blob: fd9afb7f1f84fbf30cfba87606417142e8129170 [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",
Marat Dukhan35dacfb2019-11-07 19:18:16 -080080 "src/f32-bilinear/scalar-c1.c",
81 "src/f32-bilinear/scalar-c2.c",
82 "src/f32-bilinear/scalar-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070083 "src/f32-clamp/scalar.c",
Erich Elsen563df5f2019-10-23 08:02:21 -070084 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070085 "src/f32-igemm/1x4-scalar.c",
86 "src/f32-igemm/2x4-scalar.c",
87 "src/f32-igemm/4x2-scalar.c",
88 "src/f32-igemm/4x4-scalar.c",
89 "src/f32-dwconv/up1x25-scalar.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -080090 "src/f32-dwconv/up1x25-scalar-acc2.c",
91 "src/f32-dwconv/up2x25-scalar.c",
92 "src/f32-dwconv/up2x25-scalar-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070093 "src/f32-dwconv/up1x4-scalar.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -080094 "src/f32-dwconv/up1x4-scalar-acc2.c",
95 "src/f32-dwconv/up2x4-scalar.c",
96 "src/f32-dwconv/up2x4-scalar-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070097 "src/f32-dwconv/up1x9-scalar.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -080098 "src/f32-dwconv/up1x9-scalar-acc2.c",
99 "src/f32-dwconv/up2x9-scalar.c",
100 "src/f32-dwconv/up2x9-scalar-acc2.c",
Erich Elsen0cc2c532019-10-15 04:44:18 -0700101 "src/f32-dwconv-spchw/3x3p1-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -0700102 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -0700103 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700104 "src/f32-gavgpool/mp7p7q-scalar.c",
105 "src/f32-gavgpool/up7-scalar.c",
106 "src/f32-gemm/1x4-scalar.c",
107 "src/f32-gemm/2x4-scalar.c",
108 "src/f32-gemm/4x2-scalar.c",
109 "src/f32-gemm/4x4-scalar.c",
110 "src/f32-gemminc/1x4-scalar.c",
111 "src/f32-gemminc/2x4-scalar.c",
112 "src/f32-gemminc/4x4-scalar.c",
113 "src/f32-hswish/scalar.c",
114 "src/f32-maxpool/9p8q-scalar.c",
115 "src/f32-pavgpool/mp9p8q-scalar.c",
116 "src/f32-pavgpool/up9-scalar.c",
117 "src/f32-ppmm/2x4-scalar.c",
118 "src/f32-ppmm/3x3-scalar.c",
119 "src/f32-ppmm/4x2-scalar.c",
120 "src/f32-ppmm/4x4-scalar.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800121 "src/f32-prelu/scalar-2x1.c",
122 "src/f32-prelu/scalar-2x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700123 "src/f32-rmax/scalar.c",
124 "src/f32-spmm/1x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700125 "src/f32-spmm/1x1-scalar.c",
126 "src/f32-spmm/2x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700127 "src/f32-spmm/2x1-scalar.c",
128 "src/f32-spmm/4x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700129 "src/f32-spmm/4x1-scalar.c",
130 "src/f32-spmm/8x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700131 "src/f32-spmm/8x1-scalar.c",
Erich Elsenc6afd9b2019-10-24 16:10:53 -0700132 "src/f32-spmm/8x2-scalar.c",
133 "src/f32-spmm/8x4-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700134 "src/f32-vadd/scalar.c",
135 "src/f32-vmul/scalar.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800136 "src/f32-vmulcaddc/c1-scalar-2x.c",
137 "src/f32-vmulcaddc/c2-scalar-2x.c",
138 "src/f32-vmulcaddc/c4-scalar-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700139 "src/f32-vsub/scalar.c",
140 "src/q8-avgpool/mp9p8q-scalar.c",
141 "src/q8-avgpool/up9-scalar.c",
142 "src/q8-igemm/2x2-scalar.c",
143 "src/q8-dwconv/up1x9-scalar.c",
144 "src/q8-gavgpool/mp7p7q-scalar.c",
145 "src/q8-gavgpool/up7-scalar.c",
146 "src/q8-gemm/2x2-scalar.c",
147 "src/q8-vadd/scalar.c",
148 "src/u8-clamp/scalar.c",
149 "src/u8-lut32norm/scalar.c",
150 "src/u8-maxpool/9p8q-scalar.c",
151 "src/u8-rmax/scalar.c",
152 "src/x32-packx/x2-scalar.c",
153 "src/x32-packx/x3-scalar.c",
154 "src/x32-packx/x4-scalar.c",
155 "src/x32-pad/x2-scalar.c",
156 "src/x32-unpool/scalar.c",
157 "src/x32-zip/x2-scalar.c",
158 "src/x32-zip/x3-scalar.c",
159 "src/x32-zip/x4-scalar.c",
160 "src/x32-zip/xm-scalar.c",
161 "src/x8-lut/scalar.c",
162 "src/x8-zip/x2-scalar.c",
163 "src/x8-zip/x3-scalar.c",
164 "src/x8-zip/x4-scalar.c",
165 "src/x8-zip/xm-scalar.c",
166]
167
168PSIMD_UKERNELS = [
169 "src/f32-argmaxpool/mp9p8q-psimd.c",
170 "src/f32-argmaxpool/up4-psimd.c",
171 "src/f32-argmaxpool/up9-psimd.c",
172 "src/f32-avgpool/mp9p8q-psimd.c",
173 "src/f32-avgpool/up9-psimd.c",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800174 "src/f32-bilinear/psimd-c4.c",
175 "src/f32-bilinear/psimd-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700176 "src/f32-clamp/psimd.c",
177 "src/f32-igemm/1x8-psimd-loadsplat.c",
178 "src/f32-igemm/1x8-psimd-splat.c",
179 "src/f32-igemm/1x8s4-psimd.c",
180 "src/f32-igemm/4x2c4-psimd.c",
181 "src/f32-igemm/4x8-psimd-loadsplat.c",
182 "src/f32-igemm/4x8-psimd-splat.c",
183 "src/f32-igemm/4x8s4-psimd.c",
184 "src/f32-igemm/6x8-psimd-loadsplat.c",
185 "src/f32-igemm/6x8-psimd-splat.c",
186 "src/f32-igemm/6x8s4-psimd.c",
187 "src/f32-dwconv/up4x25-psimd.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800188 "src/f32-dwconv/up4x25-psimd-acc2.c",
189 "src/f32-dwconv/up8x25-psimd.c",
190 "src/f32-dwconv/up8x25-psimd-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700191 "src/f32-dwconv/up4x4-psimd.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800192 "src/f32-dwconv/up4x4-psimd-acc2.c",
193 "src/f32-dwconv/up8x4-psimd.c",
194 "src/f32-dwconv/up8x4-psimd-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700195 "src/f32-dwconv/up4x9-psimd.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800196 "src/f32-dwconv/up4x9-psimd-acc2.c",
197 "src/f32-dwconv/up8x9-psimd.c",
198 "src/f32-dwconv/up8x9-psimd-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700199 "src/f32-gavgpool/mp7p7q-psimd.c",
200 "src/f32-gavgpool/up7-psimd.c",
201 "src/f32-gemm/1x8-psimd-loadsplat.c",
202 "src/f32-gemm/1x8-psimd-splat.c",
203 "src/f32-gemm/1x8s4-psimd.c",
204 "src/f32-gemm/4x8-psimd-loadsplat.c",
205 "src/f32-gemm/4x8-psimd-splat.c",
206 "src/f32-gemm/4x8s4-psimd.c",
207 "src/f32-gemm/6x8-psimd-loadsplat.c",
208 "src/f32-gemm/6x8-psimd-splat.c",
209 "src/f32-gemm/6x8s4-psimd.c",
210 "src/f32-gemminc/1x8-psimd-loadsplat.c",
211 "src/f32-gemminc/1x8-psimd-splat.c",
212 "src/f32-gemminc/1x8s4-psimd.c",
213 "src/f32-gemminc/4x8-psimd-loadsplat.c",
214 "src/f32-gemminc/4x8-psimd-splat.c",
215 "src/f32-gemminc/4x8s4-psimd.c",
216 "src/f32-gemminc/6x8-psimd-loadsplat.c",
217 "src/f32-gemminc/6x8-psimd-splat.c",
218 "src/f32-gemminc/6x8s4-psimd.c",
219 "src/f32-hswish/psimd.c",
220 "src/f32-maxpool/9p8q-psimd.c",
221 "src/f32-pavgpool/mp9p8q-psimd.c",
222 "src/f32-pavgpool/up9-psimd.c",
223 "src/f32-ppmm/4x8-psimd.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800224 "src/f32-prelu/psimd-2x4.c",
225 "src/f32-prelu/psimd-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700226 "src/f32-vadd/psimd.c",
227 "src/f32-vmul/psimd.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800228 "src/f32-vmulcaddc/c4-psimd-2x.c",
229 "src/f32-vmulcaddc/c8-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700230 "src/f32-vsub/psimd.c",
231 "src/x32-packx/x4-psimd.c",
232 "src/x32-pad/x2-psimd.c",
233 "src/x32-unpool/psimd.c",
234 "src/x32-zip/x2-psimd.c",
235 "src/x32-zip/x3-psimd.c",
236 "src/x32-zip/x4-psimd.c",
237 "src/x32-zip/xm-psimd.c",
238]
239
240# ISA-specific micro-kernels
241NEON_UKERNELS = [
242 "src/f32-avgpool/mp9p8q-neon.c",
243 "src/f32-avgpool/up9-neon.c",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800244 "src/f32-bilinear/neon-c4.c",
