blob: 25184d16d6abff805a84d457a14f8257c98d515d [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 Elsen38709a62019-11-08 11:58:45 -0800102 "src/f32-dwconv-spchw/5x5p2-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -0700103 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Erich Elsen38709a62019-11-08 11:58:45 -0800104 "src/f32-dwconv-spchw/5x5s2p2-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -0700105 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700106 "src/f32-gavgpool/mp7p7q-scalar.c",
107 "src/f32-gavgpool/up7-scalar.c",
108 "src/f32-gemm/1x4-scalar.c",
109 "src/f32-gemm/2x4-scalar.c",
110 "src/f32-gemm/4x2-scalar.c",
111 "src/f32-gemm/4x4-scalar.c",
112 "src/f32-gemminc/1x4-scalar.c",
113 "src/f32-gemminc/2x4-scalar.c",
114 "src/f32-gemminc/4x4-scalar.c",
115 "src/f32-hswish/scalar.c",
116 "src/f32-maxpool/9p8q-scalar.c",
117 "src/f32-pavgpool/mp9p8q-scalar.c",
118 "src/f32-pavgpool/up9-scalar.c",
119 "src/f32-ppmm/2x4-scalar.c",
120 "src/f32-ppmm/3x3-scalar.c",
121 "src/f32-ppmm/4x2-scalar.c",
122 "src/f32-ppmm/4x4-scalar.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800123 "src/f32-prelu/scalar-2x1.c",
124 "src/f32-prelu/scalar-2x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700125 "src/f32-rmax/scalar.c",
126 "src/f32-spmm/1x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700127 "src/f32-spmm/1x1-scalar.c",
128 "src/f32-spmm/2x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700129 "src/f32-spmm/2x1-scalar.c",
130 "src/f32-spmm/4x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700131 "src/f32-spmm/4x1-scalar.c",
132 "src/f32-spmm/8x1-scalar-pipelined.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700133 "src/f32-spmm/8x1-scalar.c",
Erich Elsenc6afd9b2019-10-24 16:10:53 -0700134 "src/f32-spmm/8x2-scalar.c",
135 "src/f32-spmm/8x4-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700136 "src/f32-vadd/scalar.c",
137 "src/f32-vmul/scalar.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800138 "src/f32-vmulcaddc/c1-scalar-2x.c",
139 "src/f32-vmulcaddc/c2-scalar-2x.c",
140 "src/f32-vmulcaddc/c4-scalar-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700141 "src/f32-vsub/scalar.c",
142 "src/q8-avgpool/mp9p8q-scalar.c",
143 "src/q8-avgpool/up9-scalar.c",
144 "src/q8-igemm/2x2-scalar.c",
145 "src/q8-dwconv/up1x9-scalar.c",
146 "src/q8-gavgpool/mp7p7q-scalar.c",
147 "src/q8-gavgpool/up7-scalar.c",
148 "src/q8-gemm/2x2-scalar.c",
149 "src/q8-vadd/scalar.c",
150 "src/u8-clamp/scalar.c",
151 "src/u8-lut32norm/scalar.c",
152 "src/u8-maxpool/9p8q-scalar.c",
153 "src/u8-rmax/scalar.c",
154 "src/x32-packx/x2-scalar.c",
155 "src/x32-packx/x3-scalar.c",
156 "src/x32-packx/x4-scalar.c",
157 "src/x32-pad/x2-scalar.c",
158 "src/x32-unpool/scalar.c",
159 "src/x32-zip/x2-scalar.c",
160 "src/x32-zip/x3-scalar.c",
161 "src/x32-zip/x4-scalar.c",
162 "src/x32-zip/xm-scalar.c",
163 "src/x8-lut/scalar.c",
164 "src/x8-zip/x2-scalar.c",
165 "src/x8-zip/x3-scalar.c",
166 "src/x8-zip/x4-scalar.c",
167 "src/x8-zip/xm-scalar.c",
168]
169
170PSIMD_UKERNELS = [
171 "src/f32-argmaxpool/mp9p8q-psimd.c",
172 "src/f32-argmaxpool/up4-psimd.c",
173 "src/f32-argmaxpool/up9-psimd.c",
174 "src/f32-avgpool/mp9p8q-psimd.c",
175 "src/f32-avgpool/up9-psimd.c",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800176 "src/f32-bilinear/psimd-c4.c",
177 "src/f32-bilinear/psimd-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700178 "src/f32-clamp/psimd.c",
179 "src/f32-igemm/1x8-psimd-loadsplat.c",
180 "src/f32-igemm/1x8-psimd-splat.c",
181 "src/f32-igemm/1x8s4-psimd.c",
182 "src/f32-igemm/4x2c4-psimd.c",
183 "src/f32-igemm/4x8-psimd-loadsplat.c",
184 "src/f32-igemm/4x8-psimd-splat.c",
185 "src/f32-igemm/4x8s4-psimd.c",
186 "src/f32-igemm/6x8-psimd-loadsplat.c",
187 "src/f32-igemm/6x8-psimd-splat.c",
188 "src/f32-igemm/6x8s4-psimd.c",
189 "src/f32-dwconv/up4x25-psimd.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800190 "src/f32-dwconv/up4x25-psimd-acc2.c",
191 "src/f32-dwconv/up8x25-psimd.c",
192 "src/f32-dwconv/up8x25-psimd-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700193 "src/f32-dwconv/up4x4-psimd.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800194 "src/f32-dwconv/up4x4-psimd-acc2.c",
195 "src/f32-dwconv/up8x4-psimd.c",
196 "src/f32-dwconv/up8x4-psimd-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700197 "src/f32-dwconv/up4x9-psimd.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800198 "src/f32-dwconv/up4x9-psimd-acc2.c",
199 "src/f32-dwconv/up8x9-psimd.c",
200 "src/f32-dwconv/up8x9-psimd-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700201 "src/f32-gavgpool/mp7p7q-psimd.c",
202 "src/f32-gavgpool/up7-psimd.c",
203 "src/f32-gemm/1x8-psimd-loadsplat.c",
204 "src/f32-gemm/1x8-psimd-splat.c",
205 "src/f32-gemm/1x8s4-psimd.c",
206 "src/f32-gemm/4x8-psimd-loadsplat.c",
207 "src/f32-gemm/4x8-psimd-splat.c",
208 "src/f32-gemm/4x8s4-psimd.c",
209 "src/f32-gemm/6x8-psimd-loadsplat.c",
210 "src/f32-gemm/6x8-psimd-splat.c",
211 "src/f32-gemm/6x8s4-psimd.c",
212 "src/f32-gemminc/1x8-psimd-loadsplat.c",
213 "src/f32-gemminc/1x8-psimd-splat.c",
214 "src/f32-gemminc/1x8s4-psimd.c",
215 "src/f32-gemminc/4x8-psimd-loadsplat.c",
216 "src/f32-gemminc/4x8-psimd-splat.c",
217 "src/f32-gemminc/4x8s4-psimd.c",
218 "src/f32-gemminc/6x8-psimd-loadsplat.c",
219 "src/f32-gemminc/6x8-psimd-splat.c",
220 "src/f32-gemminc/6x8s4-psimd.c",
221 "src/f32-hswish/psimd.c",
222 "src/f32-maxpool/9p8q-psimd.c",
223 "src/f32-pavgpool/mp9p8q-psimd.c",
224 "src/f32-pavgpool/up9-psimd.c",
225 "src/f32-ppmm/4x8-psimd.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800226 "src/f32-prelu/psimd-2x4.c",
227 "src/f32-prelu/psimd-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700228 "src/f32-vadd/psimd.c",
229 "src/f32-vmul/psimd.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800230 "src/f32-vmulcaddc/c4-psimd-2x.c",
231 "src/f32-vmulcaddc/c8-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700232 "src/f32-vsub/psimd.c",
233 "src/x32-packx/x4-psimd.c",
234 "src/x32-pad/x2-psimd.c",
235 "src/x32-unpool/psimd.c",
236 "src/x32-zip/x2-psimd.c",
237 "src/x32-zip/x3-psimd.c",
238 "src/x32-zip/x4-psimd.c",
239 "src/x32-zip/xm-psimd.c",
240]
241
242# ISA-specific micro-kernels
243NEON_UKERNELS = [
244 "src/f32-avgpool/mp9p8q-neon.c",
245 "src/f32-avgpool/up9-neon.c",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800246 "src/f32-bilinear/neon-c4.c",
247 "src/f32-bilinear/neon-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700248 "src/f32-clamp/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700249 "src/f32-dwconv/up4x9-neon.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800250 "src/f32-dwconv/up4x9-neon-acc2.c",
251 "src/f32-dwconv/up8x9-neon.c",
252 "src/f32-dwconv/up8x9-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700253 "src/f32-gavgpool-spchw/neon-x4.c",
254 "src/f32-gavgpool/mp7p7q-neon.c",
255 "src/f32-gavgpool/up7-neon.c",
256 "src/f32-gemm/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700257 "src/f32-gemm/4x2-neon-ld64.c",
258 "src/f32-gemm/4x8-neon-ld128.c",
259 "src/f32-gemm/4x8-neon-ld64.c",
260 "src/f32-gemm/5x8-neon-ld64.c",
261 "src/f32-gemm/6x8-neon-ld64.c",
262 "src/f32-gemminc/1x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700263 "src/f32-gemminc/4x8-neon-ld128.c",
264 "src/f32-gemminc/4x8-neon-ld64.c",
