Gitiles
Code Review Sign In
gerrit-public.fairphone.software / platform / external / XNNPACK / c3b9e862010e48bf345aed3528e98e8b167d4b44 / . / BUILD.bazel
blob: 5f56498b6b9e100a228200344b9e9b27827e1bdf [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]1# 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 -0800[diff] [blame]9load(":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 -0700[diff] [blame]11licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]15OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
19 "@pthreadpool",
20]
21
22MICROKERNEL_BENCHMARK_DEPS = [
23 ":ukernels",
24 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -0700[diff] [blame]25 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]26 "@cpuinfo",
27 "@FP16",
28 "@pthreadpool",
29]
30
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700[diff] [blame]31ACCURACY_EVAL_DEPS = [
32 ":XNNPACK",
33 ":ukernels",
34 "@FP16",
35 "@pthreadpool",
36]
37
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]38MICROKERNEL_TEST_DEPS = [
39 ":ukernels",
Frank Barchard7e955972019-10-11 10:34:25 -0700[diff] [blame]40 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]41 "@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",
Marat Dukhanca2733c2019-11-15 23:21:17 -0800[diff] [blame]68 "src/multiply.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]69 "src/prelu.c",
Marat Dukhan69722492019-11-11 19:55:50 -0800[diff] [blame]70 "src/resize-bilinear.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]71 "src/sigmoid.c",
72 "src/softargmax.c",
73 "src/unpooling.c",
74]
75
76SCALAR_UKERNELS = [
77 "src/f32-argmaxpool/mp9p8q-scalar.c",
78 "src/f32-argmaxpool/up4-scalar.c",
79 "src/f32-argmaxpool/up9-scalar.c",
80 "src/f32-avgpool/mp9p8q-scalar.c",
81 "src/f32-avgpool/up9-scalar.c",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800[diff] [blame]82 "src/f32-bilinear/scalar-c1.c",
83 "src/f32-bilinear/scalar-c2.c",
84 "src/f32-bilinear/scalar-c4.c",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -0800[diff] [blame]85 "src/f32-binop/vadd-scalar-x1.c",
86 "src/f32-binop/vadd-scalar-x2.c",
87 "src/f32-binop/vadd-scalar-x4.c",
88 "src/f32-binop/vaddc-scalar-x1.c",
89 "src/f32-binop/vaddc-scalar-x2.c",
90 "src/f32-binop/vaddc-scalar-x4.c",
91 "src/f32-binop/vmul-scalar-x1.c",
92 "src/f32-binop/vmul-scalar-x2.c",
93 "src/f32-binop/vmul-scalar-x4.c",
94 "src/f32-binop/vmulc-scalar-x1.c",
95 "src/f32-binop/vmulc-scalar-x2.c",
96 "src/f32-binop/vmulc-scalar-x4.c",
97 "src/f32-binop/vsub-scalar-x1.c",
98 "src/f32-binop/vsub-scalar-x2.c",
99 "src/f32-binop/vsub-scalar-x4.c",
100 "src/f32-binop/vsubc-scalar-x1.c",
101 "src/f32-binop/vsubc-scalar-x2.c",
102 "src/f32-binop/vsubc-scalar-x4.c",
103 "src/f32-binop/vrsubc-scalar-x1.c",
104 "src/f32-binop/vrsubc-scalar-x2.c",
105 "src/f32-binop/vrsubc-scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]106 "src/f32-clamp/scalar.c",
Erich Elsen563df5f2019-10-23 08:02:21 -0700[diff] [blame]107 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]108 "src/f32-igemm/1x4-scalar.c",
109 "src/f32-igemm/2x4-scalar.c",
110 "src/f32-igemm/4x2-scalar.c",
111 "src/f32-igemm/4x4-scalar.c",
112 "src/f32-dwconv/up1x25-scalar.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800[diff] [blame]113 "src/f32-dwconv/up1x25-scalar-acc2.c",
114 "src/f32-dwconv/up2x25-scalar.c",
115 "src/f32-dwconv/up2x25-scalar-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]116 "src/f32-dwconv/up1x4-scalar.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800[diff] [blame]117 "src/f32-dwconv/up1x4-scalar-acc2.c",
118 "src/f32-dwconv/up2x4-scalar.c",
119 "src/f32-dwconv/up2x4-scalar-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]120 "src/f32-dwconv/up1x9-scalar.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800[diff] [blame]121 "src/f32-dwconv/up1x9-scalar-acc2.c",
122 "src/f32-dwconv/up2x9-scalar.c",
123 "src/f32-dwconv/up2x9-scalar-acc2.c",
Erich Elsen0cc2c532019-10-15 04:44:18 -0700[diff] [blame]124 "src/f32-dwconv-spchw/3x3p1-scalar.c",
Erich Elsen38709a62019-11-08 11:58:45 -0800[diff] [blame]125 "src/f32-dwconv-spchw/5x5p2-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -0700[diff] [blame]126 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Erich Elsen38709a62019-11-08 11:58:45 -0800[diff] [blame]127 "src/f32-dwconv-spchw/5x5s2p2-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -0700[diff] [blame]128 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]129 "src/f32-gavgpool/mp7p7q-scalar.c",
130 "src/f32-gavgpool/up7-scalar.c",
131 "src/f32-gemm/1x4-scalar.c",
132 "src/f32-gemm/2x4-scalar.c",
133 "src/f32-gemm/4x2-scalar.c",
134 "src/f32-gemm/4x4-scalar.c",
135 "src/f32-gemminc/1x4-scalar.c",
136 "src/f32-gemminc/2x4-scalar.c",
137 "src/f32-gemminc/4x4-scalar.c",
138 "src/f32-hswish/scalar.c",
139 "src/f32-maxpool/9p8q-scalar.c",
140 "src/f32-pavgpool/mp9p8q-scalar.c",
141 "src/f32-pavgpool/up9-scalar.c",
142 "src/f32-ppmm/2x4-scalar.c",
143 "src/f32-ppmm/3x3-scalar.c",
144 "src/f32-ppmm/4x2-scalar.c",
145 "src/f32-ppmm/4x4-scalar.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800[diff] [blame]146 "src/f32-prelu/scalar-2x1.c",
147 "src/f32-prelu/scalar-2x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]148 "src/f32-rmax/scalar.c",
149 "src/f32-spmm/1x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]150 "src/f32-spmm/1x1-scalar.c",
151 "src/f32-spmm/2x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]152 "src/f32-spmm/2x1-scalar.c",
153 "src/f32-spmm/4x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]154 "src/f32-spmm/4x1-scalar.c",
155 "src/f32-spmm/8x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]156 "src/f32-spmm/8x1-scalar.c",
Erich Elsenc6afd9b2019-10-24 16:10:53 -0700[diff] [blame]157 "src/f32-spmm/8x2-scalar.c",
158 "src/f32-spmm/8x4-scalar.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800[diff] [blame]159 "src/f32-vmulcaddc/c1-scalar-2x.c",
160 "src/f32-vmulcaddc/c2-scalar-2x.c",
161 "src/f32-vmulcaddc/c4-scalar-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]162 "src/q8-avgpool/mp9p8q-scalar.c",
163 "src/q8-avgpool/up9-scalar.c",
164 "src/q8-igemm/2x2-scalar.c",
165 "src/q8-dwconv/up1x9-scalar.c",
166 "src/q8-gavgpool/mp7p7q-scalar.c",
167 "src/q8-gavgpool/up7-scalar.c",
168 "src/q8-gemm/2x2-scalar.c",
169 "src/q8-vadd/scalar.c",
170 "src/u8-clamp/scalar.c",
171 "src/u8-lut32norm/scalar.c",
172 "src/u8-maxpool/9p8q-scalar.c",
173 "src/u8-rmax/scalar.c",
174 "src/x32-packx/x2-scalar.c",
175 "src/x32-packx/x3-scalar.c",
176 "src/x32-packx/x4-scalar.c",
177 "src/x32-pad/x2-scalar.c",
178 "src/x32-unpool/scalar.c",
179 "src/x32-zip/x2-scalar.c",
180 "src/x32-zip/x3-scalar.c",
181 "src/x32-zip/x4-scalar.c",
182 "src/x32-zip/xm-scalar.c",
183 "src/x8-lut/scalar.c",
184 "src/x8-zip/x2-scalar.c",
185 "src/x8-zip/x3-scalar.c",
186 "src/x8-zip/x4-scalar.c",
187 "src/x8-zip/xm-scalar.c",
188]
189
190PSIMD_UKERNELS = [
191 "src/f32-argmaxpool/mp9p8q-psimd.c",
192 "src/f32-argmaxpool/up4-psimd.c",
193 "src/f32-argmaxpool/up9-psimd.c",
194 "src/f32-avgpool/mp9p8q-psimd.c",
195 "src/f32-avgpool/up9-psimd.c",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800[diff] [blame]196 "src/f32-bilinear/psimd-c4.c",
197 "src/f32-bilinear/psimd-c8.c",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -0800[diff] [blame]198 "src/f32-binop/vadd-psimd-x4.c",
199 "src/f32-binop/vadd-psimd-x8.c",
200 "src/f32-binop/vaddc-psimd-x4.c",
201 "src/f32-binop/vaddc-psimd-x8.c",
202 "src/f32-binop/vmul-psimd-x4.c",
203 "src/f32-binop/vmul-psimd-x8.c",
204 "src/f32-binop/vmulc-psimd-x4.c",
205 "src/f32-binop/vmulc-psimd-x8.c",
206 "src/f32-binop/vrsubc-psimd-x4.c",
207 "src/f32-binop/vrsubc-psimd-x8.c",
208 "src/f32-binop/vsub-psimd-x4.c",
209 "src/f32-binop/vsub-psimd-x8.c",
210 "src/f32-binop/vsubc-psimd-x4.c",
211 "src/f32-binop/vsubc-psimd-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]212 "src/f32-clamp/psimd.c",
213 "src/f32-igemm/1x8-psimd-loadsplat.c",
214 "src/f32-igemm/1x8-psimd-splat.c",
215 "src/f32-igemm/1x8s4-psimd.c",
216 "src/f32-igemm/4x2c4-psimd.c",
217 "src/f32-igemm/4x8-psimd-loadsplat.c",
218 "src/f32-igemm/4x8-psimd-splat.c",
219 "src/f32-igemm/4x8s4-psimd.c",
220 "src/f32-igemm/6x8-psimd-loadsplat.c",
221 "src/f32-igemm/6x8-psimd-splat.c",
222 "src/f32-igemm/6x8s4-psimd.c",
223 "src/f32-dwconv/up4x25-psimd.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800[diff] [blame]224 "src/f32-dwconv/up4x25-psimd-acc2.c",
225 "src/f32-dwconv/up8x25-psimd.c",
226 "src/f32-dwconv/up8x25-psimd-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]227 "src/f32-dwconv/up4x4-psimd.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800[diff] [blame]228 "src/f32-dwconv/up4x4-psimd-acc2.c",