245 "src/f32-bilinear/neon-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700246 "src/f32-clamp/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700247 "src/f32-dwconv/up4x9-neon.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800248 "src/f32-dwconv/up4x9-neon-acc2.c",
249 "src/f32-dwconv/up8x9-neon.c",
250 "src/f32-dwconv/up8x9-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700251 "src/f32-gavgpool-spchw/neon-x4.c",
252 "src/f32-gavgpool/mp7p7q-neon.c",
253 "src/f32-gavgpool/up7-neon.c",
254 "src/f32-gemm/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700255 "src/f32-gemm/4x2-neon-ld64.c",
256 "src/f32-gemm/4x8-neon-ld128.c",
257 "src/f32-gemm/4x8-neon-ld64.c",
258 "src/f32-gemm/5x8-neon-ld64.c",
259 "src/f32-gemm/6x8-neon-ld64.c",
260 "src/f32-gemminc/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700261 "src/f32-gemminc/4x8-neon-ld128.c",
262 "src/f32-gemminc/4x8-neon-ld64.c",
263 "src/f32-gemminc/5x8-neon-ld64.c",
264 "src/f32-gemminc/6x8-neon-ld64.c",
265 "src/f32-hswish/neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800266 "src/f32-igemm/1x8-neon-ld64.c",
267 "src/f32-igemm/4x2-neon-ld64.c",
268 "src/f32-igemm/4x4-neon-ld64.c",
269 "src/f32-igemm/4x8-neon-ld128.c",
270 "src/f32-igemm/4x8-neon-ld64.c",
271 "src/f32-igemm/6x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700272 "src/f32-pavgpool/mp9p8q-neon.c",
273 "src/f32-pavgpool/up9-neon.c",
274 "src/f32-ppmm/4x8-neon.c",
275 "src/f32-ppmm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800276 "src/f32-prelu/neon-2x4.c",
277 "src/f32-prelu/neon-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700278 "src/f32-rmax/neon.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800279 "src/f32-vmulcaddc/c4-neon-2x.c",
280 "src/f32-vmulcaddc/c8-neon-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700281 "src/q8-avgpool/mp9p8q-neon.c",
282 "src/q8-avgpool/up9-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700283 "src/q8-dwconv/up8x9-neon.c",
284 "src/q8-gavgpool/mp7p7q-neon.c",
285 "src/q8-gavgpool/up7-neon.c",
286 "src/q8-gemm/4x8-neon.c",
287 "src/q8-gemm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800288 "src/q8-igemm/4x8-neon.c",
289 "src/q8-igemm/8x8-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700290 "src/q8-vadd/neon.c",
291 "src/u8-clamp/neon.c",
292 "src/u8-maxpool/9p8q-neon.c",
293 "src/u8-rmax/neon.c",
294 "src/x32-packx/x4-neon-st4.c",
295 "src/x32-pad/x2-neon.c",
296 "src/x32-zip/x2-neon.c",
297 "src/x32-zip/x3-neon.c",
298 "src/x32-zip/x4-neon.c",
299 "src/x32-zip/xm-neon.c",
300 "src/x8-zip/x2-neon.c",
301 "src/x8-zip/x3-neon.c",
302 "src/x8-zip/x4-neon.c",
303 "src/x8-zip/xm-neon.c",
304]
305
306NEONFMA_UKERNELS = [
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800307 "src/f32-bilinear/neonfma-c4.c",
308 "src/f32-bilinear/neonfma-c8.c",
Frank Barcharddb45b6a2019-10-09 16:42:45 -0700309 "src/f32-igemm/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700310 "src/f32-igemm/4x2-neonfma-ld64.c",
311 "src/f32-igemm/4x4-neonfma-ld64.c",
312 "src/f32-igemm/4x8-neonfma-ld128.c",
313 "src/f32-igemm/4x8-neonfma-ld64.c",
314 "src/f32-igemm/6x8-neonfma-ld64.c",
315 "src/f32-dwconv/up4x9-neonfma.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800316 "src/f32-dwconv/up4x9-neonfma-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700317 "src/f32-dwconv/up8x9-neonfma.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800318 "src/f32-dwconv/up8x9-neonfma-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700319 "src/f32-gemm/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700320 "src/f32-gemm/4x2-neonfma-ld64.c",
321 "src/f32-gemm/4x8-neonfma-ld128.c",
322 "src/f32-gemm/4x8-neonfma-ld64.c",
323 "src/f32-gemm/5x8-neonfma-ld64.c",
324 "src/f32-gemm/6x8-neonfma-ld64.c",
325 "src/f32-gemminc/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700326 "src/f32-gemminc/4x8-neonfma-ld128.c",
327 "src/f32-gemminc/4x8-neonfma-ld64.c",
328 "src/f32-gemminc/5x8-neonfma-ld64.c",
329 "src/f32-gemminc/6x8-neonfma-ld64.c",
330 "src/f32-hswish/neonfma.c",
331 "src/f32-ppmm/4x8-neonfma.c",
332 "src/f32-ppmm/8x8-neonfma.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800333 "src/f32-vmulcaddc/c4-neonfma-2x.c",
334 "src/f32-vmulcaddc/c8-neonfma-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700335]
336
337AARCH64_NEONFMA_UKERNELS = [
338 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
339 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
340 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
341 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
342 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
343 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
344 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
345 "src/f32-spmm/12x1-neonfma.c",
346 "src/f32-spmm/12x2-neonfma.c",
347 "src/f32-spmm/12x4-neonfma.c",
348 "src/f32-spmm/16x1-neonfma-pipelined.c",
349 "src/f32-spmm/16x1-neonfma-unroll2.c",
350 "src/f32-spmm/16x1-neonfma.c",
351 "src/f32-spmm/16x2-neonfma.c",
352 "src/f32-spmm/16x4-neonfma.c",
353 "src/f32-spmm/4x1-neonfma-pipelined.c",
354 "src/f32-spmm/4x1-neonfma-unroll2.c",
355 "src/f32-spmm/4x1-neonfma.c",
356 "src/f32-spmm/4x2-neonfma.c",
357 "src/f32-spmm/4x4-neonfma.c",
358 "src/f32-spmm/8x1-neonfma-pipelined.c",
359 "src/f32-spmm/8x1-neonfma-unroll2.c",
360 "src/f32-spmm/8x1-neonfma.c",
361 "src/f32-spmm/8x2-neonfma.c",
362 "src/f32-spmm/8x4-neonfma.c",
363]
364
365AARCH64_NEONFP16ARITH_UKERNELS = [
366 "src/f16-gemm/4x8-neonfp16arith-ld64.c",
367 "src/f16-gemm/6x8-neonfp16arith-ld64.c",
368 "src/f16-gemm/8x8-neonfp16arith-ld64.c",
369]
370
371SSE_UKERNELS = [
372 "src/f32-avgpool/mp9p8q-sse.c",
373 "src/f32-avgpool/up9-sse.c",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800374 "src/f32-bilinear/sse-c4.c",
375 "src/f32-bilinear/sse-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700376 "src/f32-clamp/sse.c",
377 "src/f32-igemm/1x8-sse-dup.c",
378 "src/f32-igemm/1x8-sse-load1.c",
379 "src/f32-igemm/1x8s4-sse.c",
380 "src/f32-igemm/4x2c4-sse.c",
381 "src/f32-igemm/4x8-sse-dup.c",
382 "src/f32-igemm/4x8-sse-load1.c",
383 "src/f32-igemm/4x8s4-sse.c",
384 "src/f32-dwconv/up4x25-sse.c",
385 "src/f32-dwconv/up4x4-sse.c",
386 "src/f32-dwconv/up4x9-sse.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800387 "src/f32-dwconv/up4x25-sse-acc2.c",
388 "src/f32-dwconv/up4x4-sse-acc2.c",
389 "src/f32-dwconv/up4x9-sse-acc2.c",
390 "src/f32-dwconv/up8x25-sse.c",
391 "src/f32-dwconv/up8x4-sse.c",
392 "src/f32-dwconv/up8x9-sse.c",
393 "src/f32-dwconv/up8x25-sse-acc2.c",
394 "src/f32-dwconv/up8x4-sse-acc2.c",
395 "src/f32-dwconv/up8x9-sse-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700396 "src/f32-gavgpool-spchw/sse-x4.c",
397 "src/f32-gavgpool/mp7p7q-sse.c",
398 "src/f32-gavgpool/up7-sse.c",
399 "src/f32-gemm/1x8-sse-dup.c",
400 "src/f32-gemm/1x8-sse-load1.c",
401 "src/f32-gemm/1x8s4-sse.c",
402 "src/f32-gemm/4x8-sse-dup.c",
403 "src/f32-gemm/4x8-sse-load1.c",
404 "src/f32-gemm/4x8s4-sse.c",
405 "src/f32-gemminc/1x8-sse-dup.c",
406 "src/f32-gemminc/1x8-sse-load1.c",
407 "src/f32-gemminc/1x8s4-sse.c",
408 "src/f32-gemminc/4x8-sse-dup.c",
409 "src/f32-gemminc/4x8-sse-load1.c",
410 "src/f32-gemminc/4x8s4-sse.c",
411 "src/f32-hswish/sse.c",
412 "src/f32-maxpool/9p8q-sse.c",
413 "src/f32-pavgpool/mp9p8q-sse.c",
414 "src/f32-pavgpool/up9-sse.c",
415 "src/f32-dwconv-spchw/3x3p1-sse.c",
