse2.c",
411 "src/x8-zip/x4-sse2.c",
412 "src/x8-zip/xm-sse2.c",
413]
414
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800415SSE41_UKERNELS = [
416 "src/f32-prelu/sse41-2x4.c",
417 "src/f32-prelu/sse41-2x8.c",
418]
419
Marat Dukhan08c4a432019-10-03 09:29:21 -0700420AVX_UKERNELS = [
421 "src/f32-rmax/avx.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700422 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700423]
424
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700425AVX2_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700426 "src/f32-raddexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700427 "src/f32-raddextexp/avx2-p5-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700428 "src/f32-raddstoreexpminusmax/avx2-p5-unroll64.c",
429 "src/f32-vscaleexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700430 "src/f32-vscaleextexp/avx2-p5-unroll64.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700431 "src/math/exp-avx2-p5.c",
432 "src/math/exp-avx2-perm-p3.c",
433 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -0700434 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700435 "src/math/extexp-avx2-p5.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700436]
437
Marat Dukhan08c4a432019-10-03 09:29:21 -0700438AVX512F_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700439 "src/f32-raddexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700440 "src/f32-raddextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700441 "src/f32-raddstoreexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700442 "src/f32-rmax/avx512f.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700443 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700444 "src/f32-vscaleexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700445 "src/f32-vscaleextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700446 "src/math/exp-avx512f-p5-scalef.c",
447 "src/math/exp-avx512f-p5.c",
448 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -0700449 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700450 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700451]
452
453AARCH32_ASM_UKERNELS = [
454 "src/q8-dwconv/up8x9-aarch32-neon.S",
455]
456
457AARCH64_ASM_UKERNELS = [
458 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
459 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
460 "src/f32-gemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700461 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700462 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a57.S",
463 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a75.S",
464 "src/f32-gemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700465 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700466 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a57.S",
467 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a75.S",
468 "src/f32-gemm/4x8-aarch64-neonfma-ld128.S",
469 "src/f32-gemm/4x8-aarch64-neonfma-ld64.S",
470 "src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700471 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700472 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a57.S",
473 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S",
474 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S",
475 "src/f32-gemm/6x8-aarch64-neonfma-ld128.S",
476 "src/f32-gemm/6x8-aarch64-neonfma-ld64.S",
477 "src/f32-gemminc/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700478 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700479 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a57.S",
480 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a75.S",
481 "src/f32-gemminc/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700482 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700483 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a57.S",
484 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a75.S",
485 "src/f32-gemminc/4x8-aarch64-neonfma-ld128.S",
486 "src/f32-gemminc/4x8-aarch64-neonfma-ld64.S",
487 "src/f32-gemminc/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700488 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700489 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a57.S",
490 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a73.S",
491 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a75.S",
492 "src/f32-gemminc/6x8-aarch64-neonfma-ld128.S",
493 "src/f32-gemminc/6x8-aarch64-neonfma-ld64.S",
494 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700495 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700496 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S",
497 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S",
498 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700499 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700500 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
501 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700502 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700503 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
504 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
505 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
506]
507
508INTERNAL_MICROKERNEL_HDRS = [
509 "src/xnnpack/argmaxpool.h",
510 "src/xnnpack/avgpool.h",
511 "src/xnnpack/clamp.h",
512 "src/xnnpack/common.h",
513 "src/xnnpack/conv.h",