265 "src/f32-gemminc/5x8-neon-ld64.c",
266 "src/f32-gemminc/6x8-neon-ld64.c",
267 "src/f32-hswish/neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800268 "src/f32-igemm/1x8-neon-ld64.c",
269 "src/f32-igemm/4x2-neon-ld64.c",
270 "src/f32-igemm/4x4-neon-ld64.c",
271 "src/f32-igemm/4x8-neon-ld128.c",
272 "src/f32-igemm/4x8-neon-ld64.c",
273 "src/f32-igemm/6x8-neon-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700274 "src/f32-pavgpool/mp9p8q-neon.c",
275 "src/f32-pavgpool/up9-neon.c",
276 "src/f32-ppmm/4x8-neon.c",
277 "src/f32-ppmm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800278 "src/f32-prelu/neon-2x4.c",
279 "src/f32-prelu/neon-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700280 "src/f32-rmax/neon.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800281 "src/f32-vmulcaddc/c4-neon-2x.c",
282 "src/f32-vmulcaddc/c8-neon-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700283 "src/q8-avgpool/mp9p8q-neon.c",
284 "src/q8-avgpool/up9-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700285 "src/q8-dwconv/up8x9-neon.c",
286 "src/q8-gavgpool/mp7p7q-neon.c",
287 "src/q8-gavgpool/up7-neon.c",
288 "src/q8-gemm/4x8-neon.c",
289 "src/q8-gemm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800290 "src/q8-igemm/4x8-neon.c",
291 "src/q8-igemm/8x8-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700292 "src/q8-vadd/neon.c",
293 "src/u8-clamp/neon.c",
294 "src/u8-maxpool/9p8q-neon.c",
295 "src/u8-rmax/neon.c",
296 "src/x32-packx/x4-neon-st4.c",
297 "src/x32-pad/x2-neon.c",
298 "src/x32-zip/x2-neon.c",
299 "src/x32-zip/x3-neon.c",
300 "src/x32-zip/x4-neon.c",
301 "src/x32-zip/xm-neon.c",
302 "src/x8-zip/x2-neon.c",
303 "src/x8-zip/x3-neon.c",
304 "src/x8-zip/x4-neon.c",
305 "src/x8-zip/xm-neon.c",
306]
307
308NEONFMA_UKERNELS = [
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800309 "src/f32-bilinear/neonfma-c4.c",
310 "src/f32-bilinear/neonfma-c8.c",
Frank Barcharddb45b6a2019-10-09 16:42:45 -0700311 "src/f32-igemm/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700312 "src/f32-igemm/4x2-neonfma-ld64.c",
313 "src/f32-igemm/4x4-neonfma-ld64.c",
314 "src/f32-igemm/4x8-neonfma-ld128.c",
315 "src/f32-igemm/4x8-neonfma-ld64.c",
316 "src/f32-igemm/6x8-neonfma-ld64.c",
317 "src/f32-dwconv/up4x9-neonfma.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800318 "src/f32-dwconv/up4x9-neonfma-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700319 "src/f32-dwconv/up8x9-neonfma.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800320 "src/f32-dwconv/up8x9-neonfma-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700321 "src/f32-gemm/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700322 "src/f32-gemm/4x2-neonfma-ld64.c",
323 "src/f32-gemm/4x8-neonfma-ld128.c",
324 "src/f32-gemm/4x8-neonfma-ld64.c",
325 "src/f32-gemm/5x8-neonfma-ld64.c",
326 "src/f32-gemm/6x8-neonfma-ld64.c",
327 "src/f32-gemminc/1x8-neonfma-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700328 "src/f32-gemminc/4x8-neonfma-ld128.c",
329 "src/f32-gemminc/4x8-neonfma-ld64.c",
330 "src/f32-gemminc/5x8-neonfma-ld64.c",
331 "src/f32-gemminc/6x8-neonfma-ld64.c",
332 "src/f32-hswish/neonfma.c",
333 "src/f32-ppmm/4x8-neonfma.c",
334 "src/f32-ppmm/8x8-neonfma.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800335 "src/f32-vmulcaddc/c4-neonfma-2x.c",
336 "src/f32-vmulcaddc/c8-neonfma-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700337]
338
339AARCH64_NEONFMA_UKERNELS = [
340 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
341 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
342 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
343 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
344 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
345 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
346 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
347 "src/f32-spmm/12x1-neonfma.c",
348 "src/f32-spmm/12x2-neonfma.c",
349 "src/f32-spmm/12x4-neonfma.c",
350 "src/f32-spmm/16x1-neonfma-pipelined.c",
351 "src/f32-spmm/16x1-neonfma-unroll2.c",
352 "src/f32-spmm/16x1-neonfma.c",
353 "src/f32-spmm/16x2-neonfma.c",
354 "src/f32-spmm/16x4-neonfma.c",
355 "src/f32-spmm/4x1-neonfma-pipelined.c",
356 "src/f32-spmm/4x1-neonfma-unroll2.c",
357 "src/f32-spmm/4x1-neonfma.c",
358 "src/f32-spmm/4x2-neonfma.c",
359 "src/f32-spmm/4x4-neonfma.c",
360 "src/f32-spmm/8x1-neonfma-pipelined.c",
361 "src/f32-spmm/8x1-neonfma-unroll2.c",
362 "src/f32-spmm/8x1-neonfma.c",
363 "src/f32-spmm/8x2-neonfma.c",
364 "src/f32-spmm/8x4-neonfma.c",
365]
366
367AARCH64_NEONFP16ARITH_UKERNELS = [
368 "src/f16-gemm/4x8-neonfp16arith-ld64.c",
369 "src/f16-gemm/6x8-neonfp16arith-ld64.c",
370 "src/f16-gemm/8x8-neonfp16arith-ld64.c",
371]
372
373SSE_UKERNELS = [
374 "src/f32-avgpool/mp9p8q-sse.c",
375 "src/f32-avgpool/up9-sse.c",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800376 "src/f32-bilinear/sse-c4.c",
377 "src/f32-bilinear/sse-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700378 "src/f32-clamp/sse.c",
379 "src/f32-igemm/1x8-sse-dup.c",
380 "src/f32-igemm/1x8-sse-load1.c",
381 "src/f32-igemm/1x8s4-sse.c",
382 "src/f32-igemm/4x2c4-sse.c",
383 "src/f32-igemm/4x8-sse-dup.c",
384 "src/f32-igemm/4x8-sse-load1.c",
385 "src/f32-igemm/4x8s4-sse.c",
386 "src/f32-dwconv/up4x25-sse.c",
387 "src/f32-dwconv/up4x4-sse.c",
388 "src/f32-dwconv/up4x9-sse.c",
Marat Dukhan5098c3e2019-11-07 12:01:19 -0800389 "src/f32-dwconv/up4x25-sse-acc2.c",
390 "src/f32-dwconv/up4x4-sse-acc2.c",
391 "src/f32-dwconv/up4x9-sse-acc2.c",
392 "src/f32-dwconv/up8x25-sse.c",
393 "src/f32-dwconv/up8x4-sse.c",
394 "src/f32-dwconv/up8x9-sse.c",
395 "src/f32-dwconv/up8x25-sse-acc2.c",
396 "src/f32-dwconv/up8x4-sse-acc2.c",
397 "src/f32-dwconv/up8x9-sse-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700398 "src/f32-gavgpool-spchw/sse-x4.c",
399 "src/f32-gavgpool/mp7p7q-sse.c",
400 "src/f32-gavgpool/up7-sse.c",
401 "src/f32-gemm/1x8-sse-dup.c",
402 "src/f32-gemm/1x8-sse-load1.c",
403 "src/f32-gemm/1x8s4-sse.c",
404 "src/f32-gemm/4x8-sse-dup.c",
405 "src/f32-gemm/4x8-sse-load1.c",
406 "src/f32-gemm/4x8s4-sse.c",
407 "src/f32-gemminc/1x8-sse-dup.c",
408 "src/f32-gemminc/1x8-sse-load1.c",
409 "src/f32-gemminc/1x8s4-sse.c",
410 "src/f32-gemminc/4x8-sse-dup.c",
411 "src/f32-gemminc/4x8-sse-load1.c",
412 "src/f32-gemminc/4x8s4-sse.c",
413 "src/f32-hswish/sse.c",
414 "src/f32-maxpool/9p8q-sse.c",
415 "src/f32-pavgpool/mp9p8q-sse.c",
416 "src/f32-pavgpool/up9-sse.c",
417 "src/f32-dwconv-spchw/3x3p1-sse.c",
418 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
419 "src/f32-ppmm/4x8-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700420 "src/f32-rmax/sse.c",
421 "src/f32-spmm/4x1-sse.c",
422 "src/f32-spmm/8x1-sse.c",
423 "src/f32-vadd/sse.c",
424 "src/f32-vmul/sse.c",
Marat Dukhan49e6ee92019-11-06 15:55:29 -0800425 "src/f32-vmulcaddc/c4-sse-2x.c",
426 "src/f32-vmulcaddc/c8-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700427 "src/f32-vsub/sse.c",
428 "src/x32-packx/x4-sse.c",
429]
430
431SSE2_UKERNELS = [
432 "src/f32-argmaxpool/mp9p8q-sse2.c",
433 "src/f32-argmaxpool/up4-sse2.c",
434 "src/f32-argmaxpool/up9-sse2.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800435 "src/f32-prelu/sse2-2x4.c",
436 "src/f32-prelu/sse2-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700437 "src/q8-avgpool/mp9p8q-sse2.c",
438 "src/q8-avgpool/up9-sse2.c",
