blob: 4bcf91aa81b9b70b31aaa7259c48c7bda33c630c [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 = [
Marat Dukhanefc47b82019-11-18 09:25:38 -080053 "src/add-nc.c",
54 "src/argmax-pooling-nhwc.c",
55 "src/average-pooling-nhwc.c",
56 "src/channel-pad-nc.c",
57 "src/channel-shuffle-nc.c",
58 "src/clamp-nc.c",
59 "src/convolution-nchw.c",
60 "src/convolution-nhwc.c",
61 "src/deconvolution-nhwc.c",
62 "src/fully-connected-nc.c",
63 "src/global-average-pooling-ncw.c",
64 "src/global-average-pooling-nwc.c",
65 "src/hardswish-nc.c",
66 "src/leaky-relu-nc.c",
67 "src/max-pooling-nhwc.c",
68 "src/multiply-nd.c",
69 "src/prelu-nc.c",
70 "src/resize-bilinear-nhwc.c",
71 "src/sigmoid-nc.c",
72 "src/softargmax-nc.c",
73 "src/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070074]
75
76SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -080077 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080078 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -080079 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070080 "src/f32-avgpool/mp9p8q-scalar.c",
81 "src/f32-avgpool/up9-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -080082 "src/f32-bilinear/gen/scalar-c1.c",
83 "src/f32-bilinear/gen/scalar-c2.c",
84 "src/f32-bilinear/gen/scalar-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070085 "src/f32-clamp/scalar.c",
Erich Elsen563df5f2019-10-23 08:02:21 -070086 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen0cc2c532019-10-15 04:44:18 -070087 "src/f32-dwconv-spchw/3x3p1-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -070088 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080089 "src/f32-dwconv-spchw/5x5p2-scalar.c",
Erich Elsen38709a62019-11-08 11:58:45 -080090 "src/f32-dwconv-spchw/5x5s2p2-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -080091 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
92 "src/f32-dwconv/gen/up1x25-scalar.c",
93 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
94 "src/f32-dwconv/gen/up1x4-scalar.c",
95 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
96 "src/f32-dwconv/gen/up1x9-scalar.c",
97 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
98 "src/f32-dwconv/gen/up2x25-scalar.c",
99 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
100 "src/f32-dwconv/gen/up2x4-scalar.c",
101 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
102 "src/f32-dwconv/gen/up2x9-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -0700103 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700104 "src/f32-gavgpool/mp7p7q-scalar.c",
105 "src/f32-gavgpool/up7-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800106 "src/f32-gemm/gen-inc/1x4-scalar.c",
107 "src/f32-gemm/gen-inc/2x4-scalar.c",
108 "src/f32-gemm/gen-inc/4x4-scalar.c",
109 "src/f32-gemm/gen/1x4-scalar.c",
110 "src/f32-gemm/gen/2x4-scalar.c",
111 "src/f32-gemm/gen/4x2-scalar.c",
112 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700113 "src/f32-hswish/scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800114 "src/f32-igemm/gen/1x4-scalar.c",
115 "src/f32-igemm/gen/2x4-scalar.c",
116 "src/f32-igemm/gen/4x2-scalar.c",
117 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-maxpool/9p8x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700119 "src/f32-pavgpool/mp9p8q-scalar.c",
120 "src/f32-pavgpool/up9-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800121 "src/f32-ppmm/gen/2x4-scalar.c",
122 "src/f32-ppmm/gen/3x3-scalar.c",
123 "src/f32-ppmm/gen/4x2-scalar.c",
124 "src/f32-ppmm/gen/4x4-scalar.c",
125 "src/f32-prelu/gen/scalar-2x1.c",
126 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700127 "src/f32-rmax/scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800128 "src/f32-spmm/gen/1x1-scalar-pipelined.c",
129 "src/f32-spmm/gen/1x1-scalar.c",
130 "src/f32-spmm/gen/2x1-scalar-pipelined.c",