514 "src/xnnpack/dwconv.h",
515 "src/xnnpack/gavgpool.h",
516 "src/xnnpack/gemm.h",
517 "src/xnnpack/hswish.h",
518 "src/xnnpack/igemm.h",
519 "src/xnnpack/lut.h",
520 "src/xnnpack/math.h",
521 "src/xnnpack/maxpool.h",
522 "src/xnnpack/packx.h",
523 "src/xnnpack/pad.h",
524 "src/xnnpack/params.h",
525 "src/xnnpack/pavgpool.h",
526 "src/xnnpack/ppmm.h",
527 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700528 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700529 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700530 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700531 "src/xnnpack/rmax.h",
532 "src/xnnpack/scalar-utils.h",
533 "src/xnnpack/spmm.h",
534 "src/xnnpack/unpool.h",
535 "src/xnnpack/vadd.h",
536 "src/xnnpack/vmul.h",
537 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700538 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700539 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700540 "src/xnnpack/vscaleextexp.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700541 "src/xnnpack/vsub.h",
542 "src/xnnpack/zip.h",
543]
544
545INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
546 "include/xnnpack.h",
547 "src/xnnpack/allocator.h",
548 "src/xnnpack/compute.h",
549 "src/xnnpack/im2col.h",
550 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700551 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700552 "src/xnnpack/operator.h",
553 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700554 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700555 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700556 "src/xnnpack/requantization.h",
557]
558
559ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
560 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700561]
562
563MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700564 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700565 "include/xnnpack.h",
566]
567
568MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
569 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700570 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700571 "src/xnnpack/requantization.h",
572 "include/xnnpack.h",
573]
574
575OPERATOR_TEST_PARAMS_HDRS = [
576 "src/xnnpack/params.h",
577 "src/xnnpack/common.h",
578]
579
580WEIGHTS_PACK_HDRS = [
581 "src/xnnpack/pack.h",
582 "src/xnnpack/operator.h",
583 "src/xnnpack/compute.h",
584]
585
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700586LOGGING_COPTS = select({
587 # No logging in optimized mode
588 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
589 # Full logging in debug mode
590 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
591 # Error-only logging in default (fastbuild) mode
592 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
593})
594
595LOGGING_HDRS = [
596 "src/xnnpack/log.h",
597]
598
Marat Dukhan08c4a432019-10-03 09:29:21 -0700599xnnpack_cc_library(
600 name = "scalar_ukernels",
601 srcs = SCALAR_UKERNELS,
602 hdrs = INTERNAL_HDRS,
603 aarch32_copts = ["-marm"],
604 copts = xnnpack_std_copts(),
605 deps = [
606 "@FP16",
607 "@FXdiv",
608 ],
609)
610
611xnnpack_cc_library(
612 name = "psimd_ukernels",
613 srcs = PSIMD_UKERNELS,
614 hdrs = INTERNAL_HDRS,
615 aarch32_copts = [
616 "-marm",
617 "-mfpu=neon",
618 ],
619 copts = xnnpack_std_copts(),
620 optimized_copts = [
621 "-O3",
622 "-ffast-math",
623 ],
624 deps = [
625 "@FP16",
626 "@psimd",
627 ],
628)
629
630xnnpack_cc_library(
631 name = "neon_ukernels",
632 hdrs = INTERNAL_HDRS,
633 aarch32_copts = [
634 "-marm",
635 "-mfpu=neon",
636 ],
637 aarch32_srcs = NEON_UKERNELS,
638 aarch64_srcs = NEON_UKERNELS,
639 copts = xnnpack_std_copts(),
640 deps = ["@FP16"],
641)
642
643xnnpack_cc_library(
644 name = "neonfma_ukernels",
645 hdrs = INTERNAL_HDRS,
646 aarch32_copts = [
647 "-marm",
648 "-mfpu=neon-vfpv4",
649 ],
650 aarch32_srcs = NEONFMA_UKERNELS,
651 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
652 copts = xnnpack_std_copts(),
653 deps = ["@FP16"],
654)
655
656xnnpack_cc_library(
657 name = "neonfp16arith_ukernels",
658 hdrs = INTERNAL_HDRS,
659 aarch64_copts = ["-march=armv8.2-a+fp16"],
660 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
661 copts = xnnpack_std_copts(),
662 deps = ["@FP16"],
663)
664
665xnnpack_cc_library(
666 name = "sse2_ukernels",
667 hdrs = INTERNAL_HDRS,
668 copts = xnnpack_std_copts(),
669 x86_copts = ["-msse2"],
670 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
671 deps = ["@FP16"],
672)
673
674xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800675 name = "sse41_ukernels",
676 hdrs = INTERNAL_HDRS,
677 copts = xnnpack_std_copts(),
678 x86_copts = ["-msse4.1"],
679 x86_srcs = SSE41_UKERNELS,
680 deps = ["@FP16"],
681)
682
683xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700684 name = "avx_ukernels",
685 hdrs = INTERNAL_HDRS,
686 copts = xnnpack_std_copts(),
687 x86_copts = ["-mavx"],
688 x86_srcs = AVX_UKERNELS,
689 deps = ["@FP16"],
690)
691
692xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700693 name = "avx2_ukernels",
694 hdrs = INTERNAL_HDRS,
695 copts = xnnpack_std_copts(),
696 x86_copts = [
697 "-mfma",
698 "-mavx2",
699 ],
700 x86_srcs = AVX2_UKERNELS,
701 deps = ["@FP16"],
702)
703
704xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700705 name = "avx512f_ukernels",
706 hdrs = INTERNAL_HDRS,
707 copts = xnnpack_std_copts(),