439 "src/q8-igemm/4x4c2-sse2.c",
440 "src/q8-dwconv/up8x9-sse2.c",
441 "src/q8-gavgpool/mp7p7q-sse2.c",
442 "src/q8-gavgpool/up7-sse2.c",
443 "src/q8-gemm/2x4c8-sse2.c",
444 "src/q8-gemm/4x4c2-sse2.c",
445 "src/q8-vadd/sse2.c",
446 "src/u8-clamp/sse2.c",
447 "src/u8-maxpool/9p8q-sse2.c",
448 "src/u8-rmax/sse2.c",
449 "src/x32-pad/x2-sse2.c",
450 "src/x32-zip/x2-sse2.c",
451 "src/x32-zip/x3-sse2.c",
452 "src/x32-zip/x4-sse2.c",
453 "src/x32-zip/xm-sse2.c",
454 "src/x8-zip/x2-sse2.c",
455 "src/x8-zip/x3-sse2.c",
456 "src/x8-zip/x4-sse2.c",
457 "src/x8-zip/xm-sse2.c",
458]
459
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800460SSE41_UKERNELS = [
461 "src/f32-prelu/sse41-2x4.c",
462 "src/f32-prelu/sse41-2x8.c",
463]
464
Marat Dukhan08c4a432019-10-03 09:29:21 -0700465AVX_UKERNELS = [
466 "src/f32-rmax/avx.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700467 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700468]
469
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700470AVX2_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700471 "src/f32-raddexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700472 "src/f32-raddextexp/avx2-p5-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700473 "src/f32-raddstoreexpminusmax/avx2-p5-unroll64.c",
474 "src/f32-vscaleexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700475 "src/f32-vscaleextexp/avx2-p5-unroll64.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700476 "src/math/exp-avx2-p5.c",
477 "src/math/exp-avx2-perm-p3.c",
478 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -0700479 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700480 "src/math/extexp-avx2-p5.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700481]
482
Marat Dukhan08c4a432019-10-03 09:29:21 -0700483AVX512F_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700484 "src/f32-raddexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700485 "src/f32-raddextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700486 "src/f32-raddstoreexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700487 "src/f32-rmax/avx512f.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700488 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700489 "src/f32-vscaleexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700490 "src/f32-vscaleextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700491 "src/math/exp-avx512f-p5-scalef.c",
492 "src/math/exp-avx512f-p5.c",
493 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -0700494 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700495 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700496]
497
498AARCH32_ASM_UKERNELS = [
499 "src/q8-dwconv/up8x9-aarch32-neon.S",
500]
501
502AARCH64_ASM_UKERNELS = [
503 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
504 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
505 "src/f32-gemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700506 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700507 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a57.S",
508 "src/f32-gemm/1x8-aarch64-neonfma-cortex-a75.S",
509 "src/f32-gemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700510 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700511 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a57.S",
512 "src/f32-gemm/4x8-aarch64-neonfma-cortex-a75.S",
513 "src/f32-gemm/4x8-aarch64-neonfma-ld128.S",
514 "src/f32-gemm/4x8-aarch64-neonfma-ld64.S",
515 "src/f32-gemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700516 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700517 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a57.S",
518 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a73.S",
519 "src/f32-gemm/6x8-aarch64-neonfma-cortex-a75.S",
520 "src/f32-gemm/6x8-aarch64-neonfma-ld128.S",
521 "src/f32-gemm/6x8-aarch64-neonfma-ld64.S",
522 "src/f32-gemminc/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700523 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700524 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a57.S",
525 "src/f32-gemminc/1x8-aarch64-neonfma-cortex-a75.S",
526 "src/f32-gemminc/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700527 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700528 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a57.S",
529 "src/f32-gemminc/4x8-aarch64-neonfma-cortex-a75.S",
530 "src/f32-gemminc/4x8-aarch64-neonfma-ld128.S",
531 "src/f32-gemminc/4x8-aarch64-neonfma-ld64.S",
532 "src/f32-gemminc/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700533 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700534 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a57.S",
535 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a73.S",
536 "src/f32-gemminc/6x8-aarch64-neonfma-cortex-a75.S",
537 "src/f32-gemminc/6x8-aarch64-neonfma-ld128.S",
538 "src/f32-gemminc/6x8-aarch64-neonfma-ld64.S",
539 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700540 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700541 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S",
542 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S",
543 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700544 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700545 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
546 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700547 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700548 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
549 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
550 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
551]
552
553INTERNAL_MICROKERNEL_HDRS = [
554 "src/xnnpack/argmaxpool.h",
555 "src/xnnpack/avgpool.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800556 "src/xnnpack/bilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700557 "src/xnnpack/clamp.h",
558 "src/xnnpack/common.h",
559 "src/xnnpack/conv.h",
560 "src/xnnpack/dwconv.h",
561 "src/xnnpack/gavgpool.h",
562 "src/xnnpack/gemm.h",
563 "src/xnnpack/hswish.h",
564 "src/xnnpack/igemm.h",
565 "src/xnnpack/lut.h",
566 "src/xnnpack/math.h",
567 "src/xnnpack/maxpool.h",
568 "src/xnnpack/packx.h",
569 "src/xnnpack/pad.h",
570 "src/xnnpack/params.h",
571 "src/xnnpack/pavgpool.h",
572 "src/xnnpack/ppmm.h",
573 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700574 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700575 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700576 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700577 "src/xnnpack/rmax.h",
578 "src/xnnpack/scalar-utils.h",
579 "src/xnnpack/spmm.h",
580 "src/xnnpack/unpool.h",
581 "src/xnnpack/vadd.h",
582 "src/xnnpack/vmul.h",
583 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700584 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700585 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700586 "src/xnnpack/vscaleextexp.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700587 "src/xnnpack/vsub.h",
588 "src/xnnpack/zip.h",
589]
590
591INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
592 "include/xnnpack.h",
593 "src/xnnpack/allocator.h",
594 "src/xnnpack/compute.h",
595 "src/xnnpack/im2col.h",
596 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700597 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700598 "src/xnnpack/operator.h",