416 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
417 "src/f32-ppmm/4x8-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700418 "src/f32-rmax/sse.c",
419 "src/f32-spmm/4x1-sse.c",
420 "src/f32-spmm/8x1-sse.c",
421 "src/f32-vadd/sse.c",
422 "src/f32-vmul/sse.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800423 "src/f32-vmulcaddc/c4-sse-2x.c",
424 "src/f32-vmulcaddc/c8-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700425 "src/f32-vsub/sse.c",
426 "src/x32-packx/x4-sse.c",
427]
428
429SSE2_UKERNELS = [
430 "src/f32-argmaxpool/mp9p8q-sse2.c",
431 "src/f32-argmaxpool/up4-sse2.c",
432 "src/f32-argmaxpool/up9-sse2.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800433 "src/f32-prelu/sse2-2x4.c",
434 "src/f32-prelu/sse2-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700435 "src/q8-avgpool/mp9p8q-sse2.c",
436 "src/q8-avgpool/up9-sse2.c",
437 "src/q8-igemm/4x4c2-sse2.c",
438 "src/q8-dwconv/up8x9-sse2.c",
439 "src/q8-gavgpool/mp7p7q-sse2.c",
440 "src/q8-gavgpool/up7-sse2.c",
441 "src/q8-gemm/2x4c8-sse2.c",
442 "src/q8-gemm/4x4c2-sse2.c",
443 "src/q8-vadd/sse2.c",
444 "src/u8-clamp/sse2.c",
445 "src/u8-maxpool/9p8q-sse2.c",
446 "src/u8-rmax/sse2.c",
447 "src/x32-pad/x2-sse2.c",
448 "src/x32-zip/x2-sse2.c",
449 "src/x32-zip/x3-sse2.c",
450 "src/x32-zip/x4-sse2.c",
451 "src/x32-zip/xm-sse2.c",
452 "src/x8-zip/x2-sse2.c",
453 "src/x8-zip/x3-sse2.c",
454 "src/x8-zip/x4-sse2.c",
455 "src/x8-zip/xm-sse2.c",
456]
457
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800458SSE41_UKERNELS = [
459 "src/f32-prelu/sse41-2x4.c",
460 "src/f32-prelu/sse41-2x8.c",
461]
462
Marat Dukhan08c4a432019-10-03 09:29:21 -0700463AVX_UKERNELS = [
464 "src/f32-rmax/avx.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700465 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700466]
467
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700468AVX2_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700469 "src/f32-raddexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700470 "src/f32-raddextexp/avx2-p5-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700471 "src/f32-raddstoreexpminusmax/avx2-p5-unroll64.c",
472 "src/f32-vscaleexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700473 "src/f32-vscaleextexp/avx2-p5-unroll64.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700474 "src/math/exp-avx2-p5.c",
475 "src/math/exp-avx2-perm-p3.c",
476 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -0700477 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700478 "src/math/extexp-avx2-p5.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700479]
480
Marat Dukhan08c4a432019-10-03 09:29:21 -0700481AVX512F_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700482 "src/f32-raddexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700483 "src/f32-raddextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700484 "src/f32-raddstoreexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700485 "src/f32-rmax/avx512f.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700486 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700487 "src/f32-vscaleexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700488 "src/f32-vscaleextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700489 "src/math/exp-avx512f-p5-scalef.c",
490 "src/math/exp-avx512f-p5.c",
491 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -0700492 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700493 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700494]
495
496AARCH32_ASM_UKERNELS = [
497 "src/q8-dwconv/up8x9-aarch32-neon.S",
498]
499
500AARCH64_ASM_UKERNELS = [
501 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
502 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
503 "src/f32-gemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700504 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700505 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a57.S",
506 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a75.S",
507 "src/f32-gemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700508 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700509 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a57.S",
510 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a75.S",
511 "src/f32-gemm/4x8-aarch64-neonfma-ld128.S",
512 "src/f32-gemm/4x8-aarch64-neonfma-ld64.S",
513 "src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700514 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700515 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a57.S",
516 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S",
517 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S",
518 "src/f32-gemm/6x8-aarch64-neonfma-ld128.S",
519 "src/f32-gemm/6x8-aarch64-neonfma-ld64.S",
520 "src/f32-gemminc/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700521 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700522 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a57.S",
523 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a75.S",
524 "src/f32-gemminc/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700525 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700526 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a57.S",
527 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a75.S",
528 "src/f32-gemminc/4x8-aarch64-neonfma-ld128.S",
529 "src/f32-gemminc/4x8-aarch64-neonfma-ld64.S",
530 "src/f32-gemminc/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700531 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700532 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a57.S",
533 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a73.S",
534 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a75.S",
535 "src/f32-gemminc/6x8-aarch64-neonfma-ld128.S",
536 "src/f32-gemminc/6x8-aarch64-neonfma-ld64.S",
537 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700538 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700539 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S",
540 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S",
541 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700542 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700543 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
544 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700545 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700546 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
547 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
548 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
549]
550
551INTERNAL_MICROKERNEL_HDRS = [
552 "src/xnnpack/argmaxpool.h",
553 "src/xnnpack/avgpool.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800554 "src/xnnpack/bilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700555 "src/xnnpack/clamp.h",
556 "src/xnnpack/common.h",
557 "src/xnnpack/conv.h",
558 "src/xnnpack/dwconv.h",
559 "src/xnnpack/gavgpool.h",
560 "src/xnnpack/gemm.h",
561 "src/xnnpack/hswish.h",
562 "src/xnnpack/igemm.h",
563 "src/xnnpack/lut.h",
564 "src/xnnpack/math.h",
565 "src/xnnpack/maxpool.h",
566 "src/xnnpack/packx.h",
567 "src/xnnpack/pad.h",
568 "src/xnnpack/params.h",
569 "src/xnnpack/pavgpool.h",
570 "src/xnnpack/ppmm.h",
571 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700572 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700573 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700574 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700575 "src/xnnpack/rmax.h",