708 x86_copts = ["-mavx512f"],
709 x86_srcs = AVX512F_UKERNELS,
710 deps = ["@FP16"],
711)
712
713xnnpack_cc_library(
714 name = "asm_ukernels",
715 hdrs = ["src/xnnpack/assembly.h"],
716 aarch32_srcs = AARCH32_ASM_UKERNELS,
717 aarch64_srcs = AARCH64_ASM_UKERNELS,
718)
719
720xnnpack_aggregate_library(
721 name = "ukernels",
722 aarch32_deps = [
723 ":psimd_ukernels",
724 ":neon_ukernels",
725 ":neonfma_ukernels",
726 ":asm_ukernels",
727 ],
728 aarch64_deps = [
729 ":psimd_ukernels",
730 ":neon_ukernels",
731 ":neonfma_ukernels",
732 ":neonfp16arith_ukernels",
733 ":asm_ukernels",
734 ],
735 generic_deps = [":scalar_ukernels"],
736 wasmsimd_deps = [
737 ":psimd_ukernels",
738 ],
739 x86_deps = [
740 ":psimd_ukernels",
741 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800742 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700743 ":avx_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700744 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700745 ":avx512f_ukernels",
746 ],
747)
748
749xnnpack_cc_library(
750 name = "im2col",
751 srcs = ["src/im2col.c"],
752 hdrs = [
753 "src/xnnpack/common.h",
754 "src/xnnpack/im2col.h",
755 ],
756 copts = xnnpack_std_copts(),
757)
758
759xnnpack_cc_library(
760 name = "indirection",
761 srcs = ["src/indirection.c"],
762 hdrs = INTERNAL_HDRS,
763 copts = xnnpack_std_copts(),
764 deps = [
765 "@FP16",
766 "@FXdiv",
767 "@pthreadpool",
768 ],
769)
770
771xnnpack_cc_library(
772 name = "operator_run",
773 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700774 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
775 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700776 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
777 "-Wno-vla",
778 ],
779 deps = [
780 "@FP16",
781 "@FXdiv",
782 "@clog",
783 "@pthreadpool",
784 ],
785)
786
787cc_library(
788 name = "enable_assembly",
789 defines = select({
790 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
791 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -0700792 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700793 }),
794)
795
Marat Dukhancf056b22019-10-07 10:26:29 -0700796xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700797 name = "operators",
798 srcs = OPERATOR_SRCS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700799 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -0700800 ],
801 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700802 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700803 "-Isrc",
804 "-Iinclude",
805 ] + select({
806 ":debug_build": [],
807 "//conditions:default": xnnpack_min_size_copts(),
808 }),
Marat Dukhancf056b22019-10-07 10:26:29 -0700809 wasm_srcs = ["src/wasm-stubs.c"],
810 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700811 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700812 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700813 "@FP16",
814 "@FXdiv",
815 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700816 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700817 ],
818)
819
820cc_library(
821 name = "XNNPACK",
822 srcs = [
823 "src/init.c",
824 ],
825 copts = xnnpack_std_copts() + LOGGING_COPTS + [
826 "-Isrc",
827 "-Iinclude",
828 ] + select({
829 ":debug_build": [],
830 "//conditions:default": xnnpack_min_size_copts(),
831 }),
832 includes = ["include"],
833 linkstatic = True,
834 textual_hdrs = ["include/xnnpack.h"],
835 visibility = xnnpack_visibility(),
836 deps = [
837 ":enable_assembly",
838 ":ukernels",
839 ":operator_run",
840 ":operators",
841 "@clog",
842 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -0700843 ] + select({
844 ":emscripten": [],
845 "//conditions:default": ["@cpuinfo"],
846 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700847)
848
849cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700850 name = "xnnpack_operators_nhwc_f32",
851 srcs = [
852 "src/init.c",
853 ],
854 copts = xnnpack_std_copts() + LOGGING_COPTS + [
855 "-Isrc",
856 "-Iinclude",
857 ] + select({
858 ":debug_build": [],
859 "//conditions:default": xnnpack_min_size_copts(),
860 }),
861 defines = [
862 "XNN_NO_Q8_OPERATORS",
863 "XNN_NO_U8_OPERATORS",
864 "XNN_NO_X8_OPERATORS",
865 "XNN_NO_SPNCHW_OPERATORS",
866 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700867 includes = ["include"],
868 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700869 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700870 visibility = xnnpack_visibility(),
871 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700872 ":enable_assembly",
873 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700874 ":operator_run",
875 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700876 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700877 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700878 ] + select({
879 ":emscripten": [],
880 "//conditions:default": ["@cpuinfo"],
881 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700882)
883
Marat Dukhancf056b22019-10-07 10:26:29 -0700884xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700885 name = "bench_utils",
886 srcs = ["bench/utils.cc"],
887 hdrs = ["bench/utils.h"],
888 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -0800889 deps = [
890 "@com_google_benchmark//:benchmark",
891 "@cpuinfo",
892 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700893)
894