599 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700600 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700601 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700602 "src/xnnpack/requantization.h",
603]
604
605ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
606 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700607]
608
609MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700610 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700611 "include/xnnpack.h",
612]
613
614MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
615 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700616 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700617 "src/xnnpack/requantization.h",
618 "include/xnnpack.h",
619]
620
621OPERATOR_TEST_PARAMS_HDRS = [
622 "src/xnnpack/params.h",
623 "src/xnnpack/common.h",
624]
625
626WEIGHTS_PACK_HDRS = [
627 "src/xnnpack/pack.h",
628 "src/xnnpack/operator.h",
629 "src/xnnpack/compute.h",
630]
631
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700632LOGGING_COPTS = select({
633 # No logging in optimized mode
634 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
635 # Full logging in debug mode
636 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
637 # Error-only logging in default (fastbuild) mode
638 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
639})
640
641LOGGING_HDRS = [
642 "src/xnnpack/log.h",
643]
644
Marat Dukhan08c4a432019-10-03 09:29:21 -0700645xnnpack_cc_library(
646 name = "scalar_ukernels",
647 srcs = SCALAR_UKERNELS,
648 hdrs = INTERNAL_HDRS,
649 aarch32_copts = ["-marm"],
650 copts = xnnpack_std_copts(),
651 deps = [
652 "@FP16",
653 "@FXdiv",
654 ],
655)
656
657xnnpack_cc_library(
658 name = "psimd_ukernels",
659 srcs = PSIMD_UKERNELS,
660 hdrs = INTERNAL_HDRS,
661 aarch32_copts = [
662 "-marm",
663 "-mfpu=neon",
664 ],
665 copts = xnnpack_std_copts(),
666 optimized_copts = [
667 "-O3",
668 "-ffast-math",
669 ],
670 deps = [
671 "@FP16",
672 "@psimd",
673 ],
674)
675
676xnnpack_cc_library(
677 name = "neon_ukernels",
678 hdrs = INTERNAL_HDRS,
679 aarch32_copts = [
680 "-marm",
681 "-mfpu=neon",
682 ],
683 aarch32_srcs = NEON_UKERNELS,
684 aarch64_srcs = NEON_UKERNELS,
685 copts = xnnpack_std_copts(),
686 deps = ["@FP16"],
687)
688
689xnnpack_cc_library(
690 name = "neonfma_ukernels",
691 hdrs = INTERNAL_HDRS,
692 aarch32_copts = [
693 "-marm",
694 "-mfpu=neon-vfpv4",
695 ],
696 aarch32_srcs = NEONFMA_UKERNELS,
697 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
698 copts = xnnpack_std_copts(),
699 deps = ["@FP16"],
700)
701
702xnnpack_cc_library(
703 name = "neonfp16arith_ukernels",
704 hdrs = INTERNAL_HDRS,
705 aarch64_copts = ["-march=armv8.2-a+fp16"],
706 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
707 copts = xnnpack_std_copts(),
708 deps = ["@FP16"],
709)
710
711xnnpack_cc_library(
712 name = "sse2_ukernels",
713 hdrs = INTERNAL_HDRS,
714 copts = xnnpack_std_copts(),
715 x86_copts = ["-msse2"],
716 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
717 deps = ["@FP16"],
718)
719
720xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800721 name = "sse41_ukernels",
722 hdrs = INTERNAL_HDRS,
723 copts = xnnpack_std_copts(),
724 x86_copts = ["-msse4.1"],
725 x86_srcs = SSE41_UKERNELS,
726 deps = ["@FP16"],
727)
728
729xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700730 name = "avx_ukernels",
731 hdrs = INTERNAL_HDRS,
732 copts = xnnpack_std_copts(),
733 x86_copts = ["-mavx"],
734 x86_srcs = AVX_UKERNELS,
735 deps = ["@FP16"],
736)
737
738xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700739 name = "avx2_ukernels",
740 hdrs = INTERNAL_HDRS,
741 copts = xnnpack_std_copts(),
742 x86_copts = [
743 "-mfma",
744 "-mavx2",
745 ],
746 x86_srcs = AVX2_UKERNELS,
747 deps = ["@FP16"],
748)
749
750xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700751 name = "avx512f_ukernels",
752 hdrs = INTERNAL_HDRS,
753 copts = xnnpack_std_copts(),
754 x86_copts = ["-mavx512f"],
755 x86_srcs = AVX512F_UKERNELS,
756 deps = ["@FP16"],
757)
758
759xnnpack_cc_library(
760 name = "asm_ukernels",
761 hdrs = ["src/xnnpack/assembly.h"],
762 aarch32_srcs = AARCH32_ASM_UKERNELS,
763 aarch64_srcs = AARCH64_ASM_UKERNELS,
764)
765
766xnnpack_aggregate_library(
767 name = "ukernels",
768 aarch32_deps = [
769 ":psimd_ukernels",
770 ":neon_ukernels",
771 ":neonfma_ukernels",
772 ":asm_ukernels",
773 ],
774 aarch64_deps = [
775 ":psimd_ukernels",
776 ":neon_ukernels",
777 ":neonfma_ukernels",
778 ":neonfp16arith_ukernels",
779 ":asm_ukernels",
780 ],
781 generic_deps = [":scalar_ukernels"],
782 wasmsimd_deps = [
783 ":psimd_ukernels",
784 ],
785 x86_deps = [
786 ":psimd_ukernels",
787 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800788 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700789 ":avx_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700790 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700791 ":avx512f_ukernels",
792 ],
793)
794
795xnnpack_cc_library(
796 name = "im2col",
797 srcs = ["src/im2col.c"],
798 hdrs = [
799 "src/xnnpack/common.h",
800 "src/xnnpack/im2col.h",
801 ],
802 copts = xnnpack_std_copts(),
803)
804
805xnnpack_cc_library(
806 name = "indirection",
807 srcs = ["src/indirection.c"],
808 hdrs = INTERNAL_HDRS,
809 copts = xnnpack_std_copts(),
810 deps = [
811 "@FP16",
812 "@FXdiv",
813 "@pthreadpool",
814 ],
815)
816
817xnnpack_cc_library(
818 name = "operator_run",
819 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700820 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
821 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700822 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
823 "-Wno-vla",
824 ],
825 deps = [
826 "@FP16",
827 "@FXdiv",
828 "@clog",
829 "@pthreadpool",
830 ],
831)
832
833cc_library(
834 name = "enable_assembly",
835 defines = select({
836 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
837 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -0700838 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700839 }),
840)
841
Marat Dukhancf056b22019-10-07 10:26:29 -0700842xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700843 name = "operators",
844 srcs = OPERATOR_SRCS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700845 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -0700846 ],
847 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700848 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700849 "-Isrc",
850 "-Iinclude",
851 ] + select({
852 ":debug_build": [],
853 "//conditions:default": xnnpack_min_size_copts(),
854 }),
Marat Dukhancf056b22019-10-07 10:26:29 -0700855 wasm_srcs = ["src/wasm-stubs.c"],
856 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700857 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -0700858 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700859 "@FP16",
860 "@FXdiv",
861 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700862 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700863 ],
864)
865
866cc_library(
867 name = "XNNPACK",
868 srcs = [
869 "src/init.c",
870 ],
871 copts = xnnpack_std_copts() + LOGGING_COPTS + [
872 "-Isrc",
873 "-Iinclude",
874 ] + select({
875 ":debug_build": [],
876 "//conditions:default": xnnpack_min_size_copts(),
877 }),
878 includes = ["include"],
879 linkstatic = True,
880 textual_hdrs = ["include/xnnpack.h"],