576 "src/xnnpack/scalar-utils.h",
577 "src/xnnpack/spmm.h",
578 "src/xnnpack/unpool.h",
579 "src/xnnpack/vadd.h",
580 "src/xnnpack/vmul.h",
581 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700582 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700583 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700584 "src/xnnpack/vscaleextexp.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700585 "src/xnnpack/vsub.h",
586 "src/xnnpack/zip.h",
587]
588
589INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
590 "include/xnnpack.h",
591 "src/xnnpack/allocator.h",
592 "src/xnnpack/compute.h",
593 "src/xnnpack/im2col.h",
594 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700595 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700596 "src/xnnpack/operator.h",
597 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700598 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700599 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700600 "src/xnnpack/requantization.h",
601]
602
603ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
604 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700605]
606
607MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700608 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700609 "include/xnnpack.h",
610]
611
612MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
613 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700614 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700615 "src/xnnpack/requantization.h",
616 "include/xnnpack.h",
617]
618
619OPERATOR_TEST_PARAMS_HDRS = [
620 "src/xnnpack/params.h",
621 "src/xnnpack/common.h",
622]
623
624WEIGHTS_PACK_HDRS = [
625 "src/xnnpack/pack.h",
626 "src/xnnpack/operator.h",
627 "src/xnnpack/compute.h",
628]
629
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700630LOGGING_COPTS = select({
631 # No logging in optimized mode
632 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
633 # Full logging in debug mode
634 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
635 # Error-only logging in default (fastbuild) mode
636 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
637})
638
639LOGGING_HDRS = [
640 "src/xnnpack/log.h",
641]
642
Marat Dukhan08c4a432019-10-03 09:29:21 -0700643xnnpack_cc_library(
644 name = "scalar_ukernels",
645 srcs = SCALAR_UKERNELS,
646 hdrs = INTERNAL_HDRS,
647 aarch32_copts = ["-marm"],
648 copts = xnnpack_std_copts(),
649 deps = [
650 "@FP16",
651 "@FXdiv",
652 ],
653)
654
655xnnpack_cc_library(
656 name = "psimd_ukernels",
657 srcs = PSIMD_UKERNELS,
658 hdrs = INTERNAL_HDRS,
659 aarch32_copts = [
660 "-marm",
661 "-mfpu=neon",
662 ],
663 copts = xnnpack_std_copts(),
664 optimized_copts = [
665 "-O3",
666 "-ffast-math",
667 ],
668 deps = [
669 "@FP16",
670 "@psimd",
671 ],
672)
673
674xnnpack_cc_library(
675 name = "neon_ukernels",
676 hdrs = INTERNAL_HDRS,
677 aarch32_copts = [
678 "-marm",
679 "-mfpu=neon",
680 ],
681 aarch32_srcs = NEON_UKERNELS,
682 aarch64_srcs = NEON_UKERNELS,
683 copts = xnnpack_std_copts(),
684 deps = ["@FP16"],
685)
686
687xnnpack_cc_library(
688 name = "neonfma_ukernels",
689 hdrs = INTERNAL_HDRS,
690 aarch32_copts = [
691 "-marm",
692 "-mfpu=neon-vfpv4",
693 ],
694 aarch32_srcs = NEONFMA_UKERNELS,
695 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
696 copts = xnnpack_std_copts(),
697 deps = ["@FP16"],
698)
699
700xnnpack_cc_library(
701 name = "neonfp16arith_ukernels",
702 hdrs = INTERNAL_HDRS,
703 aarch64_copts = ["-march=armv8.2-a+fp16"],
704 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
705 copts = xnnpack_std_copts(),
706 deps = ["@FP16"],
707)
708
709xnnpack_cc_library(
710 name = "sse2_ukernels",
711 hdrs = INTERNAL_HDRS,
712 copts = xnnpack_std_copts(),
713 x86_copts = ["-msse2"],
714 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
715 deps = ["@FP16"],
716)
717
718xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800719 name = "sse41_ukernels",
720 hdrs = INTERNAL_HDRS,
721 copts = xnnpack_std_copts(),
722 x86_copts = ["-msse4.1"],
723 x86_srcs = SSE41_UKERNELS,
724 deps = ["@FP16"],
725)
726
727xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700728 name = "avx_ukernels",
729 hdrs = INTERNAL_HDRS,
730 copts = xnnpack_std_copts(),
731 x86_copts = ["-mavx"],
732 x86_srcs = AVX_UKERNELS,
733 deps = ["@FP16"],
734)
735
736xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700737 name = "avx2_ukernels",
738 hdrs = INTERNAL_HDRS,
739 copts = xnnpack_std_copts(),
740 x86_copts = [
741 "-mfma",
742 "-mavx2",
743 ],
744 x86_srcs = AVX2_UKERNELS,
745 deps = ["@FP16"],
746)
747
748xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700749 name = "avx512f_ukernels",
750 hdrs = INTERNAL_HDRS,
751 copts = xnnpack_std_copts(),
752 x86_copts = ["-mavx512f"],
753 x86_srcs = AVX512F_UKERNELS,
754 deps = ["@FP16"],
755)
756
757xnnpack_cc_library(
758 name = "asm_ukernels",
759 hdrs = ["src/xnnpack/assembly.h"],
760 aarch32_srcs = AARCH32_ASM_UKERNELS,
761 aarch64_srcs = AARCH64_ASM_UKERNELS,
762)
763
764xnnpack_aggregate_library(
765 name = "ukernels",
766 aarch32_deps = [
767 ":psimd_ukernels",
768 ":neon_ukernels",
769 ":neonfma_ukernels",
770 ":asm_ukernels",
771 ],
772 aarch64_deps = [
773 ":psimd_ukernels",
774 ":neon_ukernels",
775 ":neonfma_ukernels",
776 ":neonfp16arith_ukernels",
777 ":asm_ukernels",
778 ],
779 generic_deps = [":scalar_ukernels"],
780 wasmsimd_deps = [
781 ":psimd_ukernels",
782 ],
783 x86_deps = [
784 ":psimd_ukernels",
785 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800786 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700787 ":avx_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700788 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700789 ":avx512f_ukernels",
790 ],
791)
792
793xnnpack_cc_library(
794 name = "im2col",
795 srcs = ["src/im2col.c"],
796 hdrs = [
797 "src/xnnpack/common.h",
798 "src/xnnpack/im2col.h",
799 ],
800 copts = xnnpack_std_copts(),
801)
802
803xnnpack_cc_library(
804 name = "indirection",
805 srcs = ["src/indirection.c"],
806 hdrs = INTERNAL_HDRS,
807 copts = xnnpack_std_copts(),
808 deps = [
809 "@FP16",
810 "@FXdiv",
811 "@pthreadpool",
812 ],
813)
814
815xnnpack_cc_library(
816 name = "operator_run",
817 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700818 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
819 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700820 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
821 "-Wno-vla",
822 ],
823 deps = [
824 "@FP16",
825 "@FXdiv",
826 "@clog",
827 "@pthreadpool",
828 ],
829)
830
831cc_library(
832 name = "enable_assembly",
833 defines = select({
834 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
835 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -0700836 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700837 }),
838)
839
Marat Dukhancf056b22019-10-07 10:26:29 -0700840xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700841 name = "operators",
842 srcs = OPERATOR_SRCS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700843 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -0700844 ],
845 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700846 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700847 "-Isrc",
848 "-Iinclude",
849 ] + select({
850 ":debug_build": [],
851 "//conditions:default": xnnpack_min_size_copts(),
852 }),
Marat Dukhancf056b22019-10-07 10:26:29 -0700853 wasm_srcs = ["src/wasm-stubs.c"],