Frank Barchard7e955972019-10-11 10:34:25 -0700895######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -0700896
897xnnpack_benchmark(
898 name = "q8_gemm_bench",
899 srcs = [
900 "bench/gemm.h",
901 "bench/q8-gemm.cc",
902 "src/xnnpack/AlignedAllocator.h",
903 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
904 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
905 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
906)
907
908xnnpack_benchmark(
909 name = "f16_gemm_bench",
910 srcs = [
911 "bench/f16-gemm.cc",
912 "bench/gemm.h",
913 "src/xnnpack/AlignedAllocator.h",
914 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
915 copts = ["-Wno-unused-function"],
916 deps = MICROKERNEL_BENCHMARK_DEPS,
917)
918
919xnnpack_benchmark(
920 name = "f32_igemm_bench",
921 srcs = [
922 "bench/f32-igemm.cc",
923 "bench/conv.h",
924 "src/xnnpack/AlignedAllocator.h",
925 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -0700926 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700927)
928
929xnnpack_benchmark(
930 name = "f32_conv_hwc_bench",
931 srcs = [
932 "bench/f32-conv-hwc.cc",
933 "bench/dconv.h",
934 "src/xnnpack/AlignedAllocator.h",
935 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
936 copts = ["-Wno-unused-function"],
937 deps = MICROKERNEL_BENCHMARK_DEPS,
938)
939
940xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -0700941 name = "f32_conv_hwc2spchw_bench",
942 srcs = [
943 "bench/f32-conv-hwc2spchw.cc",
944 "bench/dconv.h",
945 "src/xnnpack/AlignedAllocator.h",
946 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
947 copts = ["-Wno-unused-function"],
948 deps = MICROKERNEL_BENCHMARK_DEPS,
949)
950
951xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700952 name = "f32_dwconv_bench",
953 srcs = [
954 "bench/f32-dwconv.cc",
955 "bench/dwconv.h",
956 "src/xnnpack/AlignedAllocator.h",
957 ] + MICROKERNEL_BENCHMARK_HDRS,
958 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
959)
960
961xnnpack_benchmark(
962 name = "f32_dwconv_spchw_bench",
963 srcs = [
964 "bench/f32-dwconv-spchw.cc",
965 "bench/dwconv.h",
966 "src/xnnpack/AlignedAllocator.h",
967 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
968 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
969)
970
971xnnpack_benchmark(
972 name = "f32_gemm_bench",
973 srcs = [
974 "bench/f32-gemm.cc",
975 "bench/gemm.h",
976 "src/xnnpack/AlignedAllocator.h",
977 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
978 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -0700979 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700980)
981
982xnnpack_benchmark(
983 name = "f32_rmax_bench",
984 srcs = [
985 "bench/f32-rmax.cc",
986 "src/xnnpack/AlignedAllocator.h",
987 ] + MICROKERNEL_BENCHMARK_HDRS,
988 deps = MICROKERNEL_BENCHMARK_DEPS,
989)
990
991xnnpack_benchmark(
992 name = "f32_spmm_bench",
993 srcs = [
994 "bench/f32-spmm.cc",
995 "bench/gemm.h",
996 "src/xnnpack/AlignedAllocator.h",
997 ] + MICROKERNEL_BENCHMARK_HDRS,
998 copts = ["-Wno-unused-function"],
999 deps = MICROKERNEL_BENCHMARK_DEPS,
1000)
1001
1002xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07001003 name = "f32_softargmax_bench",
1004 srcs = [
1005 "bench/f32-softargmax.cc",
1006 ] + MICROKERNEL_BENCHMARK_HDRS,
1007 copts = ["-Wno-unused-function"],
1008 deps = MICROKERNEL_BENCHMARK_DEPS,
1009)
1010
1011xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001012 name = "f32_im2col_gemm_bench",
1013 srcs = [
1014 "bench/f32-im2col-gemm.cc",
1015 "bench/conv.h",
1016 "src/xnnpack/AlignedAllocator.h",
1017 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1018 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
1019)
1020
1021########################### Benchmarks for operators ###########################
1022
1023xnnpack_benchmark(
1024 name = "add_bench",
1025 srcs = ["bench/add.cc"],
1026 deps = OPERATOR_BENCHMARK_DEPS,
1027)
1028
1029xnnpack_benchmark(
1030 name = "average_pooling_bench",
1031 srcs = ["bench/average-pooling.cc"],
1032 deps = OPERATOR_BENCHMARK_DEPS,
1033)
1034
1035xnnpack_benchmark(
1036 name = "channel_shuffle_bench",
1037 srcs = ["bench/channel-shuffle.cc"],
1038 deps = OPERATOR_BENCHMARK_DEPS,
1039)
1040
1041xnnpack_benchmark(
1042 name = "convolution_bench",
1043 srcs = ["bench/convolution.cc"],
1044 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1045 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1046)
1047
1048xnnpack_benchmark(
1049 name = "deconvolution_bench",
1050 srcs = ["bench/deconvolution.cc"],
1051 copts = xnnpack_optional_tflite_copts(),
1052 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1053)
1054
1055xnnpack_benchmark(
1056 name = "global_average_pooling_bench",
1057 srcs = ["bench/global-average-pooling.cc"],
1058 deps = OPERATOR_BENCHMARK_DEPS,
1059)
1060
1061xnnpack_benchmark(
1062 name = "max_pooling_bench",
1063 srcs = ["bench/max-pooling.cc"],
1064 deps = OPERATOR_BENCHMARK_DEPS,
1065)
1066
1067xnnpack_benchmark(
1068 name = "sigmoid_bench",
1069 srcs = ["bench/sigmoid.cc"],
1070 deps = OPERATOR_BENCHMARK_DEPS,
1071)
1072
1073xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07001074 name = "prelu_bench",
1075 srcs = ["bench/prelu.cc"],
1076 copts = xnnpack_optional_tflite_copts(),
1077 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1078)
1079
1080xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001081 name = "softargmax_bench",