229 "src/f32-dwconv/up8x4-psimd.c",
230 "src/f32-dwconv/up8x4-psimd-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]231 "src/f32-dwconv/up4x9-psimd.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800[diff] [blame]232 "src/f32-dwconv/up4x9-psimd-acc2.c",
233 "src/f32-dwconv/up8x9-psimd.c",
234 "src/f32-dwconv/up8x9-psimd-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]235 "src/f32-gavgpool/mp7p7q-psimd.c",
236 "src/f32-gavgpool/up7-psimd.c",
237 "src/f32-gemm/1x8-psimd-loadsplat.c",
238 "src/f32-gemm/1x8-psimd-splat.c",
239 "src/f32-gemm/1x8s4-psimd.c",
240 "src/f32-gemm/4x8-psimd-loadsplat.c",
241 "src/f32-gemm/4x8-psimd-splat.c",
242 "src/f32-gemm/4x8s4-psimd.c",
243 "src/f32-gemm/6x8-psimd-loadsplat.c",
244 "src/f32-gemm/6x8-psimd-splat.c",
245 "src/f32-gemm/6x8s4-psimd.c",
246 "src/f32-gemminc/1x8-psimd-loadsplat.c",
247 "src/f32-gemminc/1x8-psimd-splat.c",
248 "src/f32-gemminc/1x8s4-psimd.c",
249 "src/f32-gemminc/4x8-psimd-loadsplat.c",
250 "src/f32-gemminc/4x8-psimd-splat.c",
251 "src/f32-gemminc/4x8s4-psimd.c",
252 "src/f32-gemminc/6x8-psimd-loadsplat.c",
253 "src/f32-gemminc/6x8-psimd-splat.c",
254 "src/f32-gemminc/6x8s4-psimd.c",
255 "src/f32-hswish/psimd.c",
256 "src/f32-maxpool/9p8q-psimd.c",
257 "src/f32-pavgpool/mp9p8q-psimd.c",
258 "src/f32-pavgpool/up9-psimd.c",
259 "src/f32-ppmm/4x8-psimd.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800[diff] [blame]260 "src/f32-prelu/psimd-2x4.c",
261 "src/f32-prelu/psimd-2x8.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800[diff] [blame]262 "src/f32-vmulcaddc/c4-psimd-2x.c",
263 "src/f32-vmulcaddc/c8-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]264 "src/x32-packx/x4-psimd.c",
265 "src/x32-pad/x2-psimd.c",
266 "src/x32-unpool/psimd.c",
267 "src/x32-zip/x2-psimd.c",
268 "src/x32-zip/x3-psimd.c",
269 "src/x32-zip/x4-psimd.c",
270 "src/x32-zip/xm-psimd.c",
271]
272
273# ISA-specific micro-kernels
274NEON_UKERNELS = [
275 "src/f32-avgpool/mp9p8q-neon.c",
276 "src/f32-avgpool/up9-neon.c",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800[diff] [blame]277 "src/f32-bilinear/neon-c4.c",
278 "src/f32-bilinear/neon-c8.c",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -0800[diff] [blame]279 "src/f32-binop/vadd-neon-x4.c",
280 "src/f32-binop/vadd-neon-x8.c",
281 "src/f32-binop/vaddc-neon-x4.c",
282 "src/f32-binop/vaddc-neon-x8.c",
283 "src/f32-binop/vmul-neon-x4.c",
284 "src/f32-binop/vmul-neon-x8.c",
285 "src/f32-binop/vmulc-neon-x4.c",
286 "src/f32-binop/vmulc-neon-x8.c",
287 "src/f32-binop/vrsubc-neon-x4.c",
288 "src/f32-binop/vrsubc-neon-x8.c",
289 "src/f32-binop/vsub-neon-x4.c",
290 "src/f32-binop/vsub-neon-x8.c",
291 "src/f32-binop/vsubc-neon-x4.c",
292 "src/f32-binop/vsubc-neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]293 "src/f32-clamp/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]294 "src/f32-dwconv/up4x9-neon.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800[diff] [blame]295 "src/f32-dwconv/up4x9-neon-acc2.c",
296 "src/f32-dwconv/up8x9-neon.c",
297 "src/f32-dwconv/up8x9-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]298 "src/f32-gavgpool-spchw/neon-x4.c",
299 "src/f32-gavgpool/mp7p7q-neon.c",
300 "src/f32-gavgpool/up7-neon.c",
301 "src/f32-gemm/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]302 "src/f32-gemm/4x2-neon-ld64.c",
303 "src/f32-gemm/4x8-neon-ld128.c",
304 "src/f32-gemm/4x8-neon-ld64.c",
305 "src/f32-gemm/5x8-neon-ld64.c",
306 "src/f32-gemm/6x8-neon-ld64.c",
307 "src/f32-gemminc/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]308 "src/f32-gemminc/4x8-neon-ld128.c",
309 "src/f32-gemminc/4x8-neon-ld64.c",
310 "src/f32-gemminc/5x8-neon-ld64.c",
311 "src/f32-gemminc/6x8-neon-ld64.c",
312 "src/f32-hswish/neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800[diff] [blame]313 "src/f32-igemm/1x8-neon-ld64.c",
314 "src/f32-igemm/4x2-neon-ld64.c",
315 "src/f32-igemm/4x4-neon-ld64.c",
316 "src/f32-igemm/4x8-neon-ld128.c",
317 "src/f32-igemm/4x8-neon-ld64.c",
318 "src/f32-igemm/6x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]319 "src/f32-pavgpool/mp9p8q-neon.c",
320 "src/f32-pavgpool/up9-neon.c",
321 "src/f32-ppmm/4x8-neon.c",
322 "src/f32-ppmm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800[diff] [blame]323 "src/f32-prelu/neon-2x4.c",
324 "src/f32-prelu/neon-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]325 "src/f32-rmax/neon.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800[diff] [blame]326 "src/f32-vmulcaddc/c4-neon-2x.c",
327 "src/f32-vmulcaddc/c8-neon-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]328 "src/q8-avgpool/mp9p8q-neon.c",
329 "src/q8-avgpool/up9-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]330 "src/q8-dwconv/up8x9-neon.c",
331 "src/q8-gavgpool/mp7p7q-neon.c",
332 "src/q8-gavgpool/up7-neon.c",
333 "src/q8-gemm/4x8-neon.c",
334 "src/q8-gemm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800[diff] [blame]335 "src/q8-igemm/4x8-neon.c",
336 "src/q8-igemm/8x8-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]337 "src/q8-vadd/neon.c",
338 "src/u8-clamp/neon.c",
339 "src/u8-maxpool/9p8q-neon.c",
340 "src/u8-rmax/neon.c",
341 "src/x32-packx/x4-neon-st4.c",
342 "src/x32-pad/x2-neon.c",
343 "src/x32-zip/x2-neon.c",
344 "src/x32-zip/x3-neon.c",
345 "src/x32-zip/x4-neon.c",
346 "src/x32-zip/xm-neon.c",
347 "src/x8-zip/x2-neon.c",
348 "src/x8-zip/x3-neon.c",
349 "src/x8-zip/x4-neon.c",
350 "src/x8-zip/xm-neon.c",
351]
352
353NEONFMA_UKERNELS = [
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800[diff] [blame]354 "src/f32-bilinear/neonfma-c4.c",
355 "src/f32-bilinear/neonfma-c8.c",
Frank Barcharddb45b6a2019-10-09 16:42:45 -0700[diff] [blame]356 "src/f32-igemm/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]357 "src/f32-igemm/4x2-neonfma-ld64.c",
358 "src/f32-igemm/4x4-neonfma-ld64.c",
359 "src/f32-igemm/4x8-neonfma-ld128.c",
360 "src/f32-igemm/4x8-neonfma-ld64.c",
361 "src/f32-igemm/6x8-neonfma-ld64.c",
362 "src/f32-dwconv/up4x9-neonfma.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800[diff] [blame]363 "src/f32-dwconv/up4x9-neonfma-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]364 "src/f32-dwconv/up8x9-neonfma.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800[diff] [blame]365 "src/f32-dwconv/up8x9-neonfma-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]366 "src/f32-gemm/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]367 "src/f32-gemm/4x2-neonfma-ld64.c",
368 "src/f32-gemm/4x8-neonfma-ld128.c",
369 "src/f32-gemm/4x8-neonfma-ld64.c",
370 "src/f32-gemm/5x8-neonfma-ld64.c",
371 "src/f32-gemm/6x8-neonfma-ld64.c",
372 "src/f32-gemminc/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]373 "src/f32-gemminc/4x8-neonfma-ld128.c",
374 "src/f32-gemminc/4x8-neonfma-ld64.c",
375 "src/f32-gemminc/5x8-neonfma-ld64.c",
376 "src/f32-gemminc/6x8-neonfma-ld64.c",
377 "src/f32-hswish/neonfma.c",
378 "src/f32-ppmm/4x8-neonfma.c",
379 "src/f32-ppmm/8x8-neonfma.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800[diff] [blame]380 "src/f32-sigmoid/neonfma-p5-x16.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800[diff] [blame]381 "src/f32-vmulcaddc/c4-neonfma-2x.c",
382 "src/f32-vmulcaddc/c8-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800[diff] [blame]383 "src/math/exp-neonfma-lut64-p2.c",
384 "src/math/exp-neonfma-p5.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800[diff] [blame]385 "src/math/expminus-neonfma-p5.c",
386 "src/math/sigmoid-neonfma-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]387]
388
389AARCH64_NEONFMA_UKERNELS = [
390 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
391 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
392 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
393 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
394 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
395 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
396 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
397 "src/f32-spmm/12x1-neonfma.c",
398 "src/f32-spmm/12x2-neonfma.c",
399 "src/f32-spmm/12x4-neonfma.c",
400 "src/f32-spmm/16x1-neonfma-pipelined.c",
401 "src/f32-spmm/16x1-neonfma-unroll2.c",
402 "src/f32-spmm/16x1-neonfma.c",
403 "src/f32-spmm/16x2-neonfma.c",
404 "src/f32-spmm/16x4-neonfma.c",
405 "src/f32-spmm/4x1-neonfma-pipelined.c",
406 "src/f32-spmm/4x1-neonfma-unroll2.c",
407 "src/f32-spmm/4x1-neonfma.c",
408 "src/f32-spmm/4x2-neonfma.c",
409 "src/f32-spmm/4x4-neonfma.c",
410 "src/f32-spmm/8x1-neonfma-pipelined.c",
411 "src/f32-spmm/8x1-neonfma-unroll2.c",