854 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700855 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700856 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700857 "@FP16",
858 "@FXdiv",
859 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700860 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700861 ],
862)
863
864cc_library(
865 name = "XNNPACK",
866 srcs = [
867 "src/init.c",
868 ],
869 copts = xnnpack_std_copts() + LOGGING_COPTS + [
870 "-Isrc",
871 "-Iinclude",
872 ] + select({
873 ":debug_build": [],
874 "//conditions:default": xnnpack_min_size_copts(),
875 }),
876 includes = ["include"],
877 linkstatic = True,
878 textual_hdrs = ["include/xnnpack.h"],
879 visibility = xnnpack_visibility(),
880 deps = [
881 ":enable_assembly",
882 ":ukernels",
883 ":operator_run",
884 ":operators",
885 "@clog",
886 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -0700887 ] + select({
888 ":emscripten": [],
889 "//conditions:default": ["@cpuinfo"],
890 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700891)
892
893cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700894 name = "xnnpack_operators_nhwc_f32",
895 srcs = [
896 "src/init.c",
897 ],
898 copts = xnnpack_std_copts() + LOGGING_COPTS + [
899 "-Isrc",
900 "-Iinclude",
901 ] + select({
902 ":debug_build": [],
903 "//conditions:default": xnnpack_min_size_copts(),
904 }),
905 defines = [
906 "XNN_NO_Q8_OPERATORS",
907 "XNN_NO_U8_OPERATORS",
908 "XNN_NO_X8_OPERATORS",
909 "XNN_NO_SPNCHW_OPERATORS",
910 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700911 includes = ["include"],
912 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700913 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700914 visibility = xnnpack_visibility(),
915 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700916 ":enable_assembly",
917 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700918 ":operator_run",
919 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700920 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700921 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700922 ] + select({
923 ":emscripten": [],
924 "//conditions:default": ["@cpuinfo"],
925 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700926)
927
Marat Dukhancf056b22019-10-07 10:26:29 -0700928xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700929 name = "bench_utils",
930 srcs = ["bench/utils.cc"],
931 hdrs = ["bench/utils.h"],
932 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -0800933 deps = [
934 "@com_google_benchmark//:benchmark",
935 "@cpuinfo",
936 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700937)
938
Frank Barchard7e955972019-10-11 10:34:25 -0700939######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -0700940
941xnnpack_benchmark(
942 name = "q8_gemm_bench",
943 srcs = [
944 "bench/gemm.h",
945 "bench/q8-gemm.cc",
946 "src/xnnpack/AlignedAllocator.h",
947 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
948 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
949 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
950)
951
952xnnpack_benchmark(
953 name = "f16_gemm_bench",
954 srcs = [
955 "bench/f16-gemm.cc",
956 "bench/gemm.h",
957 "src/xnnpack/AlignedAllocator.h",
958 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
959 copts = ["-Wno-unused-function"],
960 deps = MICROKERNEL_BENCHMARK_DEPS,
961)
962
963xnnpack_benchmark(
964 name = "f32_igemm_bench",
965 srcs = [
966 "bench/f32-igemm.cc",
967 "bench/conv.h",
968 "src/xnnpack/AlignedAllocator.h",
969 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -0700970 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700971)
972
973xnnpack_benchmark(
974 name = "f32_conv_hwc_bench",
975 srcs = [
976 "bench/f32-conv-hwc.cc",
977 "bench/dconv.h",
978 "src/xnnpack/AlignedAllocator.h",
979 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
980 copts = ["-Wno-unused-function"],
981 deps = MICROKERNEL_BENCHMARK_DEPS,
982)
983
984xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -0700985 name = "f32_conv_hwc2spchw_bench",
986 srcs = [
987 "bench/f32-conv-hwc2spchw.cc",
988 "bench/dconv.h",
989 "src/xnnpack/AlignedAllocator.h",
990 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
991 copts = ["-Wno-unused-function"],
992 deps = MICROKERNEL_BENCHMARK_DEPS,
993)
994
995xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700996 name = "f32_dwconv_bench",
997 srcs = [
998 "bench/f32-dwconv.cc",
999 "bench/dwconv.h",
1000 "src/xnnpack/AlignedAllocator.h",
1001 ] + MICROKERNEL_BENCHMARK_HDRS,
1002 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1003)
1004
1005xnnpack_benchmark(
1006 name = "f32_dwconv_spchw_bench",
1007 srcs = [
1008 "bench/f32-dwconv-spchw.cc",
1009 "bench/dwconv.h",
1010 "src/xnnpack/AlignedAllocator.h",
1011 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1012 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1013)
1014
1015xnnpack_benchmark(
1016 name = "f32_gemm_bench",
1017 srcs = [
1018 "bench/f32-gemm.cc",
1019 "bench/gemm.h",
1020 "src/xnnpack/AlignedAllocator.h",
1021 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1022 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07001023 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001024)
1025
1026xnnpack_benchmark(
1027 name = "f32_rmax_bench",
1028 srcs = [
1029 "bench/f32-rmax.cc",
1030 "src/xnnpack/AlignedAllocator.h",
1031 ] + MICROKERNEL_BENCHMARK_HDRS,
1032 deps = MICROKERNEL_BENCHMARK_DEPS,
1033)
1034
1035xnnpack_benchmark(
1036 name = "f32_spmm_bench",
1037 srcs = [
1038 "bench/f32-spmm.cc",
1039 "bench/gemm.h",
1040 "src/xnnpack/AlignedAllocator.h",
1041 ] + MICROKERNEL_BENCHMARK_HDRS,
1042 copts = ["-Wno-unused-function"],
1043 deps = MICROKERNEL_BENCHMARK_DEPS,
1044)
1045
1046xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07001047 name = "f32_softargmax_bench",
1048 srcs = [
1049 "bench/f32-softargmax.cc",
1050 ] + MICROKERNEL_BENCHMARK_HDRS,
1051 copts = ["-Wno-unused-function"],
1052 deps = MICROKERNEL_BENCHMARK_DEPS,
1053)
1054
1055xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001056 name = "f32_im2col_gemm_bench",
1057 srcs = [
1058 "bench/f32-im2col-gemm.cc",
1059 "bench/conv.h",
1060 "src/xnnpack/AlignedAllocator.h",
1061 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1062 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
1063)
1064
1065########################### Benchmarks for operators ###########################
1066
1067xnnpack_benchmark(
1068 name = "add_bench",
1069 srcs = ["bench/add.cc"],
1070 deps = OPERATOR_BENCHMARK_DEPS,
1071)
1072
1073xnnpack_benchmark(
1074 name = "average_pooling_bench",
1075 srcs = ["bench/average-pooling.cc"],
1076 deps = OPERATOR_BENCHMARK_DEPS,
1077)
1078
1079xnnpack_benchmark(
1080 name = "channel_shuffle_bench",
1081 srcs = ["bench/channel-shuffle.cc"],
1082 deps = OPERATOR_BENCHMARK_DEPS,
1083)
1084
1085xnnpack_benchmark(
1086 name = "convolution_bench",
1087 srcs = ["bench/convolution.cc"],
1088 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1089 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1090)
1091
1092xnnpack_benchmark(
1093 name = "deconvolution_bench",
1094 srcs = ["bench/deconvolution.cc"],
1095 copts = xnnpack_optional_tflite_copts(),
1096 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1097)
1098
1099xnnpack_benchmark(
1100 name = "global_average_pooling_bench",
1101 srcs = ["bench/global-average-pooling.cc"],
1102 deps = OPERATOR_BENCHMARK_DEPS,