1082 srcs = ["bench/softargmax.cc"],
1083 deps = OPERATOR_BENCHMARK_DEPS,
1084)
1085
Marat Dukhanc068bb62019-10-04 13:24:39 -07001086############################# End-to-end benchmarks ############################
1087
1088cc_library(
1089 name = "mobilenet_v1",
1090 srcs = ["models/mobilenet-v1.cc"],
1091 hdrs = ["models/models.h"],
1092 linkstatic = True,
1093 deps = [
1094 ":XNNPACK",
1095 "@pthreadpool",
1096 ],
1097)
1098
1099cc_library(
1100 name = "mobilenet_v2",
1101 srcs = ["models/mobilenet-v2.cc"],
1102 hdrs = ["models/models.h"],
1103 linkstatic = True,
1104 deps = [
1105 ":XNNPACK",
1106 "@pthreadpool",
1107 ],
1108)
1109
1110xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07001111 name = "f32_dwconv_e2e_bench",
1112 srcs = ["bench/f32-dwconv-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1113 copts = ["-Wno-unused-function"],
1114 deps = MICROKERNEL_BENCHMARK_DEPS + [
1115 ":XNNPACK",
1116 ":mobilenet_v1",
1117 ":mobilenet_v2",
1118 ],
1119)
1120
1121xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07001122 name = "f32_gemm_e2e_bench",
1123 srcs = ["bench/f32-gemm-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1124 copts = ["-Wno-unused-function"],
1125 deps = MICROKERNEL_BENCHMARK_DEPS + [
1126 ":XNNPACK",
1127 ":mobilenet_v1",
1128 ":mobilenet_v2",
1129 ],
1130)
1131
1132xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07001133 name = "end2end_bench",
1134 srcs = ["bench/end2end.cc"],
1135 deps = [
1136 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07001137 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07001138 ":mobilenet_v1",
1139 ":mobilenet_v2",
1140 "@pthreadpool",
1141 ],
1142)
1143
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001144#################### Accuracy evaluation for math functions ####################
1145
1146xnnpack_benchmark(
1147 name = "f32_exp_eval",
1148 srcs = [
1149 "eval/f32-exp.cc",
1150 "src/xnnpack/AlignedAllocator.h",
1151 ] + ACCURACY_EVAL_HDRS,
1152 deps = ACCURACY_EVAL_DEPS,
1153)
1154
Marat Dukhan515c9772019-10-17 18:07:57 -07001155xnnpack_benchmark(
1156 name = "f32_expminus_eval",
1157 srcs = [
1158 "eval/f32-expminus.cc",
1159 "src/xnnpack/AlignedAllocator.h",
1160 ] + ACCURACY_EVAL_HDRS,
1161 deps = ACCURACY_EVAL_DEPS,
1162)
1163
Marat Dukhan98ba4412019-10-23 02:14:28 -07001164xnnpack_benchmark(
1165 name = "f32_extexp_eval",
1166 srcs = [
1167 "eval/f32-extexp.cc",
1168 "src/xnnpack/AlignedAllocator.h",
1169 ] + ACCURACY_EVAL_HDRS,
1170 deps = ACCURACY_EVAL_DEPS,
1171)
1172
Marat Dukhan08c4a432019-10-03 09:29:21 -07001173######################### Unit tests for micro-kernels #########################
1174
1175xnnpack_unit_test(
1176 name = "f16_gemm_test",
1177 srcs = [
1178 "test/f16-gemm.cc",
1179 "test/gemm-microkernel-tester.h",
1180 "src/xnnpack/AlignedAllocator.h",
1181 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1182 deps = MICROKERNEL_TEST_DEPS,
1183)
1184
1185xnnpack_unit_test(
1186 name = "f32_argmaxpool_test",
1187 srcs = [
1188 "test/f32-argmaxpool.cc",
1189 "test/argmaxpool-microkernel-tester.h",
1190 "src/xnnpack/AlignedAllocator.h",
1191 ] + MICROKERNEL_TEST_HDRS,
1192 deps = MICROKERNEL_TEST_DEPS,
1193)
1194
1195xnnpack_unit_test(
1196 name = "f32_avgpool_test",
1197 srcs = [
1198 "test/f32-avgpool.cc",
1199 "test/avgpool-microkernel-tester.h",
1200 "src/xnnpack/AlignedAllocator.h",
1201 ] + MICROKERNEL_TEST_HDRS,
1202 deps = MICROKERNEL_TEST_DEPS,
1203)
1204
1205xnnpack_unit_test(
1206 name = "f32_clamp_test",
1207 srcs = [
1208 "test/f32-clamp.cc",
1209 "test/clamp-microkernel-tester.h",
1210 ] + MICROKERNEL_TEST_HDRS,
1211 deps = MICROKERNEL_TEST_DEPS,
1212)
1213
1214xnnpack_unit_test(
1215 name = "f32_igemm_test",
1216 srcs = [
1217 "test/f32-igemm.cc",
1218 "test/gemm-microkernel-tester.h",
1219 "src/xnnpack/AlignedAllocator.h",
1220 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1221 deps = MICROKERNEL_TEST_DEPS,
1222)
1223
1224xnnpack_unit_test(
1225 name = "f32_conv_hwc_test",
1226 srcs = [
1227 "test/f32-conv-hwc.cc",
1228 "test/conv-hwc-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_conv_hwc2spchw_test",
1236 srcs = [
1237 "test/f32-conv-hwc2spchw.cc",
1238 "test/conv-hwc2spchw-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_dwconv_test",
1246 srcs = [
1247 "test/f32-dwconv.cc",
1248 "test/dwconv-microkernel-tester.h",
1249 "src/xnnpack/AlignedAllocator.h",
1250 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1251 deps = MICROKERNEL_TEST_DEPS,
1252)
1253
1254xnnpack_unit_test(
1255 name = "f32_dwconv_spchw_test",
1256 srcs = [
1257 "test/f32-dwconv-spchw.cc",
1258 "test/dwconv-spchw-microkernel-tester.h",
1259 "src/xnnpack/AlignedAllocator.h",
1260 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1261 deps = MICROKERNEL_TEST_DEPS,
1262)
1263
1264xnnpack_unit_test(
1265 name = "f32_gavgpool_test",
1266 srcs = [
1267 "test/f32-gavgpool.cc",
1268 "test/gavgpool-microkernel-tester.h",
1269 "src/xnnpack/AlignedAllocator.h",
1270 ] + MICROKERNEL_TEST_HDRS,
1271 deps = MICROKERNEL_TEST_DEPS,
1272)
1273
1274xnnpack_unit_test(
1275 name = "f32_gavgpool_spchw_test",
1276 srcs = [
1277 "test/f32-gavgpool-spchw.cc",
1278 "test/gavgpool-spchw-microkernel-tester.h",
1279 "src/xnnpack/AlignedAllocator.h",
1280 ] + MICROKERNEL_TEST_HDRS,
1281 deps = MICROKERNEL_TEST_DEPS,