881 visibility = xnnpack_visibility(),
882 deps = [
883 ":enable_assembly",
884 ":ukernels",
885 ":operator_run",
886 ":operators",
887 "@clog",
888 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -0700889 ] + select({
890 ":emscripten": [],
891 "//conditions:default": ["@cpuinfo"],
892 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700893)
894
895cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700896 name = "xnnpack_operators_nhwc_f32",
897 srcs = [
898 "src/init.c",
899 ],
900 copts = xnnpack_std_copts() + LOGGING_COPTS + [
901 "-Isrc",
902 "-Iinclude",
903 ] + select({
904 ":debug_build": [],
905 "//conditions:default": xnnpack_min_size_copts(),
906 }),
907 defines = [
908 "XNN_NO_Q8_OPERATORS",
909 "XNN_NO_U8_OPERATORS",
910 "XNN_NO_X8_OPERATORS",
911 "XNN_NO_SPNCHW_OPERATORS",
912 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700913 includes = ["include"],
914 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700915 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700916 visibility = xnnpack_visibility(),
917 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700918 ":enable_assembly",
919 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700920 ":operator_run",
921 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700922 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700923 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -0700924 ] + select({
925 ":emscripten": [],
926 "//conditions:default": ["@cpuinfo"],
927 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -0700928)
929
Marat Dukhancf056b22019-10-07 10:26:29 -0700930xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700931 name = "bench_utils",
932 srcs = ["bench/utils.cc"],
933 hdrs = ["bench/utils.h"],
934 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -0800935 deps = [
936 "@com_google_benchmark//:benchmark",
937 "@cpuinfo",
938 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700939)
940
Frank Barchard7e955972019-10-11 10:34:25 -0700941######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -0700942
943xnnpack_benchmark(
944 name = "q8_gemm_bench",
945 srcs = [
946 "bench/gemm.h",
947 "bench/q8-gemm.cc",
948 "src/xnnpack/AlignedAllocator.h",
949 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
950 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
951 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
952)
953
954xnnpack_benchmark(
955 name = "f16_gemm_bench",
956 srcs = [
957 "bench/f16-gemm.cc",
958 "bench/gemm.h",
959 "src/xnnpack/AlignedAllocator.h",
960 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
961 copts = ["-Wno-unused-function"],
962 deps = MICROKERNEL_BENCHMARK_DEPS,
963)
964
965xnnpack_benchmark(
966 name = "f32_igemm_bench",
967 srcs = [
968 "bench/f32-igemm.cc",
969 "bench/conv.h",
970 "src/xnnpack/AlignedAllocator.h",
971 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -0700972 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700973)
974
975xnnpack_benchmark(
976 name = "f32_conv_hwc_bench",
977 srcs = [
978 "bench/f32-conv-hwc.cc",
979 "bench/dconv.h",
980 "src/xnnpack/AlignedAllocator.h",
981 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
982 copts = ["-Wno-unused-function"],
983 deps = MICROKERNEL_BENCHMARK_DEPS,
984)
985
986xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -0700987 name = "f32_conv_hwc2spchw_bench",
988 srcs = [
989 "bench/f32-conv-hwc2spchw.cc",
990 "bench/dconv.h",
991 "src/xnnpack/AlignedAllocator.h",
992 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
993 copts = ["-Wno-unused-function"],
994 deps = MICROKERNEL_BENCHMARK_DEPS,
995)
996
997xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700998 name = "f32_dwconv_bench",
999 srcs = [
1000 "bench/f32-dwconv.cc",
1001 "bench/dwconv.h",
1002 "src/xnnpack/AlignedAllocator.h",
1003 ] + MICROKERNEL_BENCHMARK_HDRS,
1004 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1005)
1006
1007xnnpack_benchmark(
1008 name = "f32_dwconv_spchw_bench",
1009 srcs = [
1010 "bench/f32-dwconv-spchw.cc",
1011 "bench/dwconv.h",
1012 "src/xnnpack/AlignedAllocator.h",
1013 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1014 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1015)
1016
1017xnnpack_benchmark(
1018 name = "f32_gemm_bench",
1019 srcs = [
1020 "bench/f32-gemm.cc",
1021 "bench/gemm.h",
1022 "src/xnnpack/AlignedAllocator.h",
1023 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1024 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07001025 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001026)
1027
1028xnnpack_benchmark(
1029 name = "f32_rmax_bench",
1030 srcs = [
1031 "bench/f32-rmax.cc",
1032 "src/xnnpack/AlignedAllocator.h",
1033 ] + MICROKERNEL_BENCHMARK_HDRS,
1034 deps = MICROKERNEL_BENCHMARK_DEPS,
1035)
1036
1037xnnpack_benchmark(
1038 name = "f32_spmm_bench",
1039 srcs = [
1040 "bench/f32-spmm.cc",
1041 "bench/gemm.h",
1042 "src/xnnpack/AlignedAllocator.h",
1043 ] + MICROKERNEL_BENCHMARK_HDRS,
1044 copts = ["-Wno-unused-function"],
1045 deps = MICROKERNEL_BENCHMARK_DEPS,
1046)
1047
1048xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07001049 name = "f32_softargmax_bench",
1050 srcs = [
1051 "bench/f32-softargmax.cc",
1052 ] + MICROKERNEL_BENCHMARK_HDRS,
1053 copts = ["-Wno-unused-function"],
1054 deps = MICROKERNEL_BENCHMARK_DEPS,
1055)
1056
1057xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001058 name = "f32_im2col_gemm_bench",
1059 srcs = [
1060 "bench/f32-im2col-gemm.cc",
1061 "bench/conv.h",
1062 "src/xnnpack/AlignedAllocator.h",
1063 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1064 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
1065)
1066
1067########################### Benchmarks for operators ###########################
1068
1069xnnpack_benchmark(
1070 name = "add_bench",
1071 srcs = ["bench/add.cc"],
1072 deps = OPERATOR_BENCHMARK_DEPS,
1073)
1074
1075xnnpack_benchmark(
1076 name = "average_pooling_bench",
1077 srcs = ["bench/average-pooling.cc"],
1078 deps = OPERATOR_BENCHMARK_DEPS,
1079)
1080
1081xnnpack_benchmark(
1082 name = "channel_shuffle_bench",
1083 srcs = ["bench/channel-shuffle.cc"],
1084 deps = OPERATOR_BENCHMARK_DEPS,
1085)
1086
1087xnnpack_benchmark(
1088 name = "convolution_bench",
1089 srcs = ["bench/convolution.cc"],
1090 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1091 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1092)
1093
1094xnnpack_benchmark(
1095 name = "deconvolution_bench",
1096 srcs = ["bench/deconvolution.cc"],
1097 copts = xnnpack_optional_tflite_copts(),
1098 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1099)
1100
1101xnnpack_benchmark(
1102 name = "global_average_pooling_bench",
1103 srcs = ["bench/global-average-pooling.cc"],
1104 deps = OPERATOR_BENCHMARK_DEPS,
1105)
1106
1107xnnpack_benchmark(
1108 name = "max_pooling_bench",
1109 srcs = ["bench/max-pooling.cc"],
1110 deps = OPERATOR_BENCHMARK_DEPS,
1111)
1112
1113xnnpack_benchmark(
1114 name = "sigmoid_bench",
1115 srcs = ["bench/sigmoid.cc"],
1116 deps = OPERATOR_BENCHMARK_DEPS,
1117)
1118
1119xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07001120 name = "prelu_bench",
1121 srcs = ["bench/prelu.cc"],
1122 copts = xnnpack_optional_tflite_copts(),
1123 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1124)
1125
1126xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001127 name = "softargmax_bench",