131 "src/f32-spmm/gen/2x1-scalar.c",
132 "src/f32-spmm/gen/4x1-scalar-pipelined.c",
133 "src/f32-spmm/gen/4x1-scalar.c",
134 "src/f32-spmm/gen/8x1-scalar-pipelined.c",
135 "src/f32-spmm/gen/8x1-scalar.c",
136 "src/f32-spmm/gen/8x2-scalar.c",
137 "src/f32-spmm/gen/8x4-scalar.c",
138 "src/f32-vbinary/gen/vadd-scalar-x1.c",
139 "src/f32-vbinary/gen/vadd-scalar-x2.c",
140 "src/f32-vbinary/gen/vadd-scalar-x4.c",
141 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
142 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
143 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
144 "src/f32-vbinary/gen/vmul-scalar-x1.c",
145 "src/f32-vbinary/gen/vmul-scalar-x2.c",
146 "src/f32-vbinary/gen/vmul-scalar-x4.c",
147 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
148 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
149 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
150 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
151 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
152 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
153 "src/f32-vbinary/gen/vsub-scalar-x1.c",
154 "src/f32-vbinary/gen/vsub-scalar-x2.c",
155 "src/f32-vbinary/gen/vsub-scalar-x4.c",
156 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
157 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
158 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
159 "src/f32-vmulcaddc/gen/c1-scalar-2x.c",
160 "src/f32-vmulcaddc/gen/c2-scalar-2x.c",
161 "src/f32-vmulcaddc/gen/c4-scalar-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700162 "src/q8-avgpool/mp9p8q-scalar.c",
163 "src/q8-avgpool/up9-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700164 "src/q8-dwconv/up1x9-scalar.c",
165 "src/q8-gavgpool/mp7p7q-scalar.c",
166 "src/q8-gavgpool/up7-scalar.c",
167 "src/q8-gemm/2x2-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800168 "src/q8-igemm/2x2-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700169 "src/q8-vadd/scalar.c",
170 "src/u8-clamp/scalar.c",
171 "src/u8-lut32norm/scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800172 "src/u8-maxpool/9p8x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700173 "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 = [
Marat Dukhan329da642019-11-19 21:44:39 -0800191 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800192 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800193 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700194 "src/f32-avgpool/mp9p8q-psimd.c",
195 "src/f32-avgpool/up9-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800196 "src/f32-bilinear/gen/psimd-c4.c",
197 "src/f32-bilinear/gen/psimd-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700198 "src/f32-clamp/psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800199 "src/f32-dwconv/gen/up4x25-psimd-acc2.c",
200 "src/f32-dwconv/gen/up4x25-psimd.c",
201 "src/f32-dwconv/gen/up4x4-psimd-acc2.c",
202 "src/f32-dwconv/gen/up4x4-psimd.c",
203 "src/f32-dwconv/gen/up4x9-psimd-acc2.c",
204 "src/f32-dwconv/gen/up4x9-psimd.c",
205 "src/f32-dwconv/gen/up8x25-psimd-acc2.c",
206 "src/f32-dwconv/gen/up8x25-psimd.c",
207 "src/f32-dwconv/gen/up8x4-psimd-acc2.c",
208 "src/f32-dwconv/gen/up8x4-psimd.c",
209 "src/f32-dwconv/gen/up8x9-psimd-acc2.c",
210 "src/f32-dwconv/gen/up8x9-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700211 "src/f32-gavgpool/mp7p7q-psimd.c",
212 "src/f32-gavgpool/up7-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800213 "src/f32-gemm/gen/1x8-psimd-loadsplat.c",
214 "src/f32-gemm/gen/1x8-psimd-splat.c",
215 "src/f32-gemm/gen/1x8s4-psimd.c",
216 "src/f32-gemm/gen/4x8-psimd-loadsplat.c",
217 "src/f32-gemm/gen/4x8-psimd-splat.c",
218 "src/f32-gemm/gen/4x8s4-psimd.c",
219 "src/f32-gemm/gen/6x8-psimd-loadsplat.c",
220 "src/f32-gemm/gen/6x8-psimd-splat.c",
221 "src/f32-gemm/gen/6x8s4-psimd.c",