412 "src/f32-spmm/8x1-neonfma.c",
413 "src/f32-spmm/8x2-neonfma.c",
414 "src/f32-spmm/8x4-neonfma.c",
415]
416
417AARCH64_NEONFP16ARITH_UKERNELS = [
418 "src/f16-gemm/4x8-neonfp16arith-ld64.c",
419 "src/f16-gemm/6x8-neonfp16arith-ld64.c",
420 "src/f16-gemm/8x8-neonfp16arith-ld64.c",
421]
422
423SSE_UKERNELS = [
424 "src/f32-avgpool/mp9p8q-sse.c",
425 "src/f32-avgpool/up9-sse.c",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800[diff] [blame]426 "src/f32-bilinear/sse-c4.c",
427 "src/f32-bilinear/sse-c8.c",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -0800[diff] [blame]428 "src/f32-binop/vadd-sse-x4.c",
429 "src/f32-binop/vadd-sse-x8.c",
430 "src/f32-binop/vaddc-sse-x4.c",
431 "src/f32-binop/vaddc-sse-x8.c",
432 "src/f32-binop/vmul-sse-x4.c",
433 "src/f32-binop/vmul-sse-x8.c",
434 "src/f32-binop/vmulc-sse-x4.c",
435 "src/f32-binop/vmulc-sse-x8.c",
436 "src/f32-binop/vrsubc-sse-x4.c",
437 "src/f32-binop/vrsubc-sse-x8.c",
438 "src/f32-binop/vsub-sse-x4.c",
439 "src/f32-binop/vsub-sse-x8.c",
440 "src/f32-binop/vsubc-sse-x4.c",
441 "src/f32-binop/vsubc-sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]442 "src/f32-clamp/sse.c",
443 "src/f32-igemm/1x8-sse-dup.c",
444 "src/f32-igemm/1x8-sse-load1.c",
445 "src/f32-igemm/1x8s4-sse.c",
446 "src/f32-igemm/4x2c4-sse.c",
447 "src/f32-igemm/4x8-sse-dup.c",
448 "src/f32-igemm/4x8-sse-load1.c",
449 "src/f32-igemm/4x8s4-sse.c",
450 "src/f32-dwconv/up4x25-sse.c",
451 "src/f32-dwconv/up4x4-sse.c",
452 "src/f32-dwconv/up4x9-sse.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800[diff] [blame]453 "src/f32-dwconv/up4x25-sse-acc2.c",
454 "src/f32-dwconv/up4x4-sse-acc2.c",
455 "src/f32-dwconv/up4x9-sse-acc2.c",
456 "src/f32-dwconv/up8x25-sse.c",
457 "src/f32-dwconv/up8x4-sse.c",
458 "src/f32-dwconv/up8x9-sse.c",
459 "src/f32-dwconv/up8x25-sse-acc2.c",
460 "src/f32-dwconv/up8x4-sse-acc2.c",
461 "src/f32-dwconv/up8x9-sse-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]462 "src/f32-gavgpool-spchw/sse-x4.c",
463 "src/f32-gavgpool/mp7p7q-sse.c",
464 "src/f32-gavgpool/up7-sse.c",
465 "src/f32-gemm/1x8-sse-dup.c",
466 "src/f32-gemm/1x8-sse-load1.c",
467 "src/f32-gemm/1x8s4-sse.c",
468 "src/f32-gemm/4x8-sse-dup.c",
469 "src/f32-gemm/4x8-sse-load1.c",
470 "src/f32-gemm/4x8s4-sse.c",
471 "src/f32-gemminc/1x8-sse-dup.c",
472 "src/f32-gemminc/1x8-sse-load1.c",
473 "src/f32-gemminc/1x8s4-sse.c",
474 "src/f32-gemminc/4x8-sse-dup.c",
475 "src/f32-gemminc/4x8-sse-load1.c",
476 "src/f32-gemminc/4x8s4-sse.c",
477 "src/f32-hswish/sse.c",
478 "src/f32-maxpool/9p8q-sse.c",
479 "src/f32-pavgpool/mp9p8q-sse.c",
480 "src/f32-pavgpool/up9-sse.c",
481 "src/f32-dwconv-spchw/3x3p1-sse.c",
482 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
483 "src/f32-ppmm/4x8-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]484 "src/f32-rmax/sse.c",
485 "src/f32-spmm/4x1-sse.c",
486 "src/f32-spmm/8x1-sse.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800[diff] [blame]487 "src/f32-vmulcaddc/c4-sse-2x.c",
488 "src/f32-vmulcaddc/c8-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]489 "src/x32-packx/x4-sse.c",
490]
491
492SSE2_UKERNELS = [
493 "src/f32-argmaxpool/mp9p8q-sse2.c",
494 "src/f32-argmaxpool/up4-sse2.c",
495 "src/f32-argmaxpool/up9-sse2.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800[diff] [blame]496 "src/f32-prelu/sse2-2x4.c",
497 "src/f32-prelu/sse2-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]498 "src/q8-avgpool/mp9p8q-sse2.c",
499 "src/q8-avgpool/up9-sse2.c",
500 "src/q8-igemm/4x4c2-sse2.c",
501 "src/q8-dwconv/up8x9-sse2.c",
502 "src/q8-gavgpool/mp7p7q-sse2.c",
503 "src/q8-gavgpool/up7-sse2.c",
504 "src/q8-gemm/2x4c8-sse2.c",
505 "src/q8-gemm/4x4c2-sse2.c",
506 "src/q8-vadd/sse2.c",
507 "src/u8-clamp/sse2.c",
508 "src/u8-maxpool/9p8q-sse2.c",
509 "src/u8-rmax/sse2.c",
510 "src/x32-pad/x2-sse2.c",
511 "src/x32-zip/x2-sse2.c",
512 "src/x32-zip/x3-sse2.c",
513 "src/x32-zip/x4-sse2.c",
514 "src/x32-zip/xm-sse2.c",
515 "src/x8-zip/x2-sse2.c",
516 "src/x8-zip/x3-sse2.c",
517 "src/x8-zip/x4-sse2.c",
518 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -0800[diff] [blame]519 "src/math/exp-sse2-p5.c",
520 "src/math/expminus-sse2-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]521]
522
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800[diff] [blame]523SSE41_UKERNELS = [
524 "src/f32-prelu/sse41-2x4.c",
525 "src/f32-prelu/sse41-2x8.c",
526]
527
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]528AVX_UKERNELS = [
529 "src/f32-rmax/avx.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700[diff] [blame]530 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]531]
532
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700[diff] [blame]533AVX2_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700[diff] [blame]534 "src/f32-raddexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700[diff] [blame]535 "src/f32-raddextexp/avx2-p5-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700[diff] [blame]536 "src/f32-raddstoreexpminusmax/avx2-p5-unroll64.c",
537 "src/f32-vscaleexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700[diff] [blame]538 "src/f32-vscaleextexp/avx2-p5-unroll64.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700[diff] [blame]539 "src/math/exp-avx2-p5.c",
540 "src/math/exp-avx2-perm-p3.c",
541 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -0700[diff] [blame]542 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700[diff] [blame]543 "src/math/extexp-avx2-p5.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700[diff] [blame]544]
545
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]546AVX512F_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700[diff] [blame]547 "src/f32-raddexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700[diff] [blame]548 "src/f32-raddextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700[diff] [blame]549 "src/f32-raddstoreexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]550 "src/f32-rmax/avx512f.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700[diff] [blame]551 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700[diff] [blame]552 "src/f32-vscaleexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700[diff] [blame]553 "src/f32-vscaleextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700[diff] [blame]554 "src/math/exp-avx512f-p5-scalef.c",
555 "src/math/exp-avx512f-p5.c",
556 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -0700[diff] [blame]557 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700[diff] [blame]558 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]559]
560
561AARCH32_ASM_UKERNELS = [
562 "src/q8-dwconv/up8x9-aarch32-neon.S",
563]
564
565AARCH64_ASM_UKERNELS = [
566 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
567 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
568 "src/f32-gemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700[diff] [blame]569 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]570 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a57.S",
571 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a75.S",
572 "src/f32-gemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700[diff] [blame]573 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]574 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a57.S",
575 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a75.S",
576 "src/f32-gemm/4x8-aarch64-neonfma-ld128.S",
577 "src/f32-gemm/4x8-aarch64-neonfma-ld64.S",
578 "src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700[diff] [blame]579 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]580 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a57.S",
581 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S",
582 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S",
583 "src/f32-gemm/6x8-aarch64-neonfma-ld128.S",
584 "src/f32-gemm/6x8-aarch64-neonfma-ld64.S",
585 "src/f32-gemminc/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700[diff] [blame]586 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]587 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a57.S",
588 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a75.S",
589 "src/f32-gemminc/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700[diff] [blame]590 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]591 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a57.S",
592 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a75.S",
593 "src/f32-gemminc/4x8-aarch64-neonfma-ld128.S",
594 "src/f32-gemminc/4x8-aarch64-neonfma-ld64.S",