1282)
1283
1284xnnpack_unit_test(
1285 name = "f32_gemm_test",
1286 srcs = [
1287 "test/f32-gemm.cc",
1288 "test/gemm-microkernel-tester.h",
1289 "src/xnnpack/AlignedAllocator.h",
1290 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1291 deps = MICROKERNEL_TEST_DEPS,
1292)
1293
1294xnnpack_unit_test(
1295 name = "f32_gemminc_test",
1296 srcs = [
1297 "test/f32-gemminc.cc",
1298 "test/gemm-microkernel-tester.h",
1299 "src/xnnpack/AlignedAllocator.h",
1300 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1301 deps = MICROKERNEL_TEST_DEPS,
1302)
1303
1304xnnpack_unit_test(
1305 name = "f32_hswish_test",
1306 srcs = [
1307 "test/f32-hswish.cc",
1308 "test/hswish-microkernel-tester.h",
1309 ] + MICROKERNEL_TEST_HDRS,
1310 deps = MICROKERNEL_TEST_DEPS,
1311)
1312
1313xnnpack_unit_test(
1314 name = "f32_maxpool_test",
1315 srcs = [
1316 "test/f32-maxpool.cc",
1317 "test/maxpool-microkernel-tester.h",
1318 ] + MICROKERNEL_TEST_HDRS,
1319 deps = MICROKERNEL_TEST_DEPS,
1320)
1321
1322xnnpack_unit_test(
1323 name = "f32_pavgpool_test",
1324 srcs = [
1325 "test/f32-pavgpool.cc",
1326 "test/avgpool-microkernel-tester.h",
1327 "src/xnnpack/AlignedAllocator.h",
1328 ] + MICROKERNEL_TEST_HDRS,
1329 deps = MICROKERNEL_TEST_DEPS,
1330)
1331
1332xnnpack_unit_test(
1333 name = "f32_ppmm_test",
1334 srcs = [
1335 "test/f32-ppmm.cc",
1336 "test/gemm-microkernel-tester.h",
1337 "src/xnnpack/AlignedAllocator.h",
1338 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1339 deps = MICROKERNEL_TEST_DEPS,
1340)
1341
1342xnnpack_unit_test(
1343 name = "f32_prelu_test",
1344 srcs = [
1345 "test/f32-prelu.cc",
1346 "test/prelu-microkernel-tester.h",
1347 "src/xnnpack/AlignedAllocator.h",
1348 ] + MICROKERNEL_TEST_HDRS,
1349 deps = MICROKERNEL_TEST_DEPS,
1350)
1351
1352xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001353 name = "f32_raddexpminusmax_test",
1354 srcs = [
1355 "test/f32-raddexpminusmax.cc",
1356 "test/raddexpminusmax-microkernel-tester.h",
1357 ] + MICROKERNEL_TEST_HDRS,
1358 deps = MICROKERNEL_TEST_DEPS,
1359)
1360
1361xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001362 name = "f32_raddextexp_test",
1363 srcs = [
1364 "test/f32-raddextexp.cc",
1365 "test/raddextexp-microkernel-tester.h",
1366 ] + MICROKERNEL_TEST_HDRS,
1367 deps = MICROKERNEL_TEST_DEPS,
1368)
1369
1370xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001371 name = "f32_raddstoreexpminusmax_test",
1372 srcs = [
1373 "test/f32-raddstoreexpminusmax.cc",
1374 "test/raddstoreexpminusmax-microkernel-tester.h",
1375 ] + MICROKERNEL_TEST_HDRS,
1376 deps = MICROKERNEL_TEST_DEPS,
1377)
1378
1379xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001380 name = "f32_rmax_test",
1381 srcs = [
1382 "test/f32-rmax.cc",
1383 "test/rmax-microkernel-tester.h",
1384 ] + MICROKERNEL_TEST_HDRS,
1385 deps = MICROKERNEL_TEST_DEPS,
1386)
1387
1388xnnpack_unit_test(
1389 name = "f32_spmm_test",
1390 srcs = [
1391 "test/f32-spmm.cc",
1392 "test/spmm-microkernel-tester.h",
1393 "src/xnnpack/AlignedAllocator.h",
1394 ] + MICROKERNEL_TEST_HDRS,
1395 deps = MICROKERNEL_TEST_DEPS,
1396)
1397
1398xnnpack_unit_test(
1399 name = "f32_vadd_test",
1400 srcs = [
1401 "test/f32-vadd.cc",
1402 "test/vadd-microkernel-tester.h",
1403 ] + MICROKERNEL_TEST_HDRS,
1404 deps = MICROKERNEL_TEST_DEPS,
1405)
1406
1407xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001408 name = "f32_vmul_test",
1409 srcs = [
1410 "test/f32-vmul.cc",
1411 "test/vmul-microkernel-tester.h",
1412 ] + MICROKERNEL_TEST_HDRS,
1413 deps = MICROKERNEL_TEST_DEPS,
1414)
1415
1416xnnpack_unit_test(
1417 name = "f32_vmulcaddc_test",
1418 srcs = [
1419 "test/f32-vmulcaddc.cc",
1420 "test/vmulcaddc-microkernel-tester.h",
1421 "src/xnnpack/AlignedAllocator.h",
1422 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1423 deps = MICROKERNEL_TEST_DEPS,
1424)
1425
1426xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001427 name = "f32_vscale_test",
1428 srcs = [
1429 "test/f32-vscale.cc",
1430 "test/vscale-microkernel-tester.h",
1431 ] + MICROKERNEL_TEST_HDRS,
1432 deps = MICROKERNEL_TEST_DEPS,
1433)
1434
1435xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001436 name = "f32_vscaleexpminusmax_test",
1437 srcs = [
1438 "test/f32-vscaleexpminusmax.cc",
1439 "test/vscaleexpminusmax-microkernel-tester.h",
1440 ] + MICROKERNEL_TEST_HDRS,
1441 deps = MICROKERNEL_TEST_DEPS,
1442)
1443
1444xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001445 name = "f32_vscaleextexp_test",
1446 srcs = [
1447 "test/f32-vscaleextexp.cc",
1448 "test/vscaleextexp-microkernel-tester.h",
1449 ] + MICROKERNEL_TEST_HDRS,
1450 deps = MICROKERNEL_TEST_DEPS,
1451)
1452
1453xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001454 name = "f32_vsub_test",
1455 srcs = [
1456 "test/f32-vsub.cc",
1457 "test/vsub-microkernel-tester.h",
1458 ] + MICROKERNEL_TEST_HDRS,
1459 deps = MICROKERNEL_TEST_DEPS,
1460)
1461
1462xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001463 name = "q8_avgpool_test",
1464 srcs = [
1465 "test/q8-avgpool.cc",
1466 "test/avgpool-microkernel-tester.h",
1467 "src/xnnpack/AlignedAllocator.h",
1468 ] + MICROKERNEL_TEST_HDRS,
1469 deps = MICROKERNEL_TEST_DEPS,
1470)
1471
1472xnnpack_unit_test(
1473 name = "q8_igemm_test",
1474 srcs = [
1475 "test/q8-igemm.cc",
1476 "test/gemm-microkernel-tester.h",
1477 "src/xnnpack/AlignedAllocator.h",