222 "src/f32-gemm/gen-inc/1x8-psimd-loadsplat.c",
223 "src/f32-gemm/gen-inc/1x8-psimd-splat.c",
224 "src/f32-gemm/gen-inc/1x8s4-psimd.c",
225 "src/f32-gemm/gen-inc/4x8-psimd-loadsplat.c",
226 "src/f32-gemm/gen-inc/4x8-psimd-splat.c",
227 "src/f32-gemm/gen-inc/4x8s4-psimd.c",
228 "src/f32-gemm/gen-inc/6x8-psimd-loadsplat.c",
229 "src/f32-gemm/gen-inc/6x8-psimd-splat.c",
230 "src/f32-gemm/gen-inc/6x8s4-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700231 "src/f32-hswish/psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800232 "src/f32-igemm/gen/1x8-psimd-loadsplat.c",
233 "src/f32-igemm/gen/1x8-psimd-splat.c",
234 "src/f32-igemm/gen/1x8s4-psimd.c",
235 "src/f32-igemm/gen/4x2c4-psimd.c",
236 "src/f32-igemm/gen/4x8-psimd-loadsplat.c",
237 "src/f32-igemm/gen/4x8-psimd-splat.c",
238 "src/f32-igemm/gen/4x8s4-psimd.c",
239 "src/f32-igemm/gen/6x8-psimd-loadsplat.c",
240 "src/f32-igemm/gen/6x8-psimd-splat.c",
241 "src/f32-igemm/gen/6x8s4-psimd.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800242 "src/f32-maxpool/9p8x-psimd-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700243 "src/f32-pavgpool/mp9p8q-psimd.c",
244 "src/f32-pavgpool/up9-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800245 "src/f32-ppmm/gen/4x8-psimd.c",
246 "src/f32-prelu/gen/psimd-2x4.c",
247 "src/f32-prelu/gen/psimd-2x8.c",
248 "src/f32-vbinary/gen/vadd-psimd-x4.c",
249 "src/f32-vbinary/gen/vadd-psimd-x8.c",
250 "src/f32-vbinary/gen/vaddc-psimd-x4.c",
251 "src/f32-vbinary/gen/vaddc-psimd-x8.c",
252 "src/f32-vbinary/gen/vmul-psimd-x4.c",
253 "src/f32-vbinary/gen/vmul-psimd-x8.c",
254 "src/f32-vbinary/gen/vmulc-psimd-x4.c",
255 "src/f32-vbinary/gen/vmulc-psimd-x8.c",
256 "src/f32-vbinary/gen/vrsubc-psimd-x4.c",
257 "src/f32-vbinary/gen/vrsubc-psimd-x8.c",
258 "src/f32-vbinary/gen/vsub-psimd-x4.c",
259 "src/f32-vbinary/gen/vsub-psimd-x8.c",
260 "src/f32-vbinary/gen/vsubc-psimd-x4.c",
261 "src/f32-vbinary/gen/vsubc-psimd-x8.c",
262 "src/f32-vmulcaddc/gen/c4-psimd-2x.c",
263 "src/f32-vmulcaddc/gen/c8-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700264 "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 Dukhan40a672f2019-11-25 03:08:22 -0800277 "src/f32-bilinear/gen/neon-c4.c",
278 "src/f32-bilinear/gen/neon-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700279 "src/f32-clamp/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800280 "src/f32-dwconv/gen/up4x9-neon.c",
281 "src/f32-dwconv/gen/up4x9-neon-acc2.c",
282 "src/f32-dwconv/gen/up8x9-neon.c",
283 "src/f32-dwconv/gen/up8x9-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700284 "src/f32-gavgpool-spchw/neon-x4.c",
285 "src/f32-gavgpool/mp7p7q-neon.c",
286 "src/f32-gavgpool/up7-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800287 "src/f32-gemm/gen/1x8-neon-lane-ld64.c",
288 "src/f32-gemm/gen/4x2-neon-lane-ld64.c",
289 "src/f32-gemm/gen/4x8-neon-lane-ld128.c",
290 "src/f32-gemm/gen/4x8-neon-lane-ld64.c",
291 "src/f32-gemm/gen/5x8-neon-lane-ld64.c",
292 "src/f32-gemm/gen/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800293 "src/f32-gemm/gen/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800294 "src/f32-gemm/gen/1x8-neon-dup-ld64.c",
295 "src/f32-gemm/gen/4x8-neon-dup-ld128.c",
296 "src/f32-gemm/gen/4x8-neon-dup-ld64.c",
297 "src/f32-gemm/gen/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800298 "src/f32-gemm/gen/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800299 "src/f32-gemm/gen/1x8s4-neon.c",
300 "src/f32-gemm/gen/4x8s4-neon.c",
301 "src/f32-gemm/gen/6x8s4-neon.c",
302 "src/f32-gemm/gen/8x8s4-neon.c",
303 "src/f32-gemm/gen-inc/1x8-neon-lane-ld64.c",
304 "src/f32-gemm/gen-inc/4x8-neon-lane-ld128.c",
305 "src/f32-gemm/gen-inc/4x8-neon-lane-ld64.c",
306 "src/f32-gemm/gen-inc/5x8-neon-lane-ld64.c",