595 "src/f32-gemminc/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700[diff] [blame]596 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]597 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a57.S",
598 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a73.S",
599 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a75.S",
600 "src/f32-gemminc/6x8-aarch64-neonfma-ld128.S",
601 "src/f32-gemminc/6x8-aarch64-neonfma-ld64.S",
602 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700[diff] [blame]603 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]604 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S",
605 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S",
606 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700[diff] [blame]607 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]608 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
609 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700[diff] [blame]610 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]611 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
612 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
613 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
614]
615
616INTERNAL_MICROKERNEL_HDRS = [
617 "src/xnnpack/argmaxpool.h",
618 "src/xnnpack/avgpool.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800[diff] [blame]619 "src/xnnpack/bilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]620 "src/xnnpack/clamp.h",
621 "src/xnnpack/common.h",
622 "src/xnnpack/conv.h",
623 "src/xnnpack/dwconv.h",
624 "src/xnnpack/gavgpool.h",
625 "src/xnnpack/gemm.h",
626 "src/xnnpack/hswish.h",
627 "src/xnnpack/igemm.h",
628 "src/xnnpack/lut.h",
629 "src/xnnpack/math.h",
630 "src/xnnpack/maxpool.h",
631 "src/xnnpack/packx.h",
632 "src/xnnpack/pad.h",
633 "src/xnnpack/params.h",
634 "src/xnnpack/pavgpool.h",
635 "src/xnnpack/ppmm.h",
636 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700[diff] [blame]637 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700[diff] [blame]638 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700[diff] [blame]639 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]640 "src/xnnpack/rmax.h",
641 "src/xnnpack/scalar-utils.h",
642 "src/xnnpack/spmm.h",
643 "src/xnnpack/unpool.h",
644 "src/xnnpack/vadd.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -0800[diff] [blame]645 "src/xnnpack/vbinop.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]646 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700[diff] [blame]647 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700[diff] [blame]648 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700[diff] [blame]649 "src/xnnpack/vscaleextexp.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800[diff] [blame]650 "src/xnnpack/vunop.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]651 "src/xnnpack/zip.h",
652]
653
654INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
655 "include/xnnpack.h",
656 "src/xnnpack/allocator.h",
657 "src/xnnpack/compute.h",
658 "src/xnnpack/im2col.h",
659 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700[diff] [blame]660 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]661 "src/xnnpack/operator.h",
662 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700[diff] [blame]663 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]664 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700[diff] [blame]665 "src/xnnpack/requantization.h",
666]
667
668ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
669 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]670]
671
672MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700[diff] [blame]673 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]674 "include/xnnpack.h",
675]
676
677MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
678 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700[diff] [blame]679 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]680 "src/xnnpack/requantization.h",
681 "include/xnnpack.h",
682]
683
684OPERATOR_TEST_PARAMS_HDRS = [
685 "src/xnnpack/params.h",
686 "src/xnnpack/common.h",
687]
688
689WEIGHTS_PACK_HDRS = [
690 "src/xnnpack/pack.h",
691 "src/xnnpack/operator.h",
692 "src/xnnpack/compute.h",
693]
694
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700[diff] [blame]695LOGGING_COPTS = select({
696 # No logging in optimized mode
697 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
698 # Full logging in debug mode
699 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
700 # Error-only logging in default (fastbuild) mode
701 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
702})
703
704LOGGING_HDRS = [
705 "src/xnnpack/log.h",
706]
707
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]708xnnpack_cc_library(
709 name = "scalar_ukernels",
710 srcs = SCALAR_UKERNELS,
711 hdrs = INTERNAL_HDRS,
712 aarch32_copts = ["-marm"],
713 copts = xnnpack_std_copts(),
714 deps = [
715 "@FP16",
716 "@FXdiv",
717 ],
718)
719
720xnnpack_cc_library(
721 name = "psimd_ukernels",
722 srcs = PSIMD_UKERNELS,
723 hdrs = INTERNAL_HDRS,
724 aarch32_copts = [
725 "-marm",
726 "-mfpu=neon",
727 ],
728 copts = xnnpack_std_copts(),
729 optimized_copts = [
730 "-O3",
731 "-ffast-math",
732 ],
733 deps = [
734 "@FP16",
735 "@psimd",
736 ],
737)
738
739xnnpack_cc_library(
740 name = "neon_ukernels",
741 hdrs = INTERNAL_HDRS,
742 aarch32_copts = [
743 "-marm",
744 "-mfpu=neon",
745 ],
746 aarch32_srcs = NEON_UKERNELS,
747 aarch64_srcs = NEON_UKERNELS,
748 copts = xnnpack_std_copts(),
749 deps = ["@FP16"],
750)
751
752xnnpack_cc_library(
753 name = "neonfma_ukernels",
754 hdrs = INTERNAL_HDRS,
755 aarch32_copts = [
756 "-marm",
757 "-mfpu=neon-vfpv4",
758 ],
759 aarch32_srcs = NEONFMA_UKERNELS,
760 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
761 copts = xnnpack_std_copts(),
762 deps = ["@FP16"],
763)
764
765xnnpack_cc_library(
766 name = "neonfp16arith_ukernels",
767 hdrs = INTERNAL_HDRS,
768 aarch64_copts = ["-march=armv8.2-a+fp16"],
769 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
770 copts = xnnpack_std_copts(),
771 deps = ["@FP16"],
772)
773
774xnnpack_cc_library(
775 name = "sse2_ukernels",
776 hdrs = INTERNAL_HDRS,
777 copts = xnnpack_std_copts(),
778 x86_copts = ["-msse2"],
779 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
780 deps = ["@FP16"],
781)
782
783xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800[diff] [blame]784 name = "sse41_ukernels",
785 hdrs = INTERNAL_HDRS,
786 copts = xnnpack_std_copts(),
787 x86_copts = ["-msse4.1"],
788 x86_srcs = SSE41_UKERNELS,
789 deps = ["@FP16"],
790)
791
792xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]793 name = "avx_ukernels",
794 hdrs = INTERNAL_HDRS,
795 copts = xnnpack_std_copts(),
796 x86_copts = ["-mavx"],
797 x86_srcs = AVX_UKERNELS,
798 deps = ["@FP16"],
799)
800
801xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700[diff] [blame]802 name = "avx2_ukernels",
803 hdrs = INTERNAL_HDRS,
804 copts = xnnpack_std_copts(),
805 x86_copts = [
806 "-mfma",
807 "-mavx2",
808 ],
809 x86_srcs = AVX2_UKERNELS,
810 deps = ["@FP16"],
811)
812
813xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]814 name = "avx512f_ukernels",
815 hdrs = INTERNAL_HDRS,
816 copts = xnnpack_std_copts(),
817 x86_copts = ["-mavx512f"],
818 x86_srcs = AVX512F_UKERNELS,
819 deps = ["@FP16"],
820)
821
822xnnpack_cc_library(
823 name = "asm_ukernels",
824 hdrs = ["src/xnnpack/assembly.h"],
825 aarch32_srcs = AARCH32_ASM_UKERNELS,
826 aarch64_srcs = AARCH64_ASM_UKERNELS,
827)
828
829xnnpack_aggregate_library(
830 name = "ukernels",
831 aarch32_deps = [
832 ":psimd_ukernels",
833 ":neon_ukernels",
834 ":neonfma_ukernels",
835 ":asm_ukernels",
836 ],
837 aarch64_deps = [
838 ":psimd_ukernels",
839 ":neon_ukernels",
840 ":neonfma_ukernels",
841 ":neonfp16arith_ukernels",
842 ":asm_ukernels",
843 ],
844 generic_deps = [":scalar_ukernels"],
845 wasmsimd_deps = [
846 ":psimd_ukernels",
847 ],
848 x86_deps = [
849 ":psimd_ukernels",
850 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800[diff] [blame]851 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]852 ":avx_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700[diff] [blame]853 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]854 ":avx512f_ukernels",
855 ],
856)
857
858xnnpack_cc_library(
859 name = "im2col",
860 srcs = ["src/im2col.c"],
861 hdrs = [
862 "src/xnnpack/common.h",
863 "src/xnnpack/im2col.h",
864 ],
865 copts = xnnpack_std_copts(),
866)
867
868xnnpack_cc_library(
869 name = "indirection",
870 srcs = ["src/indirection.c"],
871 hdrs = INTERNAL_HDRS,
872 copts = xnnpack_std_copts(),
873 deps = [
874 "@FP16",
875 "@FXdiv",
876 "@pthreadpool",
877 ],
878)
879