1478 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1479 deps = MICROKERNEL_TEST_DEPS,
1480)
1481
1482xnnpack_unit_test(
1483 name = "q8_dwconv_test",
1484 srcs = [
1485 "test/q8-dwconv.cc",
1486 "test/dwconv-microkernel-tester.h",
1487 "src/xnnpack/AlignedAllocator.h",
1488 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1489 deps = MICROKERNEL_TEST_DEPS,
1490)
1491
1492xnnpack_unit_test(
1493 name = "q8_gavgpool_test",
1494 srcs = [
1495 "test/q8-gavgpool.cc",
1496 "test/gavgpool-microkernel-tester.h",
1497 "src/xnnpack/AlignedAllocator.h",
1498 ] + MICROKERNEL_TEST_HDRS,
1499 deps = MICROKERNEL_TEST_DEPS,
1500)
1501
1502xnnpack_unit_test(
1503 name = "q8_gemm_test",
1504 srcs = [
1505 "test/q8-gemm.cc",
1506 "test/gemm-microkernel-tester.h",
1507 "src/xnnpack/AlignedAllocator.h",
1508 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1509 deps = MICROKERNEL_TEST_DEPS,
1510)
1511
1512xnnpack_unit_test(
1513 name = "q8_vadd_test",
1514 srcs = [
1515 "test/q8-vadd.cc",
1516 "test/vadd-microkernel-tester.h",
1517 ] + MICROKERNEL_TEST_HDRS,
1518 deps = MICROKERNEL_TEST_DEPS,
1519)
1520
1521xnnpack_unit_test(
1522 name = "u8_clamp_test",
1523 srcs = [
1524 "test/u8-clamp.cc",
1525 "test/clamp-microkernel-tester.h",
1526 ] + MICROKERNEL_TEST_HDRS,
1527 deps = MICROKERNEL_TEST_DEPS,
1528)
1529
1530xnnpack_unit_test(
1531 name = "u8_lut32norm_test",
1532 srcs = [
1533 "test/u8-lut32norm.cc",
1534 "test/lut-norm-microkernel-tester.h",
1535 ] + MICROKERNEL_TEST_HDRS,
1536 deps = MICROKERNEL_TEST_DEPS,
1537)
1538
1539xnnpack_unit_test(
1540 name = "u8_maxpool_test",
1541 srcs = [
1542 "test/u8-maxpool.cc",
1543 "test/maxpool-microkernel-tester.h",
1544 ] + MICROKERNEL_TEST_HDRS,
1545 deps = MICROKERNEL_TEST_DEPS,
1546)
1547
1548xnnpack_unit_test(
1549 name = "u8_rmax_test",
1550 srcs = [
1551 "test/u8-rmax.cc",
1552 "test/rmax-microkernel-tester.h",
1553 ] + MICROKERNEL_TEST_HDRS,
1554 deps = MICROKERNEL_TEST_DEPS,
1555)
1556
1557xnnpack_unit_test(
1558 name = "x32_packx_test",
1559 srcs = [
1560 "test/x32-packx.cc",
1561 "test/pack-microkernel-tester.h",
1562 "src/xnnpack/AlignedAllocator.h",
1563 ] + MICROKERNEL_TEST_HDRS,
1564 deps = MICROKERNEL_TEST_DEPS,
1565)
1566
1567xnnpack_unit_test(
1568 name = "x32_pad_test",
1569 srcs = [
1570 "test/x32-pad.cc",
1571 "test/pad-microkernel-tester.h",
1572 ] + MICROKERNEL_TEST_HDRS,
1573 deps = MICROKERNEL_TEST_DEPS,
1574)
1575
1576xnnpack_unit_test(
1577 name = "x32_unpool_test",
1578 srcs = [
1579 "test/x32-unpool.cc",
1580 "test/unpool-microkernel-tester.h",
1581 ] + MICROKERNEL_TEST_HDRS,
1582 deps = MICROKERNEL_TEST_DEPS,
1583)
1584
1585xnnpack_unit_test(
1586 name = "x32_zip_test",
1587 srcs = [
1588 "test/x32-zip.cc",
1589 "test/zip-microkernel-tester.h",
1590 ] + MICROKERNEL_TEST_HDRS,
1591 deps = MICROKERNEL_TEST_DEPS,
1592)
1593
1594xnnpack_unit_test(
1595 name = "x8_lut_test",
1596 srcs = [
1597 "test/x8-lut.cc",
1598 "test/lut-microkernel-tester.h",
1599 ] + MICROKERNEL_TEST_HDRS,
1600 deps = MICROKERNEL_TEST_DEPS,
1601)
1602
1603xnnpack_unit_test(
1604 name = "x8_zip_test",
1605 srcs = [
1606 "test/x8-zip.cc",
1607 "test/zip-microkernel-tester.h",
1608 ] + MICROKERNEL_TEST_HDRS,
1609 deps = MICROKERNEL_TEST_DEPS,
1610)
1611
1612########################### Size test for the library ##########################
1613
1614xnnpack_binary(
1615 name = "size_test",
1616 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001617 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001618)
1619
1620########################### Unit tests for operators ###########################
1621
1622xnnpack_unit_test(
1623 name = "add_test",
1624 srcs = [
1625 "test/add.cc",
1626 "test/add-operator-tester.h",
1627 ],
1628 deps = OPERATOR_TEST_DEPS,
1629)
1630
1631xnnpack_unit_test(
1632 name = "argmax_pooling_test",
1633 srcs = [
1634 "test/argmax-pooling.cc",
1635 "test/argmax-pooling-operator-tester.h",
1636 ] + OPERATOR_TEST_PARAMS_HDRS,
1637 deps = OPERATOR_TEST_DEPS,
1638)
1639
1640xnnpack_unit_test(
1641 name = "average_pooling_test",
1642 srcs = [
1643 "test/average-pooling.cc",
1644 "test/average-pooling-operator-tester.h",
1645 ] + OPERATOR_TEST_PARAMS_HDRS,
1646 deps = OPERATOR_TEST_DEPS,
1647)
1648
1649xnnpack_unit_test(
1650 name = "channel_pad_test",
1651 srcs = [
1652 "test/channel-pad.cc",
1653 "test/channel-pad-operator-tester.h",
1654 ] + OPERATOR_TEST_PARAMS_HDRS,
1655 deps = OPERATOR_TEST_DEPS,
1656)
1657
1658xnnpack_unit_test(
1659 name = "channel_shuffle_test",
1660 srcs = [
1661 "test/channel-shuffle.cc",
1662 "test/channel-shuffle-operator-tester.h",
1663 ],
1664 deps = OPERATOR_TEST_DEPS,
1665)
1666
1667xnnpack_unit_test(
1668 name = "clamp_test",
1669 srcs = [
1670 "test/clamp.cc",
1671 "test/clamp-operator-tester.h",
1672 ],
1673 deps = OPERATOR_TEST_DEPS,
1674)
1675
1676xnnpack_unit_test(
1677 name = "convolution_test",
1678 srcs = [
1679 "test/convolution.cc",
1680 "test/convolution-operator-tester.h",
1681 ],
1682 deps = OPERATOR_TEST_DEPS,
1683)
1684
1685xnnpack_unit_test(
1686 name = "convolution_spnchw_test",
1687 srcs = [
1688 "test/convolution-spnchw.cc",
1689 "test/convolution-spnchw-operator-tester.h",
1690 ],
1691 deps = OPERATOR_TEST_DEPS,
1692)
1693
1694xnnpack_unit_test(
1695 name = "deconvolution_test",
1696 srcs = [
1697 "test/deconvolution.cc",
1698 "test/deconvolution-operator-tester.h",