307 "src/f32-gemm/gen-inc/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800308 "src/f32-gemm/gen-inc/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800309 "src/f32-gemm/gen-inc/1x8-neon-dup-ld64.c",
310 "src/f32-gemm/gen-inc/4x8-neon-dup-ld128.c",
311 "src/f32-gemm/gen-inc/4x8-neon-dup-ld64.c",
312 "src/f32-gemm/gen-inc/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800313 "src/f32-gemm/gen-inc/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800314 "src/f32-gemm/gen-inc/1x8s4-neon.c",
315 "src/f32-gemm/gen-inc/4x8s4-neon.c",
316 "src/f32-gemm/gen-inc/6x8s4-neon.c",
317 "src/f32-gemm/gen-inc/8x8s4-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700318 "src/f32-hswish/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800319 "src/f32-igemm/gen/1x8-neon-lane-ld64.c",
320 "src/f32-igemm/gen/4x2-neon-lane-ld64.c",
321 "src/f32-igemm/gen/4x4-neon-lane-ld64.c",
322 "src/f32-igemm/gen/4x8-neon-lane-ld128.c",
323 "src/f32-igemm/gen/4x8-neon-lane-ld64.c",
324 "src/f32-igemm/gen/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800325 "src/f32-igemm/gen/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800326 "src/f32-igemm/gen/1x8-neon-dup-ld64.c",
327 "src/f32-igemm/gen/4x8-neon-dup-ld128.c",
328 "src/f32-igemm/gen/4x8-neon-dup-ld64.c",
329 "src/f32-igemm/gen/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800330 "src/f32-igemm/gen/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800331 "src/f32-igemm/gen/1x8s4-neon.c",
332 "src/f32-igemm/gen/4x8s4-neon.c",
333 "src/f32-igemm/gen/6x8s4-neon.c",
334 "src/f32-igemm/gen/8x8s4-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700335 "src/f32-pavgpool/mp9p8q-neon.c",
336 "src/f32-pavgpool/up9-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800337 "src/f32-ppmm/gen/4x8-neon.c",
338 "src/f32-ppmm/gen/8x8-neon.c",
339 "src/f32-prelu/gen/neon-2x4.c",
340 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700341 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800342 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
343 "src/f32-vbinary/gen/vadd-neon-x4.c",
344 "src/f32-vbinary/gen/vadd-neon-x8.c",
345 "src/f32-vbinary/gen/vaddc-neon-x4.c",
346 "src/f32-vbinary/gen/vaddc-neon-x8.c",
347 "src/f32-vbinary/gen/vmul-neon-x4.c",
348 "src/f32-vbinary/gen/vmul-neon-x8.c",
349 "src/f32-vbinary/gen/vmulc-neon-x4.c",
350 "src/f32-vbinary/gen/vmulc-neon-x8.c",
351 "src/f32-vbinary/gen/vrsubc-neon-x4.c",
352 "src/f32-vbinary/gen/vrsubc-neon-x8.c",
353 "src/f32-vbinary/gen/vsub-neon-x4.c",
354 "src/f32-vbinary/gen/vsub-neon-x8.c",
355 "src/f32-vbinary/gen/vsubc-neon-x4.c",
356 "src/f32-vbinary/gen/vsubc-neon-x8.c",
357 "src/f32-vmulcaddc/gen/c4-neon-2x.c",
358 "src/f32-vmulcaddc/gen/c8-neon-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700359 "src/q8-avgpool/mp9p8q-neon.c",
360 "src/q8-avgpool/up9-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700361 "src/q8-dwconv/up8x9-neon.c",
362 "src/q8-gavgpool/mp7p7q-neon.c",
363 "src/q8-gavgpool/up7-neon.c",
364 "src/q8-gemm/4x8-neon.c",
365 "src/q8-gemm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800366 "src/q8-igemm/4x8-neon.c",
367 "src/q8-igemm/8x8-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700368 "src/q8-vadd/neon.c",
369 "src/u8-clamp/neon.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800370 "src/u8-maxpool/9p8x-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700371 "src/u8-rmax/neon.c",
372 "src/x32-packx/x4-neon-st4.c",
373 "src/x32-pad/x2-neon.c",
374 "src/x32-zip/x2-neon.c",
375 "src/x32-zip/x3-neon.c",
376 "src/x32-zip/x4-neon.c",
377 "src/x32-zip/xm-neon.c",
378 "src/x8-zip/x2-neon.c",
379 "src/x8-zip/x3-neon.c",
380 "src/x8-zip/x4-neon.c",
381 "src/x8-zip/xm-neon.c",
382]
383
384NEONFMA_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800385 "src/f32-bilinear/gen/neonfma-c4.c",
386 "src/f32-bilinear/gen/neonfma-c8.c",