880xnnpack_cc_library(
881 name = "operator_run",
882 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700[diff] [blame]883 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
884 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]885 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
886 "-Wno-vla",
887 ],
888 deps = [
889 "@FP16",
890 "@FXdiv",
891 "@clog",
892 "@pthreadpool",
893 ],
894)
895
896cc_library(
897 name = "enable_assembly",
898 defines = select({
899 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
900 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -0700[diff] [blame]901 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]902 }),
903)
904
Marat Dukhancf056b22019-10-07 10:26:29 -0700[diff] [blame]905xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]906 name = "operators",
907 srcs = OPERATOR_SRCS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]908 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -0700[diff] [blame]909 ],
910 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700[diff] [blame]911 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]912 "-Isrc",
913 "-Iinclude",
914 ] + select({
915 ":debug_build": [],
916 "//conditions:default": xnnpack_min_size_copts(),
917 }),
Marat Dukhancf056b22019-10-07 10:26:29 -0700[diff] [blame]918 wasm_srcs = ["src/wasm-stubs.c"],
919 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]920 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]921 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]922 "@FP16",
923 "@FXdiv",
924 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]925 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700[diff] [blame]926 ],
927)
928
929cc_library(
930 name = "XNNPACK",
931 srcs = [
932 "src/init.c",
933 ],
934 copts = xnnpack_std_copts() + LOGGING_COPTS + [
935 "-Isrc",
936 "-Iinclude",
937 ] + select({
938 ":debug_build": [],
939 "//conditions:default": xnnpack_min_size_copts(),
940 }),
941 includes = ["include"],
942 linkstatic = True,
943 textual_hdrs = ["include/xnnpack.h"],
944 visibility = xnnpack_visibility(),
945 deps = [
946 ":enable_assembly",
947 ":ukernels",
948 ":operator_run",
949 ":operators",
950 "@clog",
951 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -0700[diff] [blame]952 ] + select({
953 ":emscripten": [],
954 "//conditions:default": ["@cpuinfo"],
955 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]956)
957
958cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700[diff] [blame]959 name = "xnnpack_operators_nhwc_f32",
960 srcs = [
961 "src/init.c",
962 ],
963 copts = xnnpack_std_copts() + LOGGING_COPTS + [
964 "-Isrc",
965 "-Iinclude",
966 ] + select({
967 ":debug_build": [],
968 "//conditions:default": xnnpack_min_size_copts(),
969 }),
970 defines = [
971 "XNN_NO_Q8_OPERATORS",
972 "XNN_NO_U8_OPERATORS",
973 "XNN_NO_X8_OPERATORS",
974 "XNN_NO_SPNCHW_OPERATORS",
975 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]976 includes = ["include"],
977 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700[diff] [blame]978 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]979 visibility = xnnpack_visibility(),
980 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700[diff] [blame]981 ":enable_assembly",
982 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]983 ":operator_run",
984 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700[diff] [blame]985 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]986 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700[diff] [blame]987 ] + select({
988 ":emscripten": [],
989 "//conditions:default": ["@cpuinfo"],
990 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]991)
992
Marat Dukhancf056b22019-10-07 10:26:29 -0700[diff] [blame]993xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]994 name = "bench_utils",
995 srcs = ["bench/utils.cc"],
996 hdrs = ["bench/utils.h"],
997 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -0800[diff] [blame]998 deps = [
999 "@com_google_benchmark//:benchmark",
1000 "@cpuinfo",
1001 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]1002)
1003
Frank Barchard7e955972019-10-11 10:34:25 -0700[diff] [blame]1004######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]1005
1006xnnpack_benchmark(
1007 name = "q8_gemm_bench",
1008 srcs = [
1009 "bench/gemm.h",
1010 "bench/q8-gemm.cc",
1011 "src/xnnpack/AlignedAllocator.h",
1012 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1013 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
1014 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
1015)
1016
1017xnnpack_benchmark(
1018 name = "f16_gemm_bench",
1019 srcs = [
1020 "bench/f16-gemm.cc",
1021 "bench/gemm.h",
1022 "src/xnnpack/AlignedAllocator.h",
1023 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1024 copts = ["-Wno-unused-function"],
1025 deps = MICROKERNEL_BENCHMARK_DEPS,
1026)
1027
1028xnnpack_benchmark(
1029 name = "f32_igemm_bench",
1030 srcs = [
1031 "bench/f32-igemm.cc",
1032 "bench/conv.h",
1033 "src/xnnpack/AlignedAllocator.h",
1034 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -0700[diff] [blame]1035 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]1036)
1037
1038xnnpack_benchmark(
1039 name = "f32_conv_hwc_bench",
1040 srcs = [
1041 "bench/f32-conv-hwc.cc",
1042 "bench/dconv.h",
1043 "src/xnnpack/AlignedAllocator.h",
1044 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1045 copts = ["-Wno-unused-function"],
1046 deps = MICROKERNEL_BENCHMARK_DEPS,
1047)
1048
1049xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -0700[diff] [blame]1050 name = "f32_conv_hwc2spchw_bench",
1051 srcs = [
1052 "bench/f32-conv-hwc2spchw.cc",
1053 "bench/dconv.h",
1054 "src/xnnpack/AlignedAllocator.h",
1055 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1056 copts = ["-Wno-unused-function"],
1057 deps = MICROKERNEL_BENCHMARK_DEPS,
1058)
1059
1060xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]1061 name = "f32_dwconv_bench",
1062 srcs = [
1063 "bench/f32-dwconv.cc",
1064 "bench/dwconv.h",
1065 "src/xnnpack/AlignedAllocator.h",
1066 ] + MICROKERNEL_BENCHMARK_HDRS,
1067 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1068)
1069
1070xnnpack_benchmark(
1071 name = "f32_dwconv_spchw_bench",
1072 srcs = [
1073 "bench/f32-dwconv-spchw.cc",
1074 "bench/dwconv.h",
1075 "src/xnnpack/AlignedAllocator.h",
1076 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1077 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1078)
1079
1080xnnpack_benchmark(
1081 name = "f32_gemm_bench",
1082 srcs = [
1083 "bench/f32-gemm.cc",
1084 "bench/gemm.h",
1085 "src/xnnpack/AlignedAllocator.h",
1086 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1087 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -0700[diff] [blame]1088 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]1089)
1090
1091xnnpack_benchmark(
1092 name = "f32_rmax_bench",
1093 srcs = [
1094 "bench/f32-rmax.cc",
1095 "src/xnnpack/AlignedAllocator.h",
1096 ] + MICROKERNEL_BENCHMARK_HDRS,
1097 deps = MICROKERNEL_BENCHMARK_DEPS,
1098)
1099
1100xnnpack_benchmark(
1101 name = "f32_spmm_bench",
1102 srcs = [
1103 "bench/f32-spmm.cc",
1104 "bench/gemm.h",
1105 "src/xnnpack/AlignedAllocator.h",
1106 ] + MICROKERNEL_BENCHMARK_HDRS,
1107 copts = ["-Wno-unused-function"],
1108 deps = MICROKERNEL_BENCHMARK_DEPS,
1109)
1110
1111xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -0700[diff] [blame]1112 name = "f32_softargmax_bench",
1113 srcs = [
1114 "bench/f32-softargmax.cc",
1115 ] + MICROKERNEL_BENCHMARK_HDRS,
1116 copts = ["-Wno-unused-function"],
1117 deps = MICROKERNEL_BENCHMARK_DEPS,
1118)
1119
1120xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]1121 name = "f32_im2col_gemm_bench",
1122 srcs = [
1123 "bench/f32-im2col-gemm.cc",
1124 "bench/conv.h",
1125 "src/xnnpack/AlignedAllocator.h",
1126 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1127 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
1128)
1129
1130########################### Benchmarks for operators ###########################
1131
1132xnnpack_benchmark(
1133 name = "add_bench",
1134 srcs = ["bench/add.cc"],
1135 deps = OPERATOR_BENCHMARK_DEPS,
1136)
1137
1138xnnpack_benchmark(
1139 name = "average_pooling_bench",
1140 srcs = ["bench/average-pooling.cc"],
1141 deps = OPERATOR_BENCHMARK_DEPS,
1142)
1143
1144xnnpack_benchmark(
1145 name = "channel_shuffle_bench",
1146 srcs = ["bench/channel-shuffle.cc"],
1147 deps = OPERATOR_BENCHMARK_DEPS,
1148)
1149
1150xnnpack_benchmark(
1151 name = "convolution_bench",
1152 srcs = ["bench/convolution.cc"],
1153 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1154 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1155)