1699 ] + OPERATOR_TEST_PARAMS_HDRS,
1700 deps = OPERATOR_TEST_DEPS,
1701)
1702
1703xnnpack_unit_test(
1704 name = "fully_connected_test",
1705 srcs = [
1706 "test/fully-connected.cc",
1707 "test/fully-connected-operator-tester.h",
1708 ],
1709 deps = OPERATOR_TEST_DEPS,
1710)
1711
1712xnnpack_unit_test(
1713 name = "global_average_pooling_test",
1714 srcs = [
1715 "test/global-average-pooling.cc",
1716 "test/global-average-pooling-operator-tester.h",
1717 ] + OPERATOR_TEST_PARAMS_HDRS,
1718 deps = OPERATOR_TEST_DEPS,
1719)
1720
1721xnnpack_unit_test(
1722 name = "global_average_pooling_spnchw_test",
1723 srcs = [
1724 "test/global-average-pooling-spnchw.cc",
1725 "test/global-average-pooling-spnchw-operator-tester.h",
1726 ],
1727 deps = OPERATOR_TEST_DEPS,
1728)
1729
1730xnnpack_unit_test(
1731 name = "hardswish_test",
1732 srcs = [
1733 "test/hardswish.cc",
1734 "test/hardswish-operator-tester.h",
1735 ],
1736 deps = OPERATOR_TEST_DEPS,
1737)
1738
1739xnnpack_unit_test(
1740 name = "leaky_relu_test",
1741 srcs = [
1742 "test/leaky-relu.cc",
1743 "test/leaky-relu-operator-tester.h",
1744 ],
1745 deps = OPERATOR_TEST_DEPS,
1746)
1747
1748xnnpack_unit_test(
1749 name = "max_pooling_test",
1750 srcs = [
1751 "test/max-pooling.cc",
1752 "test/max-pooling-operator-tester.h",
1753 ] + OPERATOR_TEST_PARAMS_HDRS,
1754 deps = OPERATOR_TEST_DEPS,
1755)
1756
1757xnnpack_unit_test(
1758 name = "prelu_test",
1759 srcs = [
1760 "test/prelu.cc",
1761 "test/prelu-operator-tester.h",
1762 ] + OPERATOR_TEST_PARAMS_HDRS,
1763 deps = OPERATOR_TEST_DEPS,
1764)
1765
1766xnnpack_unit_test(
1767 name = "sigmoid_test",
1768 srcs = [
1769 "test/sigmoid.cc",
1770 "test/sigmoid-operator-tester.h",
1771 ],
1772 deps = OPERATOR_TEST_DEPS,
1773)
1774
1775xnnpack_unit_test(
1776 name = "softargmax_test",
1777 srcs = [
1778 "test/softargmax.cc",
1779 "test/softargmax-operator-tester.h",
1780 ],
1781 deps = OPERATOR_TEST_DEPS,
1782)
1783
1784xnnpack_unit_test(
1785 name = "unpooling_test",
1786 srcs = [
1787 "test/unpooling.cc",
1788 "test/unpooling-operator-tester.h",
1789 ],
1790 deps = OPERATOR_TEST_DEPS,
1791)
1792
1793############################# Build configurations #############################
1794
Marat Dukhanb8642352019-10-30 15:43:02 -07001795# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07001796config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07001797 name = "xnn_enable_assembly_explicit_true",
1798 define_values = {"xnn_enable_assembly": "true"},
1799)
1800
1801# Disables usage of assembly kernels.
1802config_setting(
1803 name = "xnn_enable_assembly_explicit_false",
1804 define_values = {"xnn_enable_assembly": "false"},
1805)
1806
1807# Builds with -c dbg
1808config_setting(
1809 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001810 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07001811 "compilation_mode": "dbg",
1812 },
1813)
1814
1815# Builds with -c opt
1816config_setting(
1817 name = "optimized_build",
1818 values = {
1819 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001820 },
1821)
1822
1823config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07001824 name = "linux_k8",
1825 values = {"cpu": "k8"},
1826)
1827
1828config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07001829 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07001830 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07001831)
1832
1833config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001834 name = "android",
1835 values = {"crosstool_top": "//external:android/crosstool"},
1836)
1837
1838config_setting(
1839 name = "android_armv7",
1840 values = {
1841 "crosstool_top": "//external:android/crosstool",
1842 "cpu": "armeabi-v7a",
1843 },
1844)
1845
1846config_setting(
1847 name = "android_arm64",
1848 values = {
1849 "crosstool_top": "//external:android/crosstool",
1850 "cpu": "arm64-v8a",
1851 },
1852)
1853
1854config_setting(
1855 name = "android_x86",
1856 values = {
1857 "crosstool_top": "//external:android/crosstool",
1858 "cpu": "x86",
1859 },
1860)
1861
1862config_setting(
1863 name = "android_x86_64",
1864 values = {
1865 "crosstool_top": "//external:android/crosstool",
1866 "cpu": "x86_64",
1867 },
1868)
1869
1870config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07001871 name = "macos_x86_64",
1872 values = {
1873 "apple_platform_type": "macos",
1874 "cpu": "darwin",
1875 },
1876)
1877
1878config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001879 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001880 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07001881)
1882
1883config_setting(
1884 name = "emscripten_wasm",
1885 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001886 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001887 "cpu": "wasm",
1888 },
1889)
1890
1891config_setting(
1892 name = "emscripten_wasmsimd",
1893 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001894 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001895 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07001896 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001897 },
1898)
1899
1900config_setting(
1901 name = "emscripten_asmjs",
1902 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001903 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001904 "cpu": "asmjs",
1905 },
1906)