1103)
1104
1105xnnpack_benchmark(
1106 name = "max_pooling_bench",
1107 srcs = ["bench/max-pooling.cc"],
1108 deps = OPERATOR_BENCHMARK_DEPS,
1109)
1110
1111xnnpack_benchmark(
1112 name = "sigmoid_bench",
1113 srcs = ["bench/sigmoid.cc"],
1114 deps = OPERATOR_BENCHMARK_DEPS,
1115)
1116
1117xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07001118 name = "prelu_bench",
1119 srcs = ["bench/prelu.cc"],
1120 copts = xnnpack_optional_tflite_copts(),
1121 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1122)
1123
1124xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001125 name = "softargmax_bench",
1126 srcs = ["bench/softargmax.cc"],
1127 deps = OPERATOR_BENCHMARK_DEPS,
1128)
1129
Marat Dukhanc068bb62019-10-04 13:24:39 -07001130############################# End-to-end benchmarks ############################
1131
1132cc_library(
1133 name = "mobilenet_v1",
1134 srcs = ["models/mobilenet-v1.cc"],
1135 hdrs = ["models/models.h"],
1136 linkstatic = True,
1137 deps = [
1138 ":XNNPACK",
1139 "@pthreadpool",
1140 ],
1141)
1142
1143cc_library(
1144 name = "mobilenet_v2",
1145 srcs = ["models/mobilenet-v2.cc"],
1146 hdrs = ["models/models.h"],
1147 linkstatic = True,
1148 deps = [
1149 ":XNNPACK",
1150 "@pthreadpool",
1151 ],
1152)
1153
1154xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07001155 name = "f32_dwconv_e2e_bench",
1156 srcs = ["bench/f32-dwconv-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1157 copts = ["-Wno-unused-function"],
1158 deps = MICROKERNEL_BENCHMARK_DEPS + [
1159 ":XNNPACK",
1160 ":mobilenet_v1",
1161 ":mobilenet_v2",
1162 ],
1163)
1164
1165xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07001166 name = "f32_gemm_e2e_bench",
1167 srcs = ["bench/f32-gemm-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1168 copts = ["-Wno-unused-function"],
1169 deps = MICROKERNEL_BENCHMARK_DEPS + [
1170 ":XNNPACK",
1171 ":mobilenet_v1",
1172 ":mobilenet_v2",
1173 ],
1174)
1175
1176xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07001177 name = "end2end_bench",
1178 srcs = ["bench/end2end.cc"],
1179 deps = [
1180 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07001181 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07001182 ":mobilenet_v1",
1183 ":mobilenet_v2",
1184 "@pthreadpool",
1185 ],
1186)
1187
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001188#################### Accuracy evaluation for math functions ####################
1189
1190xnnpack_benchmark(
1191 name = "f32_exp_eval",
1192 srcs = [
1193 "eval/f32-exp.cc",
1194 "src/xnnpack/AlignedAllocator.h",
1195 ] + ACCURACY_EVAL_HDRS,
1196 deps = ACCURACY_EVAL_DEPS,
1197)
1198
Marat Dukhan515c9772019-10-17 18:07:57 -07001199xnnpack_benchmark(
1200 name = "f32_expminus_eval",
1201 srcs = [
1202 "eval/f32-expminus.cc",
1203 "src/xnnpack/AlignedAllocator.h",
1204 ] + ACCURACY_EVAL_HDRS,
1205 deps = ACCURACY_EVAL_DEPS,
1206)
1207
Marat Dukhan98ba4412019-10-23 02:14:28 -07001208xnnpack_benchmark(
1209 name = "f32_extexp_eval",
1210 srcs = [
1211 "eval/f32-extexp.cc",
1212 "src/xnnpack/AlignedAllocator.h",
1213 ] + ACCURACY_EVAL_HDRS,
1214 deps = ACCURACY_EVAL_DEPS,
1215)
1216
Marat Dukhan08c4a432019-10-03 09:29:21 -07001217######################### Unit tests for micro-kernels #########################
1218
1219xnnpack_unit_test(
1220 name = "f16_gemm_test",
1221 srcs = [
1222 "test/f16-gemm.cc",
1223 "test/gemm-microkernel-tester.h",
1224 "src/xnnpack/AlignedAllocator.h",
1225 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1226 deps = MICROKERNEL_TEST_DEPS,
1227)
1228
1229xnnpack_unit_test(
1230 name = "f32_argmaxpool_test",
1231 srcs = [
1232 "test/f32-argmaxpool.cc",
1233 "test/argmaxpool-microkernel-tester.h",
1234 "src/xnnpack/AlignedAllocator.h",
1235 ] + MICROKERNEL_TEST_HDRS,
1236 deps = MICROKERNEL_TEST_DEPS,
1237)
1238
1239xnnpack_unit_test(
1240 name = "f32_avgpool_test",
1241 srcs = [
1242 "test/f32-avgpool.cc",
1243 "test/avgpool-microkernel-tester.h",
1244 "src/xnnpack/AlignedAllocator.h",
1245 ] + MICROKERNEL_TEST_HDRS,
1246 deps = MICROKERNEL_TEST_DEPS,
1247)
1248
1249xnnpack_unit_test(
Marat Dukhan35dacfb2019-11-07 19:18:16 -08001250 name = "f32_bilinear_test",
1251 srcs = [
1252 "test/f32-bilinear.cc",
1253 "test/bilinear-microkernel-tester.h",
1254 "src/xnnpack/AlignedAllocator.h",
1255 ] + MICROKERNEL_TEST_HDRS,
1256 deps = MICROKERNEL_TEST_DEPS,
1257)
1258
1259xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001260 name = "f32_clamp_test",
1261 srcs = [
1262 "test/f32-clamp.cc",
1263 "test/clamp-microkernel-tester.h",
1264 ] + MICROKERNEL_TEST_HDRS,
1265 deps = MICROKERNEL_TEST_DEPS,
1266)
1267
1268xnnpack_unit_test(
1269 name = "f32_igemm_test",
1270 srcs = [
1271 "test/f32-igemm.cc",
1272 "test/gemm-microkernel-tester.h",
1273 "src/xnnpack/AlignedAllocator.h",
1274 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1275 deps = MICROKERNEL_TEST_DEPS,
1276)
1277
1278xnnpack_unit_test(
1279 name = "f32_conv_hwc_test",
1280 srcs = [
1281 "test/f32-conv-hwc.cc",
1282 "test/conv-hwc-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_conv_hwc2spchw_test",
1290 srcs = [
1291 "test/f32-conv-hwc2spchw.cc",
1292 "test/conv-hwc2spchw-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_dwconv_test",
1300 srcs = [
1301 "test/f32-dwconv.cc",
1302 "test/dwconv-microkernel-tester.h",
1303 "src/xnnpack/AlignedAllocator.h",
1304 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1305 deps = MICROKERNEL_TEST_DEPS,
1306)
1307
1308xnnpack_unit_test(
1309 name = "f32_dwconv_spchw_test",
1310 srcs = [
1311 "test/f32-dwconv-spchw.cc",
1312 "test/dwconv-spchw-microkernel-tester.h",
1313 "src/xnnpack/AlignedAllocator.h",
1314 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1315 deps = MICROKERNEL_TEST_DEPS,
1316)
1317
1318xnnpack_unit_test(
1319 name = "f32_gavgpool_test",
1320 srcs = [
1321 "test/f32-gavgpool.cc",
1322 "test/gavgpool-microkernel-tester.h",
1323 "src/xnnpack/AlignedAllocator.h",
1324 ] + MICROKERNEL_TEST_HDRS,
1325 deps = MICROKERNEL_TEST_DEPS,
1326)
1327
1328xnnpack_unit_test(
1329 name = "f32_gavgpool_spchw_test",
1330 srcs = [
1331 "test/f32-gavgpool-spchw.cc",
1332 "test/gavgpool-spchw-microkernel-tester.h",
1333 "src/xnnpack/AlignedAllocator.h",
1334 ] + MICROKERNEL_TEST_HDRS,
1335 deps = MICROKERNEL_TEST_DEPS,
1336)
1337
1338xnnpack_unit_test(
1339 name = "f32_gemm_test",
1340 srcs = [
1341 "test/f32-gemm.cc",
1342 "test/gemm-microkernel-tester.h",
1343 "src/xnnpack/AlignedAllocator.h",
1344 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1345 deps = MICROKERNEL_TEST_DEPS,
1346)
1347
1348xnnpack_unit_test(
1349 name = "f32_gemminc_test",
1350 srcs = [
1351 "test/f32-gemminc.cc",
1352 "test/gemm-microkernel-tester.h",
1353 "src/xnnpack/AlignedAllocator.h",
1354 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1355 deps = MICROKERNEL_TEST_DEPS,
1356)
1357
1358xnnpack_unit_test(
1359 name = "f32_hswish_test",
1360 srcs = [
1361 "test/f32-hswish.cc",
1362 "test/hswish-microkernel-tester.h",
1363 ] + MICROKERNEL_TEST_HDRS,
1364 deps = MICROKERNEL_TEST_DEPS,
1365)
1366
1367xnnpack_unit_test(
1368 name = "f32_maxpool_test",
1369 srcs = [
1370 "test/f32-maxpool.cc",
1371 "test/maxpool-microkernel-tester.h",
1372 ] + MICROKERNEL_TEST_HDRS,
1373 deps = MICROKERNEL_TEST_DEPS,
1374)
1375
1376xnnpack_unit_test(
1377 name = "f32_pavgpool_test",
1378 srcs = [
1379 "test/f32-pavgpool.cc",
1380 "test/avgpool-microkernel-tester.h",
1381 "src/xnnpack/AlignedAllocator.h",