1156
1157xnnpack_benchmark(
1158 name = "deconvolution_bench",
1159 srcs = ["bench/deconvolution.cc"],
1160 copts = xnnpack_optional_tflite_copts(),
1161 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1162)
1163
1164xnnpack_benchmark(
1165 name = "global_average_pooling_bench",
1166 srcs = ["bench/global-average-pooling.cc"],
1167 deps = OPERATOR_BENCHMARK_DEPS,
1168)
1169
1170xnnpack_benchmark(
1171 name = "max_pooling_bench",
1172 srcs = ["bench/max-pooling.cc"],
1173 deps = OPERATOR_BENCHMARK_DEPS,
1174)
1175
1176xnnpack_benchmark(
1177 name = "sigmoid_bench",
1178 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -0800[diff] [blame^]1179 copts = xnnpack_optional_tflite_copts(),
1180 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]1181)
1182
1183xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -0700[diff] [blame]1184 name = "prelu_bench",
1185 srcs = ["bench/prelu.cc"],
1186 copts = xnnpack_optional_tflite_copts(),
1187 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1188)
1189
1190xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]1191 name = "softargmax_bench",
1192 srcs = ["bench/softargmax.cc"],
1193 deps = OPERATOR_BENCHMARK_DEPS,
1194)
1195
Marat Dukhanc068bb62019-10-04 13:24:39 -0700[diff] [blame]1196############################# End-to-end benchmarks ############################
1197
1198cc_library(
1199 name = "mobilenet_v1",
1200 srcs = ["models/mobilenet-v1.cc"],
1201 hdrs = ["models/models.h"],
1202 linkstatic = True,
1203 deps = [
1204 ":XNNPACK",
1205 "@pthreadpool",
1206 ],
1207)
1208
1209cc_library(
1210 name = "mobilenet_v2",
1211 srcs = ["models/mobilenet-v2.cc"],
1212 hdrs = ["models/models.h"],
1213 linkstatic = True,
1214 deps = [
1215 ":XNNPACK",
1216 "@pthreadpool",
1217 ],
1218)
1219
1220xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -0700[diff] [blame]1221 name = "f32_dwconv_e2e_bench",
1222 srcs = ["bench/f32-dwconv-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1223 copts = ["-Wno-unused-function"],
1224 deps = MICROKERNEL_BENCHMARK_DEPS + [
1225 ":XNNPACK",
1226 ":mobilenet_v1",
1227 ":mobilenet_v2",
1228 ],
1229)
1230
1231xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -0700[diff] [blame]1232 name = "f32_gemm_e2e_bench",
1233 srcs = ["bench/f32-gemm-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1234 copts = ["-Wno-unused-function"],
1235 deps = MICROKERNEL_BENCHMARK_DEPS + [
1236 ":XNNPACK",
1237 ":mobilenet_v1",
1238 ":mobilenet_v2",
1239 ],
1240)
1241
1242xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -0700[diff] [blame]1243 name = "end2end_bench",
1244 srcs = ["bench/end2end.cc"],
1245 deps = [
1246 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -0700[diff] [blame]1247 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -0700[diff] [blame]1248 ":mobilenet_v1",
1249 ":mobilenet_v2",
1250 "@pthreadpool",
1251 ],
1252)
1253
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700[diff] [blame]1254#################### Accuracy evaluation for math functions ####################
1255
1256xnnpack_benchmark(
1257 name = "f32_exp_eval",
1258 srcs = [
1259 "eval/f32-exp.cc",
1260 "src/xnnpack/AlignedAllocator.h",
1261 ] + ACCURACY_EVAL_HDRS,
1262 deps = ACCURACY_EVAL_DEPS,
1263)
1264
Marat Dukhan515c9772019-10-17 18:07:57 -0700[diff] [blame]1265xnnpack_benchmark(
1266 name = "f32_expminus_eval",
1267 srcs = [
1268 "eval/f32-expminus.cc",
1269 "src/xnnpack/AlignedAllocator.h",
1270 ] + ACCURACY_EVAL_HDRS,
1271 deps = ACCURACY_EVAL_DEPS,
1272)
1273
Marat Dukhan98ba4412019-10-23 02:14:28 -0700[diff] [blame]1274xnnpack_benchmark(
1275 name = "f32_extexp_eval",
1276 srcs = [
1277 "eval/f32-extexp.cc",
1278 "src/xnnpack/AlignedAllocator.h",
1279 ] + ACCURACY_EVAL_HDRS,
1280 deps = ACCURACY_EVAL_DEPS,
1281)
1282
Marat Dukhan346a9e52019-11-15 09:06:30 -0800[diff] [blame]1283xnnpack_benchmark(
1284 name = "f32_sigmoid_eval",
1285 srcs = [
1286 "eval/f32-sigmoid.cc",
1287 "src/xnnpack/AlignedAllocator.h",
1288 ] + ACCURACY_EVAL_HDRS,
1289 deps = ACCURACY_EVAL_DEPS,
1290)
1291
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]1292######################### Unit tests for micro-kernels #########################
1293
1294xnnpack_unit_test(
1295 name = "f16_gemm_test",
1296 srcs = [
1297 "test/f16-gemm.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_argmaxpool_test",
1306 srcs = [
1307 "test/f32-argmaxpool.cc",
1308 "test/argmaxpool-microkernel-tester.h",
1309 "src/xnnpack/AlignedAllocator.h",
1310 ] + MICROKERNEL_TEST_HDRS,
1311 deps = MICROKERNEL_TEST_DEPS,
1312)
1313
1314xnnpack_unit_test(
1315 name = "f32_avgpool_test",
1316 srcs = [
1317 "test/f32-avgpool.cc",
1318 "test/avgpool-microkernel-tester.h",
1319 "src/xnnpack/AlignedAllocator.h",
1320 ] + MICROKERNEL_TEST_HDRS,
1321 deps = MICROKERNEL_TEST_DEPS,
1322)
1323
1324xnnpack_unit_test(
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800[diff] [blame]1325 name = "f32_bilinear_test",
1326 srcs = [
1327 "test/f32-bilinear.cc",
1328 "test/bilinear-microkernel-tester.h",
1329 "src/xnnpack/AlignedAllocator.h",
1330 ] + MICROKERNEL_TEST_HDRS,
1331 deps = MICROKERNEL_TEST_DEPS,
1332)
1333
1334xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]1335 name = "f32_clamp_test",
1336 srcs = [
1337 "test/f32-clamp.cc",
1338 "test/clamp-microkernel-tester.h",
1339 ] + MICROKERNEL_TEST_HDRS,
1340 deps = MICROKERNEL_TEST_DEPS,
1341)
1342
1343xnnpack_unit_test(
1344 name = "f32_igemm_test",
1345 srcs = [
1346 "test/f32-igemm.cc",
1347 "test/gemm-microkernel-tester.h",
1348 "src/xnnpack/AlignedAllocator.h",
1349 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1350 deps = MICROKERNEL_TEST_DEPS,
1351)
1352
1353xnnpack_unit_test(
1354 name = "f32_conv_hwc_test",
1355 srcs = [
1356 "test/f32-conv-hwc.cc",
1357 "test/conv-hwc-microkernel-tester.h",
1358 "src/xnnpack/AlignedAllocator.h",
1359 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1360 deps = MICROKERNEL_TEST_DEPS,
1361)
1362
1363xnnpack_unit_test(
1364 name = "f32_conv_hwc2spchw_test",
1365 srcs = [
1366 "test/f32-conv-hwc2spchw.cc",
1367 "test/conv-hwc2spchw-microkernel-tester.h",
1368 "src/xnnpack/AlignedAllocator.h",
1369 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1370 deps = MICROKERNEL_TEST_DEPS,
1371)
1372
1373xnnpack_unit_test(
1374 name = "f32_dwconv_test",
1375 srcs = [
1376 "test/f32-dwconv.cc",
1377 "test/dwconv-microkernel-tester.h",
1378 "src/xnnpack/AlignedAllocator.h",
1379 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1380 deps = MICROKERNEL_TEST_DEPS,
1381)
1382
1383xnnpack_unit_test(
1384 name = "f32_dwconv_spchw_test",
1385 srcs = [
1386 "test/f32-dwconv-spchw.cc",
1387 "test/dwconv-spchw-microkernel-tester.h",
1388 "src/xnnpack/AlignedAllocator.h",
1389 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1390 deps = MICROKERNEL_TEST_DEPS,
1391)
1392
1393xnnpack_unit_test(
1394 name = "f32_gavgpool_test",
1395 srcs = [
1396 "test/f32-gavgpool.cc",
1397 "test/gavgpool-microkernel-tester.h",
1398 "src/xnnpack/AlignedAllocator.h",
1399 ] + MICROKERNEL_TEST_HDRS,
1400 deps = MICROKERNEL_TEST_DEPS,
1401)
1402
1403xnnpack_unit_test(
1404 name = "f32_gavgpool_spchw_test",
1405 srcs = [
1406 "test/f32-gavgpool-spchw.cc",
1407 "test/gavgpool-spchw-microkernel-tester.h",
1408 "src/xnnpack/AlignedAllocator.h",
1409 ] + MICROKERNEL_TEST_HDRS,
1410 deps = MICROKERNEL_TEST_DEPS,
1411)
1412
1413xnnpack_unit_test(
1414 name = "f32_gemm_test",
1415 srcs = [
1416 "test/f32-gemm.cc",
1417 "test/gemm-microkernel-tester.h",
1418 "src/xnnpack/AlignedAllocator.h",
1419 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1420 deps = MICROKERNEL_TEST_DEPS,
1421)
1422
1423xnnpack_unit_test(
1424 name = "f32_gemminc_test",
1425 srcs = [
1426 "test/f32-gemminc.cc",
1427 "test/gemm-microkernel-tester.h",
1428 "src/xnnpack/AlignedAllocator.h",
1429 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1430 deps = MICROKERNEL_TEST_DEPS,
1431)
1432
1433xnnpack_unit_test(
1434 name = "f32_hswish_test",
1435 srcs = [
1436 "test/f32-hswish.cc",
1437 "test/hswish-microkernel-tester.h",
1438 ] + MICROKERNEL_TEST_HDRS,
1439 deps = MICROKERNEL_TEST_DEPS,
1440)
1441
1442xnnpack_unit_test(
1443 name = "f32_maxpool_test",
1444 srcs = [
1445 "test/f32-maxpool.cc",
1446 "test/maxpool-microkernel-tester.h",
1447 ] + MICROKERNEL_TEST_HDRS,
1448 deps = MICROKERNEL_TEST_DEPS,
1449)
1450
1451xnnpack_unit_test(
1452 name = "f32_pavgpool_test",
1453 srcs = [
1454 "test/f32-pavgpool.cc",
1455 "test/avgpool-microkernel-tester.h",
1456 "src/xnnpack/AlignedAllocator.h",
1457 ] + MICROKERNEL_TEST_HDRS,
1458 deps = MICROKERNEL_TEST_DEPS,