1128 srcs = ["bench/softargmax.cc"],
1129 deps = OPERATOR_BENCHMARK_DEPS,
1130)
1131
Marat Dukhanc068bb62019-10-04 13:24:39 -07001132############################# End-to-end benchmarks ############################
1133
1134cc_library(
1135 name = "mobilenet_v1",
1136 srcs = ["models/mobilenet-v1.cc"],
1137 hdrs = ["models/models.h"],
1138 linkstatic = True,
1139 deps = [
1140 ":XNNPACK",
1141 "@pthreadpool",
1142 ],
1143)
1144
1145cc_library(
1146 name = "mobilenet_v2",
1147 srcs = ["models/mobilenet-v2.cc"],
1148 hdrs = ["models/models.h"],
1149 linkstatic = True,
1150 deps = [
1151 ":XNNPACK",
1152 "@pthreadpool",
1153 ],
1154)
1155
1156xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07001157 name = "f32_dwconv_e2e_bench",
1158 srcs = ["bench/f32-dwconv-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1159 copts = ["-Wno-unused-function"],
1160 deps = MICROKERNEL_BENCHMARK_DEPS + [
1161 ":XNNPACK",
1162 ":mobilenet_v1",
1163 ":mobilenet_v2",
1164 ],
1165)
1166
1167xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07001168 name = "f32_gemm_e2e_bench",
1169 srcs = ["bench/f32-gemm-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1170 copts = ["-Wno-unused-function"],
1171 deps = MICROKERNEL_BENCHMARK_DEPS + [
1172 ":XNNPACK",
1173 ":mobilenet_v1",
1174 ":mobilenet_v2",
1175 ],
1176)
1177
1178xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07001179 name = "end2end_bench",
1180 srcs = ["bench/end2end.cc"],
1181 deps = [
1182 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07001183 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07001184 ":mobilenet_v1",
1185 ":mobilenet_v2",
1186 "@pthreadpool",
1187 ],
1188)
1189
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001190#################### Accuracy evaluation for math functions ####################
1191
1192xnnpack_benchmark(
1193 name = "f32_exp_eval",
1194 srcs = [
1195 "eval/f32-exp.cc",
1196 "src/xnnpack/AlignedAllocator.h",
1197 ] + ACCURACY_EVAL_HDRS,
1198 deps = ACCURACY_EVAL_DEPS,
1199)
1200
Marat Dukhan515c9772019-10-17 18:07:57 -07001201xnnpack_benchmark(
1202 name = "f32_expminus_eval",
1203 srcs = [
1204 "eval/f32-expminus.cc",
1205 "src/xnnpack/AlignedAllocator.h",
1206 ] + ACCURACY_EVAL_HDRS,
1207 deps = ACCURACY_EVAL_DEPS,
1208)
1209
Marat Dukhan98ba4412019-10-23 02:14:28 -07001210xnnpack_benchmark(
1211 name = "f32_extexp_eval",
1212 srcs = [
1213 "eval/f32-extexp.cc",
1214 "src/xnnpack/AlignedAllocator.h",
1215 ] + ACCURACY_EVAL_HDRS,
1216 deps = ACCURACY_EVAL_DEPS,
1217)
1218
Marat Dukhan08c4a432019-10-03 09:29:21 -07001219######################### Unit tests for micro-kernels #########################
1220
1221xnnpack_unit_test(
1222 name = "f16_gemm_test",
1223 srcs = [
1224 "test/f16-gemm.cc",
1225 "test/gemm-microkernel-tester.h",
1226 "src/xnnpack/AlignedAllocator.h",
1227 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1228 deps = MICROKERNEL_TEST_DEPS,
1229)
1230
1231xnnpack_unit_test(
1232 name = "f32_argmaxpool_test",
1233 srcs = [
1234 "test/f32-argmaxpool.cc",
1235 "test/argmaxpool-microkernel-tester.h",
1236 "src/xnnpack/AlignedAllocator.h",
1237 ] + MICROKERNEL_TEST_HDRS,
1238 deps = MICROKERNEL_TEST_DEPS,
1239)
1240
1241xnnpack_unit_test(
1242 name = "f32_avgpool_test",
1243 srcs = [
1244 "test/f32-avgpool.cc",
1245 "test/avgpool-microkernel-tester.h",
1246 "src/xnnpack/AlignedAllocator.h",
1247 ] + MICROKERNEL_TEST_HDRS,
1248 deps = MICROKERNEL_TEST_DEPS,
1249)
1250
1251xnnpack_unit_test(
Marat Dukhan35dacfb2019-11-07 19:18:16 -08001252 name = "f32_bilinear_test",
1253 srcs = [
1254 "test/f32-bilinear.cc",
1255 "test/bilinear-microkernel-tester.h",
1256 "src/xnnpack/AlignedAllocator.h",
1257 ] + MICROKERNEL_TEST_HDRS,
1258 deps = MICROKERNEL_TEST_DEPS,
1259)
1260
1261xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001262 name = "f32_clamp_test",
1263 srcs = [
1264 "test/f32-clamp.cc",
1265 "test/clamp-microkernel-tester.h",
1266 ] + MICROKERNEL_TEST_HDRS,
1267 deps = MICROKERNEL_TEST_DEPS,
1268)
1269
1270xnnpack_unit_test(
1271 name = "f32_igemm_test",
1272 srcs = [
1273 "test/f32-igemm.cc",
1274 "test/gemm-microkernel-tester.h",
1275 "src/xnnpack/AlignedAllocator.h",
1276 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1277 deps = MICROKERNEL_TEST_DEPS,
1278)
1279
1280xnnpack_unit_test(
1281 name = "f32_conv_hwc_test",
1282 srcs = [
1283 "test/f32-conv-hwc.cc",
1284 "test/conv-hwc-microkernel-tester.h",
1285 "src/xnnpack/AlignedAllocator.h",
1286 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1287 deps = MICROKERNEL_TEST_DEPS,
1288)
1289
1290xnnpack_unit_test(
1291 name = "f32_conv_hwc2spchw_test",
1292 srcs = [
1293 "test/f32-conv-hwc2spchw.cc",
1294 "test/conv-hwc2spchw-microkernel-tester.h",
1295 "src/xnnpack/AlignedAllocator.h",
1296 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1297 deps = MICROKERNEL_TEST_DEPS,
1298)
1299
1300xnnpack_unit_test(
1301 name = "f32_dwconv_test",
1302 srcs = [
1303 "test/f32-dwconv.cc",
1304 "test/dwconv-microkernel-tester.h",
1305 "src/xnnpack/AlignedAllocator.h",
1306 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1307 deps = MICROKERNEL_TEST_DEPS,
1308)
1309
1310xnnpack_unit_test(
1311 name = "f32_dwconv_spchw_test",
1312 srcs = [
1313 "test/f32-dwconv-spchw.cc",
1314 "test/dwconv-spchw-microkernel-tester.h",
1315 "src/xnnpack/AlignedAllocator.h",
1316 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1317 deps = MICROKERNEL_TEST_DEPS,
1318)
1319
1320xnnpack_unit_test(
1321 name = "f32_gavgpool_test",
1322 srcs = [
1323 "test/f32-gavgpool.cc",
1324 "test/gavgpool-microkernel-tester.h",
1325 "src/xnnpack/AlignedAllocator.h",
1326 ] + MICROKERNEL_TEST_HDRS,
1327 deps = MICROKERNEL_TEST_DEPS,
1328)
1329
1330xnnpack_unit_test(
1331 name = "f32_gavgpool_spchw_test",
1332 srcs = [
1333 "test/f32-gavgpool-spchw.cc",
1334 "test/gavgpool-spchw-microkernel-tester.h",
1335 "src/xnnpack/AlignedAllocator.h",
1336 ] + MICROKERNEL_TEST_HDRS,
1337 deps = MICROKERNEL_TEST_DEPS,
1338)
1339
1340xnnpack_unit_test(
1341 name = "f32_gemm_test",
1342 srcs = [
1343 "test/f32-gemm.cc",
1344 "test/gemm-microkernel-tester.h",
1345 "src/xnnpack/AlignedAllocator.h",
1346 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1347 deps = MICROKERNEL_TEST_DEPS,
1348)
1349
1350xnnpack_unit_test(
1351 name = "f32_gemminc_test",
1352 srcs = [
1353 "test/f32-gemminc.cc",
1354 "test/gemm-microkernel-tester.h",
1355 "src/xnnpack/AlignedAllocator.h",
1356 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1357 deps = MICROKERNEL_TEST_DEPS,
1358)
1359
1360xnnpack_unit_test(
1361 name = "f32_hswish_test",
1362 srcs = [
1363 "test/f32-hswish.cc",
1364 "test/hswish-microkernel-tester.h",
1365 ] + MICROKERNEL_TEST_HDRS,
1366 deps = MICROKERNEL_TEST_DEPS,
1367)
1368
1369xnnpack_unit_test(
1370 name = "f32_maxpool_test",
1371 srcs = [
1372 "test/f32-maxpool.cc",
1373 "test/maxpool-microkernel-tester.h",
1374 ] + MICROKERNEL_TEST_HDRS,
1375 deps = MICROKERNEL_TEST_DEPS,
1376)
1377
1378xnnpack_unit_test(
1379 name = "f32_pavgpool_test",
1380 srcs = [
1381 "test/f32-pavgpool.cc",
1382 "test/avgpool-microkernel-tester.h",
1383 "src/xnnpack/AlignedAllocator.h",
1384 ] + MICROKERNEL_TEST_HDRS,
1385 deps = MICROKERNEL_TEST_DEPS,
1386)
1387
1388xnnpack_unit_test(
1389 name = "f32_ppmm_test",
1390 srcs = [
1391 "test/f32-ppmm.cc",