1382 ] + MICROKERNEL_TEST_HDRS,
1383 deps = MICROKERNEL_TEST_DEPS,
1384)
1385
1386xnnpack_unit_test(
1387 name = "f32_ppmm_test",
1388 srcs = [
1389 "test/f32-ppmm.cc",
1390 "test/gemm-microkernel-tester.h",
1391 "src/xnnpack/AlignedAllocator.h",
1392 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1393 deps = MICROKERNEL_TEST_DEPS,
1394)
1395
1396xnnpack_unit_test(
1397 name = "f32_prelu_test",
1398 srcs = [
1399 "test/f32-prelu.cc",
1400 "test/prelu-microkernel-tester.h",
1401 "src/xnnpack/AlignedAllocator.h",
1402 ] + MICROKERNEL_TEST_HDRS,
1403 deps = MICROKERNEL_TEST_DEPS,
1404)
1405
1406xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001407 name = "f32_raddexpminusmax_test",
1408 srcs = [
1409 "test/f32-raddexpminusmax.cc",
1410 "test/raddexpminusmax-microkernel-tester.h",
1411 ] + MICROKERNEL_TEST_HDRS,
1412 deps = MICROKERNEL_TEST_DEPS,
1413)
1414
1415xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001416 name = "f32_raddextexp_test",
1417 srcs = [
1418 "test/f32-raddextexp.cc",
1419 "test/raddextexp-microkernel-tester.h",
1420 ] + MICROKERNEL_TEST_HDRS,
1421 deps = MICROKERNEL_TEST_DEPS,
1422)
1423
1424xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001425 name = "f32_raddstoreexpminusmax_test",
1426 srcs = [
1427 "test/f32-raddstoreexpminusmax.cc",
1428 "test/raddstoreexpminusmax-microkernel-tester.h",
1429 ] + MICROKERNEL_TEST_HDRS,
1430 deps = MICROKERNEL_TEST_DEPS,
1431)
1432
1433xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001434 name = "f32_rmax_test",
1435 srcs = [
1436 "test/f32-rmax.cc",
1437 "test/rmax-microkernel-tester.h",
1438 ] + MICROKERNEL_TEST_HDRS,
1439 deps = MICROKERNEL_TEST_DEPS,
1440)
1441
1442xnnpack_unit_test(
1443 name = "f32_spmm_test",
1444 srcs = [
1445 "test/f32-spmm.cc",
1446 "test/spmm-microkernel-tester.h",
1447 "src/xnnpack/AlignedAllocator.h",
1448 ] + MICROKERNEL_TEST_HDRS,
1449 deps = MICROKERNEL_TEST_DEPS,
1450)
1451
1452xnnpack_unit_test(
1453 name = "f32_vadd_test",
1454 srcs = [
1455 "test/f32-vadd.cc",
1456 "test/vadd-microkernel-tester.h",
1457 ] + MICROKERNEL_TEST_HDRS,
1458 deps = MICROKERNEL_TEST_DEPS,
1459)
1460
1461xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001462 name = "f32_vmul_test",
1463 srcs = [
1464 "test/f32-vmul.cc",
1465 "test/vmul-microkernel-tester.h",
1466 ] + MICROKERNEL_TEST_HDRS,
1467 deps = MICROKERNEL_TEST_DEPS,
1468)
1469
1470xnnpack_unit_test(
1471 name = "f32_vmulcaddc_test",
1472 srcs = [
1473 "test/f32-vmulcaddc.cc",
1474 "test/vmulcaddc-microkernel-tester.h",
1475 "src/xnnpack/AlignedAllocator.h",
1476 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1477 deps = MICROKERNEL_TEST_DEPS,
1478)
1479
1480xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001481 name = "f32_vscale_test",
1482 srcs = [
1483 "test/f32-vscale.cc",
1484 "test/vscale-microkernel-tester.h",
1485 ] + MICROKERNEL_TEST_HDRS,
1486 deps = MICROKERNEL_TEST_DEPS,
1487)
1488
1489xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001490 name = "f32_vscaleexpminusmax_test",
1491 srcs = [
1492 "test/f32-vscaleexpminusmax.cc",
1493 "test/vscaleexpminusmax-microkernel-tester.h",
1494 ] + MICROKERNEL_TEST_HDRS,
1495 deps = MICROKERNEL_TEST_DEPS,
1496)
1497
1498xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001499 name = "f32_vscaleextexp_test",
1500 srcs = [
1501 "test/f32-vscaleextexp.cc",
1502 "test/vscaleextexp-microkernel-tester.h",
1503 ] + MICROKERNEL_TEST_HDRS,
1504 deps = MICROKERNEL_TEST_DEPS,
1505)
1506
1507xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001508 name = "f32_vsub_test",
1509 srcs = [
1510 "test/f32-vsub.cc",
1511 "test/vsub-microkernel-tester.h",
1512 ] + MICROKERNEL_TEST_HDRS,
1513 deps = MICROKERNEL_TEST_DEPS,
1514)
1515
1516xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001517 name = "q8_avgpool_test",
1518 srcs = [
1519 "test/q8-avgpool.cc",
1520 "test/avgpool-microkernel-tester.h",
1521 "src/xnnpack/AlignedAllocator.h",
1522 ] + MICROKERNEL_TEST_HDRS,
1523 deps = MICROKERNEL_TEST_DEPS,
1524)
1525
1526xnnpack_unit_test(
1527 name = "q8_igemm_test",
1528 srcs = [
1529 "test/q8-igemm.cc",
1530 "test/gemm-microkernel-tester.h",
1531 "src/xnnpack/AlignedAllocator.h",
1532 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1533 deps = MICROKERNEL_TEST_DEPS,
1534)
1535
1536xnnpack_unit_test(
1537 name = "q8_dwconv_test",
1538 srcs = [
1539 "test/q8-dwconv.cc",
1540 "test/dwconv-microkernel-tester.h",
1541 "src/xnnpack/AlignedAllocator.h",
1542 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1543 deps = MICROKERNEL_TEST_DEPS,
1544)
1545
1546xnnpack_unit_test(
1547 name = "q8_gavgpool_test",
1548 srcs = [
1549 "test/q8-gavgpool.cc",
1550 "test/gavgpool-microkernel-tester.h",
1551 "src/xnnpack/AlignedAllocator.h",
1552 ] + MICROKERNEL_TEST_HDRS,
1553 deps = MICROKERNEL_TEST_DEPS,
1554)
1555
1556xnnpack_unit_test(
1557 name = "q8_gemm_test",
1558 srcs = [
1559 "test/q8-gemm.cc",
1560 "test/gemm-microkernel-tester.h",
1561 "src/xnnpack/AlignedAllocator.h",
1562 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1563 deps = MICROKERNEL_TEST_DEPS,
1564)
1565
1566xnnpack_unit_test(
1567 name = "q8_vadd_test",
1568 srcs = [
1569 "test/q8-vadd.cc",
1570 "test/vadd-microkernel-tester.h",
1571 ] + MICROKERNEL_TEST_HDRS,
1572 deps = MICROKERNEL_TEST_DEPS,
1573)
1574
1575xnnpack_unit_test(
1576 name = "u8_clamp_test",
1577 srcs = [
1578 "test/u8-clamp.cc",
1579 "test/clamp-microkernel-tester.h",
1580 ] + MICROKERNEL_TEST_HDRS,
1581 deps = MICROKERNEL_TEST_DEPS,
1582)
1583
1584xnnpack_unit_test(
1585 name = "u8_lut32norm_test",
1586 srcs = [
1587 "test/u8-lut32norm.cc",
1588 "test/lut-norm-microkernel-tester.h",
1589 ] + MICROKERNEL_TEST_HDRS,
1590 deps = MICROKERNEL_TEST_DEPS,
1591)
1592
1593xnnpack_unit_test(
1594 name = "u8_maxpool_test",
1595 srcs = [
1596 "test/u8-maxpool.cc",
1597 "test/maxpool-microkernel-tester.h",
1598 ] + MICROKERNEL_TEST_HDRS,
1599 deps = MICROKERNEL_TEST_DEPS,
1600)
1601
1602xnnpack_unit_test(
1603 name = "u8_rmax_test",
1604 srcs = [
1605 "test/u8-rmax.cc",
1606 "test/rmax-microkernel-tester.h",
1607 ] + MICROKERNEL_TEST_HDRS,
1608 deps = MICROKERNEL_TEST_DEPS,
1609)
1610
1611xnnpack_unit_test(
1612 name = "x32_packx_test",
1613 srcs = [
1614 "test/x32-packx.cc",
1615 "test/pack-microkernel-tester.h",
1616 "src/xnnpack/AlignedAllocator.h",
1617 ] + MICROKERNEL_TEST_HDRS,
1618 deps = MICROKERNEL_TEST_DEPS,
1619)
1620
1621xnnpack_unit_test(
1622 name = "x32_pad_test",
1623 srcs = [
1624 "test/x32-pad.cc",
1625 "test/pad-microkernel-tester.h",
1626 ] + MICROKERNEL_TEST_HDRS,
1627 deps = MICROKERNEL_TEST_DEPS,
1628)
1629
1630xnnpack_unit_test(
1631 name = "x32_unpool_test",
1632 srcs = [
1633 "test/x32-unpool.cc",
1634 "test/unpool-microkernel-tester.h",
1635 ] + MICROKERNEL_TEST_HDRS,
1636 deps = MICROKERNEL_TEST_DEPS,
1637)
1638
1639xnnpack_unit_test(
1640 name = "x32_zip_test",
1641 srcs = [
1642 "test/x32-zip.cc",
1643 "test/zip-microkernel-tester.h",
1644 ] + MICROKERNEL_TEST_HDRS,
1645 deps = MICROKERNEL_TEST_DEPS,
1646)
1647
1648xnnpack_unit_test(
1649 name = "x8_lut_test",
1650 srcs = [
1651 "test/x8-lut.cc",
1652 "test/lut-microkernel-tester.h",
1653 ] + MICROKERNEL_TEST_HDRS,
1654 deps = MICROKERNEL_TEST_DEPS,
1655)
1656
1657xnnpack_unit_test(
1658 name = "x8_zip_test",
1659 srcs = [
1660 "test/x8-zip.cc",
1661 "test/zip-microkernel-tester.h",
1662 ] + MICROKERNEL_TEST_HDRS,
1663 deps = MICROKERNEL_TEST_DEPS,
1664)
1665