1459)
1460
1461xnnpack_unit_test(
1462 name = "f32_ppmm_test",
1463 srcs = [
1464 "test/f32-ppmm.cc",
1465 "test/gemm-microkernel-tester.h",
1466 "src/xnnpack/AlignedAllocator.h",
1467 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1468 deps = MICROKERNEL_TEST_DEPS,
1469)
1470
1471xnnpack_unit_test(
1472 name = "f32_prelu_test",
1473 srcs = [
1474 "test/f32-prelu.cc",
1475 "test/prelu-microkernel-tester.h",
1476 "src/xnnpack/AlignedAllocator.h",
1477 ] + MICROKERNEL_TEST_HDRS,
1478 deps = MICROKERNEL_TEST_DEPS,
1479)
1480
1481xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -0700[diff] [blame]1482 name = "f32_raddexpminusmax_test",
1483 srcs = [
1484 "test/f32-raddexpminusmax.cc",
1485 "test/raddexpminusmax-microkernel-tester.h",
1486 ] + MICROKERNEL_TEST_HDRS,
1487 deps = MICROKERNEL_TEST_DEPS,
1488)
1489
1490xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700[diff] [blame]1491 name = "f32_raddextexp_test",
1492 srcs = [
1493 "test/f32-raddextexp.cc",
1494 "test/raddextexp-microkernel-tester.h",
1495 ] + MICROKERNEL_TEST_HDRS,
1496 deps = MICROKERNEL_TEST_DEPS,
1497)
1498
1499xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -0700[diff] [blame]1500 name = "f32_raddstoreexpminusmax_test",
1501 srcs = [
1502 "test/f32-raddstoreexpminusmax.cc",
1503 "test/raddstoreexpminusmax-microkernel-tester.h",
1504 ] + MICROKERNEL_TEST_HDRS,
1505 deps = MICROKERNEL_TEST_DEPS,
1506)
1507
1508xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]1509 name = "f32_rmax_test",
1510 srcs = [
1511 "test/f32-rmax.cc",
1512 "test/rmax-microkernel-tester.h",
1513 ] + MICROKERNEL_TEST_HDRS,
1514 deps = MICROKERNEL_TEST_DEPS,
1515)
1516
1517xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -0800[diff] [blame]1518 name = "f32_sigmoid_test",
1519 srcs = [
1520 "test/f32-sigmoid.cc",
1521 "test/vunop-microkernel-tester.h",
1522 ] + MICROKERNEL_TEST_HDRS,
1523 deps = MICROKERNEL_TEST_DEPS,
1524)
1525
1526xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]1527 name = "f32_spmm_test",
1528 srcs = [
1529 "test/f32-spmm.cc",
1530 "test/spmm-microkernel-tester.h",
1531 "src/xnnpack/AlignedAllocator.h",
1532 ] + MICROKERNEL_TEST_HDRS,
1533 deps = MICROKERNEL_TEST_DEPS,
1534)
1535
1536xnnpack_unit_test(
1537 name = "f32_vadd_test",
1538 srcs = [
1539 "test/f32-vadd.cc",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -0800[diff] [blame]1540 "test/vbinop-microkernel-tester.h",
1541 ] + MICROKERNEL_TEST_HDRS,
1542 deps = MICROKERNEL_TEST_DEPS,
1543)
1544
1545xnnpack_unit_test(
1546 name = "f32_vaddc_test",
1547 srcs = [
1548 "test/f32-vaddc.cc",
1549 "test/vbinopc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]1550 ] + MICROKERNEL_TEST_HDRS,
1551 deps = MICROKERNEL_TEST_DEPS,
1552)
1553
1554xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]1555 name = "f32_vmul_test",
1556 srcs = [
1557 "test/f32-vmul.cc",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -0800[diff] [blame]1558 "test/vbinop-microkernel-tester.h",
1559 ] + MICROKERNEL_TEST_HDRS,
1560 deps = MICROKERNEL_TEST_DEPS,
1561)
1562
1563xnnpack_unit_test(
1564 name = "f32_vmulc_test",
1565 srcs = [
1566 "test/f32-vmulc.cc",
1567 "test/vbinopc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]1568 ] + MICROKERNEL_TEST_HDRS,
1569 deps = MICROKERNEL_TEST_DEPS,
1570)
1571
1572xnnpack_unit_test(
1573 name = "f32_vmulcaddc_test",
1574 srcs = [
1575 "test/f32-vmulcaddc.cc",
1576 "test/vmulcaddc-microkernel-tester.h",
1577 "src/xnnpack/AlignedAllocator.h",
1578 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1579 deps = MICROKERNEL_TEST_DEPS,
1580)
1581
1582xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700[diff] [blame]1583 name = "f32_vscale_test",
1584 srcs = [
1585 "test/f32-vscale.cc",
1586 "test/vscale-microkernel-tester.h",
1587 ] + MICROKERNEL_TEST_HDRS,
1588 deps = MICROKERNEL_TEST_DEPS,
1589)
1590
1591xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -0700[diff] [blame]1592 name = "f32_vscaleexpminusmax_test",
1593 srcs = [
1594 "test/f32-vscaleexpminusmax.cc",
1595 "test/vscaleexpminusmax-microkernel-tester.h",
1596 ] + MICROKERNEL_TEST_HDRS,
1597 deps = MICROKERNEL_TEST_DEPS,
1598)
1599
1600xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700[diff] [blame]1601 name = "f32_vscaleextexp_test",
1602 srcs = [
1603 "test/f32-vscaleextexp.cc",
1604 "test/vscaleextexp-microkernel-tester.h",
1605 ] + MICROKERNEL_TEST_HDRS,
1606 deps = MICROKERNEL_TEST_DEPS,
1607)
1608
1609xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -0700[diff] [blame]1610 name = "f32_vsub_test",
1611 srcs = [
1612 "test/f32-vsub.cc",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -0800[diff] [blame]1613 "test/vbinop-microkernel-tester.h",
1614 ] + MICROKERNEL_TEST_HDRS,
1615 deps = MICROKERNEL_TEST_DEPS,
1616)
1617
1618xnnpack_unit_test(
1619 name = "f32_vsubc_test",
1620 srcs = [
1621 "test/f32-vsubc.cc",
1622 "test/vbinopc-microkernel-tester.h",
1623 ] + MICROKERNEL_TEST_HDRS,
1624 deps = MICROKERNEL_TEST_DEPS,
1625)
1626
1627xnnpack_unit_test(
1628 name = "f32_vrsubc_test",
1629 srcs = [
1630 "test/f32-vrsubc.cc",
1631 "test/vbinopc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700[diff] [blame]1632 ] + MICROKERNEL_TEST_HDRS,
1633 deps = MICROKERNEL_TEST_DEPS,
1634)
1635
1636xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]1637 name = "q8_avgpool_test",
1638 srcs = [
1639 "test/q8-avgpool.cc",
1640 "test/avgpool-microkernel-tester.h",
1641 "src/xnnpack/AlignedAllocator.h",
1642 ] + MICROKERNEL_TEST_HDRS,
1643 deps = MICROKERNEL_TEST_DEPS,
1644)
1645
1646xnnpack_unit_test(
1647 name = "q8_igemm_test",
1648 srcs = [
1649 "test/q8-igemm.cc",
1650 "test/gemm-microkernel-tester.h",
1651 "src/xnnpack/AlignedAllocator.h",
1652 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1653 deps = MICROKERNEL_TEST_DEPS,
1654)
1655
1656xnnpack_unit_test(
1657 name = "q8_dwconv_test",
1658 srcs = [
1659 "test/q8-dwconv.cc",
1660 "test/dwconv-microkernel-tester.h",
1661 "src/xnnpack/AlignedAllocator.h",
1662 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1663 deps = MICROKERNEL_TEST_DEPS,
1664)
1665
1666xnnpack_unit_test(
1667 name = "q8_gavgpool_test",
1668 srcs = [
1669 "test/q8-gavgpool.cc",
1670 "test/gavgpool-microkernel-tester.h",
1671 "src/xnnpack/AlignedAllocator.h",
1672 ] + MICROKERNEL_TEST_HDRS,
1673 deps = MICROKERNEL_TEST_DEPS,
1674)
1675
1676xnnpack_unit_test(
1677 name = "q8_gemm_test",
1678 srcs = [
1679 "test/q8-gemm.cc",
1680 "test/gemm-microkernel-tester.h",
1681 "src/xnnpack/AlignedAllocator.h",
1682 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1683 deps = MICROKERNEL_TEST_DEPS,
1684)
1685
1686xnnpack_unit_test(
1687 name = "q8_vadd_test",
1688 srcs = [
1689 "test/q8-vadd.cc",
1690 "test/vadd-microkernel-tester.h",
1691 ] + MICROKERNEL_TEST_HDRS,
1692 deps = MICROKERNEL_TEST_DEPS,
1693)
1694
1695xnnpack_unit_test(
1696 name = "u8_clamp_test",
1697 srcs = [
1698 "test/u8-clamp.cc",
1699 "test/clamp-microkernel-tester.h",
1700 ] + MICROKERNEL_TEST_HDRS,
1701 deps = MICROKERNEL_TEST_DEPS,
1702)
1703
1704xnnpack_unit_test(
1705 name = "u8_lut32norm_test",
1706 srcs = [
1707 "test/u8-lut32norm.cc",
1708 "test/lut-norm-microkernel-tester.h",
1709 ] + MICROKERNEL_TEST_HDRS,
1710 deps = MICROKERNEL_TEST_DEPS,
1711)
1712
1713xnnpack_unit_test(
1714 name = "u8_maxpool_test",
1715 srcs = [
1716 "test/u8-maxpool.cc",
1717 "test/maxpool-microkernel-tester.h",
1718 ] + MICROKERNEL_TEST_HDRS,
1719 deps = MICROKERNEL_TEST_DEPS,
1720)
1721
1722xnnpack_unit_test(
1723 name = "u8_rmax_test",
1724 srcs = [
1725 "test/u8-rmax.cc",
1726 "test/rmax-microkernel-tester.h",
1727 ] + MICROKERNEL_TEST_HDRS,
1728 deps = MICROKERNEL_TEST_DEPS,
1729)
1730
1731xnnpack_unit_test(
1732 name = "x32_packx_test",
1733 srcs = [
1734 "test/x32-packx.cc",
1735 "test/pack-microkernel-tester.h",
1736 "src/xnnpack/AlignedAllocator.h",
1737 ] + MICROKERNEL_TEST_HDRS,
1738 deps = MICROKERNEL_TEST_DEPS,
1739)
1740
1741xnnpack_unit_test(
1742 name = "x32_pad_test",
1743 srcs = [
1744 "test/x32-pad.cc",
1745 "test/pad-microkernel-tester.h",
1746 ] + MICROKERNEL_TEST_HDRS,
1747 deps = MICROKERNEL_TEST_DEPS,
1748)
1749
1750xnnpack_unit_test(
1751 name = "x32_unpool_test",
1752 srcs = [
1753 "test/x32-unpool.cc",
1754 "test/unpool-microkernel-tester.h",
1755 ] + MICROKERNEL_TEST_HDRS,
1756 deps = MICROKERNEL_TEST_DEPS,
1757)
1758
1759xnnpack_unit_test(
1760 name = "x32_zip_test",
1761 srcs = [
1762 "test/x32-zip.cc",
1763 "test/zip-microkernel-tester.h",
1764 ] + MICROKERNEL_TEST_HDRS,
1765 deps = MICROKERNEL_TEST_DEPS,
1766)
1767
1768xnnpack_unit_test(
1769 name = "x8_lut_test",
1770 srcs = [