1392 "test/gemm-microkernel-tester.h",
1393 "src/xnnpack/AlignedAllocator.h",
1394 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1395 deps = MICROKERNEL_TEST_DEPS,
1396)
1397
1398xnnpack_unit_test(
1399 name = "f32_prelu_test",
1400 srcs = [
1401 "test/f32-prelu.cc",
1402 "test/prelu-microkernel-tester.h",
1403 "src/xnnpack/AlignedAllocator.h",
1404 ] + MICROKERNEL_TEST_HDRS,
1405 deps = MICROKERNEL_TEST_DEPS,
1406)
1407
1408xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001409 name = "f32_raddexpminusmax_test",
1410 srcs = [
1411 "test/f32-raddexpminusmax.cc",
1412 "test/raddexpminusmax-microkernel-tester.h",
1413 ] + MICROKERNEL_TEST_HDRS,
1414 deps = MICROKERNEL_TEST_DEPS,
1415)
1416
1417xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001418 name = "f32_raddextexp_test",
1419 srcs = [
1420 "test/f32-raddextexp.cc",
1421 "test/raddextexp-microkernel-tester.h",
1422 ] + MICROKERNEL_TEST_HDRS,
1423 deps = MICROKERNEL_TEST_DEPS,
1424)
1425
1426xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001427 name = "f32_raddstoreexpminusmax_test",
1428 srcs = [
1429 "test/f32-raddstoreexpminusmax.cc",
1430 "test/raddstoreexpminusmax-microkernel-tester.h",
1431 ] + MICROKERNEL_TEST_HDRS,
1432 deps = MICROKERNEL_TEST_DEPS,
1433)
1434
1435xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001436 name = "f32_rmax_test",
1437 srcs = [
1438 "test/f32-rmax.cc",
1439 "test/rmax-microkernel-tester.h",
1440 ] + MICROKERNEL_TEST_HDRS,
1441 deps = MICROKERNEL_TEST_DEPS,
1442)
1443
1444xnnpack_unit_test(
1445 name = "f32_spmm_test",
1446 srcs = [
1447 "test/f32-spmm.cc",
1448 "test/spmm-microkernel-tester.h",
1449 "src/xnnpack/AlignedAllocator.h",
1450 ] + MICROKERNEL_TEST_HDRS,
1451 deps = MICROKERNEL_TEST_DEPS,
1452)
1453
1454xnnpack_unit_test(
1455 name = "f32_vadd_test",
1456 srcs = [
1457 "test/f32-vadd.cc",
1458 "test/vadd-microkernel-tester.h",
1459 ] + MICROKERNEL_TEST_HDRS,
1460 deps = MICROKERNEL_TEST_DEPS,
1461)
1462
1463xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001464 name = "f32_vmul_test",
1465 srcs = [
1466 "test/f32-vmul.cc",
1467 "test/vmul-microkernel-tester.h",
1468 ] + MICROKERNEL_TEST_HDRS,
1469 deps = MICROKERNEL_TEST_DEPS,
1470)
1471
1472xnnpack_unit_test(
1473 name = "f32_vmulcaddc_test",
1474 srcs = [
1475 "test/f32-vmulcaddc.cc",
1476 "test/vmulcaddc-microkernel-tester.h",
1477 "src/xnnpack/AlignedAllocator.h",
1478 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1479 deps = MICROKERNEL_TEST_DEPS,
1480)
1481
1482xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001483 name = "f32_vscale_test",
1484 srcs = [
1485 "test/f32-vscale.cc",
1486 "test/vscale-microkernel-tester.h",
1487 ] + MICROKERNEL_TEST_HDRS,
1488 deps = MICROKERNEL_TEST_DEPS,
1489)
1490
1491xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001492 name = "f32_vscaleexpminusmax_test",
1493 srcs = [
1494 "test/f32-vscaleexpminusmax.cc",
1495 "test/vscaleexpminusmax-microkernel-tester.h",
1496 ] + MICROKERNEL_TEST_HDRS,
1497 deps = MICROKERNEL_TEST_DEPS,
1498)
1499
1500xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001501 name = "f32_vscaleextexp_test",
1502 srcs = [
1503 "test/f32-vscaleextexp.cc",
1504 "test/vscaleextexp-microkernel-tester.h",
1505 ] + MICROKERNEL_TEST_HDRS,
1506 deps = MICROKERNEL_TEST_DEPS,
1507)
1508
1509xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001510 name = "f32_vsub_test",
1511 srcs = [
1512 "test/f32-vsub.cc",
1513 "test/vsub-microkernel-tester.h",
1514 ] + MICROKERNEL_TEST_HDRS,
1515 deps = MICROKERNEL_TEST_DEPS,
1516)
1517
1518xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001519 name = "q8_avgpool_test",
1520 srcs = [
1521 "test/q8-avgpool.cc",
1522 "test/avgpool-microkernel-tester.h",
1523 "src/xnnpack/AlignedAllocator.h",
1524 ] + MICROKERNEL_TEST_HDRS,
1525 deps = MICROKERNEL_TEST_DEPS,
1526)
1527
1528xnnpack_unit_test(
1529 name = "q8_igemm_test",
1530 srcs = [
1531 "test/q8-igemm.cc",
1532 "test/gemm-microkernel-tester.h",
1533 "src/xnnpack/AlignedAllocator.h",
1534 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1535 deps = MICROKERNEL_TEST_DEPS,
1536)
1537
1538xnnpack_unit_test(
1539 name = "q8_dwconv_test",
1540 srcs = [
1541 "test/q8-dwconv.cc",
1542 "test/dwconv-microkernel-tester.h",
1543 "src/xnnpack/AlignedAllocator.h",
1544 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1545 deps = MICROKERNEL_TEST_DEPS,
1546)
1547
1548xnnpack_unit_test(
1549 name = "q8_gavgpool_test",
1550 srcs = [
1551 "test/q8-gavgpool.cc",
1552 "test/gavgpool-microkernel-tester.h",
1553 "src/xnnpack/AlignedAllocator.h",
1554 ] + MICROKERNEL_TEST_HDRS,
1555 deps = MICROKERNEL_TEST_DEPS,
1556)
1557
1558xnnpack_unit_test(
1559 name = "q8_gemm_test",
1560 srcs = [
1561 "test/q8-gemm.cc",
1562 "test/gemm-microkernel-tester.h",
1563 "src/xnnpack/AlignedAllocator.h",
1564 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1565 deps = MICROKERNEL_TEST_DEPS,
1566)
1567
1568xnnpack_unit_test(
1569 name = "q8_vadd_test",
1570 srcs = [
1571 "test/q8-vadd.cc",
1572 "test/vadd-microkernel-tester.h",
1573 ] + MICROKERNEL_TEST_HDRS,
1574 deps = MICROKERNEL_TEST_DEPS,
1575)
1576
1577xnnpack_unit_test(
1578 name = "u8_clamp_test",
1579 srcs = [
1580 "test/u8-clamp.cc",
1581 "test/clamp-microkernel-tester.h",
1582 ] + MICROKERNEL_TEST_HDRS,
1583 deps = MICROKERNEL_TEST_DEPS,
1584)
1585
1586xnnpack_unit_test(
1587 name = "u8_lut32norm_test",
1588 srcs = [
1589 "test/u8-lut32norm.cc",
1590 "test/lut-norm-microkernel-tester.h",
1591 ] + MICROKERNEL_TEST_HDRS,
1592 deps = MICROKERNEL_TEST_DEPS,
1593)
1594
1595xnnpack_unit_test(
1596 name = "u8_maxpool_test",
1597 srcs = [
1598 "test/u8-maxpool.cc",
1599 "test/maxpool-microkernel-tester.h",
1600 ] + MICROKERNEL_TEST_HDRS,
1601 deps = MICROKERNEL_TEST_DEPS,
1602)
1603
1604xnnpack_unit_test(
1605 name = "u8_rmax_test",
1606 srcs = [
1607 "test/u8-rmax.cc",
1608 "test/rmax-microkernel-tester.h",
1609 ] + MICROKERNEL_TEST_HDRS,
1610 deps = MICROKERNEL_TEST_DEPS,
1611)
1612
1613xnnpack_unit_test(
1614 name = "x32_packx_test",
1615 srcs = [
1616 "test/x32-packx.cc",
1617 "test/pack-microkernel-tester.h",
1618 "src/xnnpack/AlignedAllocator.h",
1619 ] + MICROKERNEL_TEST_HDRS,
1620 deps = MICROKERNEL_TEST_DEPS,
1621)
1622
1623xnnpack_unit_test(
1624 name = "x32_pad_test",
1625 srcs = [
1626 "test/x32-pad.cc",
1627 "test/pad-microkernel-tester.h",
1628 ] + MICROKERNEL_TEST_HDRS,
1629 deps = MICROKERNEL_TEST_DEPS,
1630)
1631
1632xnnpack_unit_test(
1633 name = "x32_unpool_test",
1634 srcs = [
1635 "test/x32-unpool.cc",
1636 "test/unpool-microkernel-tester.h",
1637 ] + MICROKERNEL_TEST_HDRS,
1638 deps = MICROKERNEL_TEST_DEPS,
1639)
1640
1641xnnpack_unit_test(
1642 name = "x32_zip_test",
1643 srcs = [
1644 "test/x32-zip.cc",
1645 "test/zip-microkernel-tester.h",
1646 ] + MICROKERNEL_TEST_HDRS,
1647 deps = MICROKERNEL_TEST_DEPS,
1648)
1649
1650xnnpack_unit_test(
1651 name = "x8_lut_test",
1652 srcs = [
1653 "test/x8-lut.cc",
1654 "test/lut-microkernel-tester.h",
1655 ] + MICROKERNEL_TEST_HDRS,
1656 deps = MICROKERNEL_TEST_DEPS,
1657)
1658
1659xnnpack_unit_test(
1660 name = "x8_zip_test",
1661 srcs = [
1662 "test/x8-zip.cc",
1663 "test/zip-microkernel-tester.h",
1664 ] + MICROKERNEL_TEST_HDRS,
1665 deps = MICROKERNEL_TEST_DEPS,
1666)
1667
1668########################### Size test for the library ##########################
1669
1670xnnpack_binary(