1666########################### Size test for the library ##########################
1667
1668xnnpack_binary(
1669 name = "size_test",
1670 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001671 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001672)
1673
1674########################### Unit tests for operators ###########################
1675
1676xnnpack_unit_test(
1677 name = "add_test",
1678 srcs = [
1679 "test/add.cc",
1680 "test/add-operator-tester.h",
1681 ],
1682 deps = OPERATOR_TEST_DEPS,
1683)
1684
1685xnnpack_unit_test(
1686 name = "argmax_pooling_test",
1687 srcs = [
1688 "test/argmax-pooling.cc",
1689 "test/argmax-pooling-operator-tester.h",
1690 ] + OPERATOR_TEST_PARAMS_HDRS,
1691 deps = OPERATOR_TEST_DEPS,
1692)
1693
1694xnnpack_unit_test(
1695 name = "average_pooling_test",
1696 srcs = [
1697 "test/average-pooling.cc",
1698 "test/average-pooling-operator-tester.h",
1699 ] + OPERATOR_TEST_PARAMS_HDRS,
1700 deps = OPERATOR_TEST_DEPS,
1701)
1702
1703xnnpack_unit_test(
1704 name = "channel_pad_test",
1705 srcs = [
1706 "test/channel-pad.cc",
1707 "test/channel-pad-operator-tester.h",
1708 ] + OPERATOR_TEST_PARAMS_HDRS,
1709 deps = OPERATOR_TEST_DEPS,
1710)
1711
1712xnnpack_unit_test(
1713 name = "channel_shuffle_test",
1714 srcs = [
1715 "test/channel-shuffle.cc",
1716 "test/channel-shuffle-operator-tester.h",
1717 ],
1718 deps = OPERATOR_TEST_DEPS,
1719)
1720
1721xnnpack_unit_test(
1722 name = "clamp_test",
1723 srcs = [
1724 "test/clamp.cc",
1725 "test/clamp-operator-tester.h",
1726 ],
1727 deps = OPERATOR_TEST_DEPS,
1728)
1729
1730xnnpack_unit_test(
1731 name = "convolution_test",
1732 srcs = [
1733 "test/convolution.cc",
1734 "test/convolution-operator-tester.h",
1735 ],
1736 deps = OPERATOR_TEST_DEPS,
1737)
1738
1739xnnpack_unit_test(
1740 name = "convolution_spnchw_test",
1741 srcs = [
1742 "test/convolution-spnchw.cc",
1743 "test/convolution-spnchw-operator-tester.h",
1744 ],
1745 deps = OPERATOR_TEST_DEPS,
1746)
1747
1748xnnpack_unit_test(
1749 name = "deconvolution_test",
1750 srcs = [
1751 "test/deconvolution.cc",
1752 "test/deconvolution-operator-tester.h",
1753 ] + OPERATOR_TEST_PARAMS_HDRS,
1754 deps = OPERATOR_TEST_DEPS,
1755)
1756
1757xnnpack_unit_test(
1758 name = "fully_connected_test",
1759 srcs = [
1760 "test/fully-connected.cc",
1761 "test/fully-connected-operator-tester.h",
1762 ],
1763 deps = OPERATOR_TEST_DEPS,
1764)
1765
1766xnnpack_unit_test(
1767 name = "global_average_pooling_test",
1768 srcs = [
1769 "test/global-average-pooling.cc",
1770 "test/global-average-pooling-operator-tester.h",
1771 ] + OPERATOR_TEST_PARAMS_HDRS,
1772 deps = OPERATOR_TEST_DEPS,
1773)
1774
1775xnnpack_unit_test(
1776 name = "global_average_pooling_spnchw_test",
1777 srcs = [
1778 "test/global-average-pooling-spnchw.cc",
1779 "test/global-average-pooling-spnchw-operator-tester.h",
1780 ],
1781 deps = OPERATOR_TEST_DEPS,
1782)
1783
1784xnnpack_unit_test(
1785 name = "hardswish_test",
1786 srcs = [
1787 "test/hardswish.cc",
1788 "test/hardswish-operator-tester.h",
1789 ],
1790 deps = OPERATOR_TEST_DEPS,
1791)
1792
1793xnnpack_unit_test(
1794 name = "leaky_relu_test",
1795 srcs = [
1796 "test/leaky-relu.cc",
1797 "test/leaky-relu-operator-tester.h",
1798 ],
1799 deps = OPERATOR_TEST_DEPS,
1800)
1801
1802xnnpack_unit_test(
1803 name = "max_pooling_test",
1804 srcs = [
1805 "test/max-pooling.cc",
1806 "test/max-pooling-operator-tester.h",
1807 ] + OPERATOR_TEST_PARAMS_HDRS,
1808 deps = OPERATOR_TEST_DEPS,
1809)
1810
1811xnnpack_unit_test(
1812 name = "prelu_test",
1813 srcs = [
1814 "test/prelu.cc",
1815 "test/prelu-operator-tester.h",
1816 ] + OPERATOR_TEST_PARAMS_HDRS,
1817 deps = OPERATOR_TEST_DEPS,
1818)
1819
1820xnnpack_unit_test(
1821 name = "sigmoid_test",
1822 srcs = [
1823 "test/sigmoid.cc",
1824 "test/sigmoid-operator-tester.h",
1825 ],
1826 deps = OPERATOR_TEST_DEPS,
1827)
1828
1829xnnpack_unit_test(
1830 name = "softargmax_test",
1831 srcs = [
1832 "test/softargmax.cc",
1833 "test/softargmax-operator-tester.h",
1834 ],
1835 deps = OPERATOR_TEST_DEPS,
1836)
1837
1838xnnpack_unit_test(
1839 name = "unpooling_test",
1840 srcs = [
1841 "test/unpooling.cc",
1842 "test/unpooling-operator-tester.h",
1843 ],
1844 deps = OPERATOR_TEST_DEPS,
1845)
1846
1847############################# Build configurations #############################
1848
Marat Dukhanb8642352019-10-30 15:43:02 -07001849# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07001850config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07001851 name = "xnn_enable_assembly_explicit_true",
1852 define_values = {"xnn_enable_assembly": "true"},
1853)
1854
1855# Disables usage of assembly kernels.
1856config_setting(
1857 name = "xnn_enable_assembly_explicit_false",
1858 define_values = {"xnn_enable_assembly": "false"},
1859)
1860
1861# Builds with -c dbg
1862config_setting(
1863 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001864 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07001865 "compilation_mode": "dbg",
1866 },
1867)
1868
1869# Builds with -c opt
1870config_setting(
1871 name = "optimized_build",
1872 values = {
1873 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001874 },
1875)
1876
1877config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07001878 name = "linux_k8",
1879 values = {"cpu": "k8"},
1880)
1881
1882config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07001883 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07001884 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07001885)
1886
1887config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001888 name = "android",
1889 values = {"crosstool_top": "//external:android/crosstool"},
1890)
1891
1892config_setting(
1893 name = "android_armv7",
1894 values = {
1895 "crosstool_top": "//external:android/crosstool",
1896 "cpu": "armeabi-v7a",
1897 },
1898)
1899
1900config_setting(
1901 name = "android_arm64",
1902 values = {
1903 "crosstool_top": "//external:android/crosstool",
1904 "cpu": "arm64-v8a",
1905 },
1906)
1907
1908config_setting(
1909 name = "android_x86",
1910 values = {
1911 "crosstool_top": "//external:android/crosstool",
1912 "cpu": "x86",
1913 },
1914)
1915
1916config_setting(
1917 name = "android_x86_64",
1918 values = {
1919 "crosstool_top": "//external:android/crosstool",
1920 "cpu": "x86_64",
1921 },
1922)
1923
1924config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07001925 name = "macos_x86_64",
1926 values = {
1927 "apple_platform_type": "macos",
1928 "cpu": "darwin",
1929 },
1930)
1931
1932config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001933 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001934 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07001935)
1936
1937config_setting(
1938 name = "emscripten_wasm",
1939 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001940 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001941 "cpu": "wasm",
1942 },
1943)
1944
1945config_setting(
1946 name = "emscripten_wasmsimd",
1947 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001948 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001949 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07001950 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001951 },
1952)
1953
1954config_setting(
1955 name = "emscripten_asmjs",
1956 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001957 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001958 "cpu": "asmjs",
1959 },
1960)