1771 "test/x8-lut.cc",
1772 "test/lut-microkernel-tester.h",
1773 ] + MICROKERNEL_TEST_HDRS,
1774 deps = MICROKERNEL_TEST_DEPS,
1775)
1776
1777xnnpack_unit_test(
1778 name = "x8_zip_test",
1779 srcs = [
1780 "test/x8-zip.cc",
1781 "test/zip-microkernel-tester.h",
1782 ] + MICROKERNEL_TEST_HDRS,
1783 deps = MICROKERNEL_TEST_DEPS,
1784)
1785
1786########################### Size test for the library ##########################
1787
1788xnnpack_binary(
1789 name = "size_test",
1790 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700[diff] [blame]1791 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]1792)
1793
1794########################### Unit tests for operators ###########################
1795
1796xnnpack_unit_test(
1797 name = "add_test",
1798 srcs = [
1799 "test/add.cc",
1800 "test/add-operator-tester.h",
1801 ],
1802 deps = OPERATOR_TEST_DEPS,
1803)
1804
1805xnnpack_unit_test(
1806 name = "argmax_pooling_test",
1807 srcs = [
1808 "test/argmax-pooling.cc",
1809 "test/argmax-pooling-operator-tester.h",
1810 ] + OPERATOR_TEST_PARAMS_HDRS,
1811 deps = OPERATOR_TEST_DEPS,
1812)
1813
1814xnnpack_unit_test(
1815 name = "average_pooling_test",
1816 srcs = [
1817 "test/average-pooling.cc",
1818 "test/average-pooling-operator-tester.h",
1819 ] + OPERATOR_TEST_PARAMS_HDRS,
1820 deps = OPERATOR_TEST_DEPS,
1821)
1822
1823xnnpack_unit_test(
1824 name = "channel_pad_test",
1825 srcs = [
1826 "test/channel-pad.cc",
1827 "test/channel-pad-operator-tester.h",
1828 ] + OPERATOR_TEST_PARAMS_HDRS,
1829 deps = OPERATOR_TEST_DEPS,
1830)
1831
1832xnnpack_unit_test(
1833 name = "channel_shuffle_test",
1834 srcs = [
1835 "test/channel-shuffle.cc",
1836 "test/channel-shuffle-operator-tester.h",
1837 ],
1838 deps = OPERATOR_TEST_DEPS,
1839)
1840
1841xnnpack_unit_test(
1842 name = "clamp_test",
1843 srcs = [
1844 "test/clamp.cc",
1845 "test/clamp-operator-tester.h",
1846 ],
1847 deps = OPERATOR_TEST_DEPS,
1848)
1849
1850xnnpack_unit_test(
1851 name = "convolution_test",
1852 srcs = [
1853 "test/convolution.cc",
1854 "test/convolution-operator-tester.h",
1855 ],
1856 deps = OPERATOR_TEST_DEPS,
1857)
1858
1859xnnpack_unit_test(
1860 name = "convolution_spnchw_test",
1861 srcs = [
1862 "test/convolution-spnchw.cc",
1863 "test/convolution-spnchw-operator-tester.h",
1864 ],
1865 deps = OPERATOR_TEST_DEPS,
1866)
1867
1868xnnpack_unit_test(
1869 name = "deconvolution_test",
1870 srcs = [
1871 "test/deconvolution.cc",
1872 "test/deconvolution-operator-tester.h",
1873 ] + OPERATOR_TEST_PARAMS_HDRS,
1874 deps = OPERATOR_TEST_DEPS,
1875)
1876
1877xnnpack_unit_test(
1878 name = "fully_connected_test",
1879 srcs = [
1880 "test/fully-connected.cc",
1881 "test/fully-connected-operator-tester.h",
1882 ],
1883 deps = OPERATOR_TEST_DEPS,
1884)
1885
1886xnnpack_unit_test(
1887 name = "global_average_pooling_test",
1888 srcs = [
1889 "test/global-average-pooling.cc",
1890 "test/global-average-pooling-operator-tester.h",
1891 ] + OPERATOR_TEST_PARAMS_HDRS,
1892 deps = OPERATOR_TEST_DEPS,
1893)
1894
1895xnnpack_unit_test(
1896 name = "global_average_pooling_spnchw_test",
1897 srcs = [
1898 "test/global-average-pooling-spnchw.cc",
1899 "test/global-average-pooling-spnchw-operator-tester.h",
1900 ],
1901 deps = OPERATOR_TEST_DEPS,
1902)
1903
1904xnnpack_unit_test(
1905 name = "hardswish_test",
1906 srcs = [
1907 "test/hardswish.cc",
1908 "test/hardswish-operator-tester.h",
1909 ],
1910 deps = OPERATOR_TEST_DEPS,
1911)
1912
1913xnnpack_unit_test(
1914 name = "leaky_relu_test",
1915 srcs = [
1916 "test/leaky-relu.cc",
1917 "test/leaky-relu-operator-tester.h",
1918 ],
1919 deps = OPERATOR_TEST_DEPS,
1920)
1921
1922xnnpack_unit_test(
1923 name = "max_pooling_test",
1924 srcs = [
1925 "test/max-pooling.cc",
1926 "test/max-pooling-operator-tester.h",
1927 ] + OPERATOR_TEST_PARAMS_HDRS,
1928 deps = OPERATOR_TEST_DEPS,
1929)
1930
1931xnnpack_unit_test(
Marat Dukhanca2733c2019-11-15 23:21:17 -0800[diff] [blame]1932 name = "multiply_test",
1933 srcs = [
1934 "test/multiply.cc",
1935 "test/multiply-operator-tester.h",
1936 ],
1937 deps = OPERATOR_TEST_DEPS,
1938)
1939
1940xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]1941 name = "prelu_test",
1942 srcs = [
1943 "test/prelu.cc",
1944 "test/prelu-operator-tester.h",
1945 ] + OPERATOR_TEST_PARAMS_HDRS,
1946 deps = OPERATOR_TEST_DEPS,
1947)
1948
1949xnnpack_unit_test(
Marat Dukhan69722492019-11-11 19:55:50 -0800[diff] [blame]1950 name = "resize_bilinear_test",
1951 srcs = [
1952 "test/resize-bilinear.cc",
1953 "test/resize-bilinear-operator-tester.h",
1954 ] + OPERATOR_TEST_PARAMS_HDRS,
1955 deps = OPERATOR_TEST_DEPS,
1956)
1957
1958xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]1959 name = "sigmoid_test",
1960 srcs = [
1961 "test/sigmoid.cc",
1962 "test/sigmoid-operator-tester.h",
1963 ],
1964 deps = OPERATOR_TEST_DEPS,
1965)
1966
1967xnnpack_unit_test(
1968 name = "softargmax_test",
1969 srcs = [
1970 "test/softargmax.cc",
1971 "test/softargmax-operator-tester.h",
1972 ],
1973 deps = OPERATOR_TEST_DEPS,
1974)
1975
1976xnnpack_unit_test(
1977 name = "unpooling_test",
1978 srcs = [
1979 "test/unpooling.cc",
1980 "test/unpooling-operator-tester.h",
1981 ],
1982 deps = OPERATOR_TEST_DEPS,
1983)
1984
1985############################# Build configurations #############################
1986
Marat Dukhanb8642352019-10-30 15:43:02 -0700[diff] [blame]1987# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]1988config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -0700[diff] [blame]1989 name = "xnn_enable_assembly_explicit_true",
1990 define_values = {"xnn_enable_assembly": "true"},
1991)
1992
1993# Disables usage of assembly kernels.
1994config_setting(
1995 name = "xnn_enable_assembly_explicit_false",
1996 define_values = {"xnn_enable_assembly": "false"},
1997)
1998
1999# Builds with -c dbg
2000config_setting(
2001 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]2002 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -0700[diff] [blame]2003 "compilation_mode": "dbg",
2004 },
2005)
2006
2007# Builds with -c opt
2008config_setting(
2009 name = "optimized_build",
2010 values = {
2011 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]2012 },
2013)
2014
2015config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -0700[diff] [blame]2016 name = "linux_k8",
2017 values = {"cpu": "k8"},
2018)
2019
2020config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -0700[diff] [blame]2021 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -0700[diff] [blame]2022 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -0700[diff] [blame]2023)
2024
2025config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]2026 name = "android",
2027 values = {"crosstool_top": "//external:android/crosstool"},
2028)
2029
2030config_setting(
2031 name = "android_armv7",
2032 values = {
2033 "crosstool_top": "//external:android/crosstool",
2034 "cpu": "armeabi-v7a",
2035 },
2036)
2037
2038config_setting(
2039 name = "android_arm64",
2040 values = {
2041 "crosstool_top": "//external:android/crosstool",
2042 "cpu": "arm64-v8a",
2043 },
2044)
2045
2046config_setting(
2047 name = "android_x86",
2048 values = {
2049 "crosstool_top": "//external:android/crosstool",
2050 "cpu": "x86",
2051 },
2052)
2053
2054config_setting(
2055 name = "android_x86_64",
2056 values = {
2057 "crosstool_top": "//external:android/crosstool",
2058 "cpu": "x86_64",
2059 },
2060)
2061
2062config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -0700[diff] [blame]2063 name = "macos_x86_64",
2064 values = {
2065 "apple_platform_type": "macos",
2066 "cpu": "darwin",
2067 },
2068)
2069
2070config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]2071 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -0700[diff] [blame]2072 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]2073)
2074
2075config_setting(
2076 name = "emscripten_wasm",
2077 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -0700[diff] [blame]2078 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]2079 "cpu": "wasm",
2080 },
2081)
2082
2083config_setting(
2084 name = "emscripten_wasmsimd",
2085 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -0700[diff] [blame]2086 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]2087 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -0700[diff] [blame]2088 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]2089 },
2090)
2091
2092config_setting(
2093 name = "emscripten_asmjs",
2094 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -0700[diff] [blame]2095 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700[diff] [blame]2096 "cpu": "asmjs",
2097 },
2098)