1671 name = "size_test",
1672 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001673 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001674)
1675
1676########################### Unit tests for operators ###########################
1677
1678xnnpack_unit_test(
1679 name = "add_test",
1680 srcs = [
1681 "test/add.cc",
1682 "test/add-operator-tester.h",
1683 ],
1684 deps = OPERATOR_TEST_DEPS,
1685)
1686
1687xnnpack_unit_test(
1688 name = "argmax_pooling_test",
1689 srcs = [
1690 "test/argmax-pooling.cc",
1691 "test/argmax-pooling-operator-tester.h",
1692 ] + OPERATOR_TEST_PARAMS_HDRS,
1693 deps = OPERATOR_TEST_DEPS,
1694)
1695
1696xnnpack_unit_test(
1697 name = "average_pooling_test",
1698 srcs = [
1699 "test/average-pooling.cc",
1700 "test/average-pooling-operator-tester.h",
1701 ] + OPERATOR_TEST_PARAMS_HDRS,
1702 deps = OPERATOR_TEST_DEPS,
1703)
1704
1705xnnpack_unit_test(
1706 name = "channel_pad_test",
1707 srcs = [
1708 "test/channel-pad.cc",
1709 "test/channel-pad-operator-tester.h",
1710 ] + OPERATOR_TEST_PARAMS_HDRS,
1711 deps = OPERATOR_TEST_DEPS,
1712)
1713
1714xnnpack_unit_test(
1715 name = "channel_shuffle_test",
1716 srcs = [
1717 "test/channel-shuffle.cc",
1718 "test/channel-shuffle-operator-tester.h",
1719 ],
1720 deps = OPERATOR_TEST_DEPS,
1721)
1722
1723xnnpack_unit_test(
1724 name = "clamp_test",
1725 srcs = [
1726 "test/clamp.cc",
1727 "test/clamp-operator-tester.h",
1728 ],
1729 deps = OPERATOR_TEST_DEPS,
1730)
1731
1732xnnpack_unit_test(
1733 name = "convolution_test",
1734 srcs = [
1735 "test/convolution.cc",
1736 "test/convolution-operator-tester.h",
1737 ],
1738 deps = OPERATOR_TEST_DEPS,
1739)
1740
1741xnnpack_unit_test(
1742 name = "convolution_spnchw_test",
1743 srcs = [
1744 "test/convolution-spnchw.cc",
1745 "test/convolution-spnchw-operator-tester.h",
1746 ],
1747 deps = OPERATOR_TEST_DEPS,
1748)
1749
1750xnnpack_unit_test(
1751 name = "deconvolution_test",
1752 srcs = [
1753 "test/deconvolution.cc",
1754 "test/deconvolution-operator-tester.h",
1755 ] + OPERATOR_TEST_PARAMS_HDRS,
1756 deps = OPERATOR_TEST_DEPS,
1757)
1758
1759xnnpack_unit_test(
1760 name = "fully_connected_test",
1761 srcs = [
1762 "test/fully-connected.cc",
1763 "test/fully-connected-operator-tester.h",
1764 ],
1765 deps = OPERATOR_TEST_DEPS,
1766)
1767
1768xnnpack_unit_test(
1769 name = "global_average_pooling_test",
1770 srcs = [
1771 "test/global-average-pooling.cc",
1772 "test/global-average-pooling-operator-tester.h",
1773 ] + OPERATOR_TEST_PARAMS_HDRS,
1774 deps = OPERATOR_TEST_DEPS,
1775)
1776
1777xnnpack_unit_test(
1778 name = "global_average_pooling_spnchw_test",
1779 srcs = [
1780 "test/global-average-pooling-spnchw.cc",
1781 "test/global-average-pooling-spnchw-operator-tester.h",
1782 ],
1783 deps = OPERATOR_TEST_DEPS,
1784)
1785
1786xnnpack_unit_test(
1787 name = "hardswish_test",
1788 srcs = [
1789 "test/hardswish.cc",
1790 "test/hardswish-operator-tester.h",
1791 ],
1792 deps = OPERATOR_TEST_DEPS,
1793)
1794
1795xnnpack_unit_test(
1796 name = "leaky_relu_test",
1797 srcs = [
1798 "test/leaky-relu.cc",
1799 "test/leaky-relu-operator-tester.h",
1800 ],
1801 deps = OPERATOR_TEST_DEPS,
1802)
1803
1804xnnpack_unit_test(
1805 name = "max_pooling_test",
1806 srcs = [
1807 "test/max-pooling.cc",
1808 "test/max-pooling-operator-tester.h",
1809 ] + OPERATOR_TEST_PARAMS_HDRS,
1810 deps = OPERATOR_TEST_DEPS,
1811)
1812
1813xnnpack_unit_test(
1814 name = "prelu_test",
1815 srcs = [
1816 "test/prelu.cc",
1817 "test/prelu-operator-tester.h",
1818 ] + OPERATOR_TEST_PARAMS_HDRS,
1819 deps = OPERATOR_TEST_DEPS,
1820)
1821
1822xnnpack_unit_test(
1823 name = "sigmoid_test",
1824 srcs = [
1825 "test/sigmoid.cc",
1826 "test/sigmoid-operator-tester.h",
1827 ],
1828 deps = OPERATOR_TEST_DEPS,
1829)
1830
1831xnnpack_unit_test(
1832 name = "softargmax_test",
1833 srcs = [
1834 "test/softargmax.cc",
1835 "test/softargmax-operator-tester.h",
1836 ],
1837 deps = OPERATOR_TEST_DEPS,
1838)
1839
1840xnnpack_unit_test(
1841 name = "unpooling_test",
1842 srcs = [
1843 "test/unpooling.cc",
1844 "test/unpooling-operator-tester.h",
1845 ],
1846 deps = OPERATOR_TEST_DEPS,
1847)
1848
1849############################# Build configurations #############################
1850
Marat Dukhanb8642352019-10-30 15:43:02 -07001851# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07001852config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07001853 name = "xnn_enable_assembly_explicit_true",
1854 define_values = {"xnn_enable_assembly": "true"},
1855)
1856
1857# Disables usage of assembly kernels.
1858config_setting(
1859 name = "xnn_enable_assembly_explicit_false",
1860 define_values = {"xnn_enable_assembly": "false"},
1861)
1862
1863# Builds with -c dbg
1864config_setting(
1865 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001866 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07001867 "compilation_mode": "dbg",
1868 },
1869)
1870
1871# Builds with -c opt
1872config_setting(
1873 name = "optimized_build",
1874 values = {
1875 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001876 },
1877)
1878
1879config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07001880 name = "linux_k8",
1881 values = {"cpu": "k8"},
1882)
1883
1884config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07001885 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07001886 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07001887)
1888
1889config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001890 name = "android",
1891 values = {"crosstool_top": "//external:android/crosstool"},
1892)
1893
1894config_setting(
1895 name = "android_armv7",
1896 values = {
1897 "crosstool_top": "//external:android/crosstool",
1898 "cpu": "armeabi-v7a",
1899 },
1900)
1901
1902config_setting(
1903 name = "android_arm64",
1904 values = {
1905 "crosstool_top": "//external:android/crosstool",
1906 "cpu": "arm64-v8a",
1907 },
1908)
1909
1910config_setting(
1911 name = "android_x86",
1912 values = {
1913 "crosstool_top": "//external:android/crosstool",
1914 "cpu": "x86",
1915 },
1916)
1917
1918config_setting(
1919 name = "android_x86_64",
1920 values = {
1921 "crosstool_top": "//external:android/crosstool",
1922 "cpu": "x86_64",
1923 },
1924)
1925
1926config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07001927 name = "macos_x86_64",
1928 values = {
1929 "apple_platform_type": "macos",
1930 "cpu": "darwin",
1931 },
1932)
1933
1934config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001935 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001936 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07001937)
1938
1939config_setting(
1940 name = "emscripten_wasm",
1941 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001942 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001943 "cpu": "wasm",
1944 },
1945)
1946
1947config_setting(
1948 name = "emscripten_wasmsimd",
1949 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001950 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001951 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07001952 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001953 },
1954)
1955
1956config_setting(
1957 name = "emscripten_asmjs",
1958 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07001959 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001960 "cpu": "asmjs",
1961 },
1962)