387 "src/f32-igemm/gen/1x8-neonfma-dup-ld64.c",
388 "src/f32-igemm/gen/4x8-neonfma-dup-ld128.c",
389 "src/f32-igemm/gen/4x8-neonfma-dup-ld64.c",
390 "src/f32-igemm/gen/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800391 "src/f32-igemm/gen/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800392 "src/f32-igemm/gen/1x8s4-neonfma.c",
393 "src/f32-igemm/gen/4x8s4-neonfma.c",
394 "src/f32-igemm/gen/6x8s4-neonfma.c",
395 "src/f32-igemm/gen/8x8s4-neonfma.c",
396 "src/f32-dwconv/gen/up4x9-neonfma.c",
397 "src/f32-dwconv/gen/up4x9-neonfma-acc2.c",
398 "src/f32-dwconv/gen/up8x9-neonfma.c",
399 "src/f32-dwconv/gen/up8x9-neonfma-acc2.c",
400 "src/f32-gemm/gen/1x8-neonfma-dup-ld64.c",
401 "src/f32-gemm/gen/4x8-neonfma-dup-ld128.c",
402 "src/f32-gemm/gen/4x8-neonfma-dup-ld64.c",
403 "src/f32-gemm/gen/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800404 "src/f32-gemm/gen/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800405 "src/f32-gemm/gen/1x8s4-neonfma.c",
406 "src/f32-gemm/gen/4x8s4-neonfma.c",
407 "src/f32-gemm/gen/6x8s4-neonfma.c",
408 "src/f32-gemm/gen/8x8s4-neonfma.c",
409 "src/f32-gemm/gen-inc/1x8-neonfma-dup-ld64.c",
410 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld128.c",
411 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld64.c",
412 "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800413 "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800414 "src/f32-gemm/gen-inc/1x8s4-neonfma.c",
415 "src/f32-gemm/gen-inc/4x8s4-neonfma.c",
416 "src/f32-gemm/gen-inc/6x8s4-neonfma.c",
417 "src/f32-gemm/gen-inc/8x8s4-neonfma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700418 "src/f32-hswish/neonfma.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800419 "src/f32-ppmm/gen/4x8-neonfma.c",
420 "src/f32-ppmm/gen/8x8-neonfma.c",
421 "src/f32-sigmoid/gen/neonfma-p5-nr2fma-x16.c",
422 "src/f32-vmulcaddc/gen/c4-neonfma-2x.c",
423 "src/f32-vmulcaddc/gen/c8-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800424 "src/math/exp-neonfma-lut64-p2.c",
425 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800426 "src/math/expminus-neonfma-lut2048-p1.c",
427 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800428 "src/math/expminus-neonfma-p5.c",
Marat Dukhan91f8d862019-11-27 12:25:42 -0800429 "src/math/sigmoid-neonfma-lut2048-p1-nr1recps1fma.c",
430 "src/math/sigmoid-neonfma-lut2048-p1-nr2fma.c",
431 "src/math/sigmoid-neonfma-lut2048-p1-nr2recps.c",
Marat Dukhan22aae132019-11-22 17:10:29 -0800432 "src/math/sigmoid-neonfma-p5-nr1recps1fma.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -0800433 "src/math/sigmoid-neonfma-p5-nr2fma.c",
Marat Dukhan22aae132019-11-22 17:10:29 -0800434 "src/math/sigmoid-neonfma-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700435]
436
437AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800438 "src/f32-gemm/gen/1x8-neonfma-lane-ld64.c",
439 "src/f32-gemm/gen/4x2-neonfma-lane-ld64.c",
440 "src/f32-gemm/gen/4x8-neonfma-lane-ld128.c",
441 "src/f32-gemm/gen/4x8-neonfma-lane-ld64.c",
442 "src/f32-gemm/gen/5x8-neonfma-lane-ld64.c",
443 "src/f32-gemm/gen/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800444 "src/f32-gemm/gen/6x8-neonfma-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800445 "src/f32-gemm/gen-inc/1x8-neonfma-lane-ld64.c",
446 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld128.c",
447 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld64.c",
448 "src/f32-gemm/gen-inc/5x8-neonfma-lane-ld64.c",
449 "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800450 "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800451 "src/f32-igemm/gen/1x8-neonfma-lane-ld64.c",
452 "src/f32-igemm/gen/4x2-neonfma-lane-ld64.c",
453 "src/f32-igemm/gen/4x4-neonfma-lane-ld64.c",
454 "src/f32-igemm/gen/4x8-neonfma-lane-ld128.c",
455 "src/f32-igemm/gen/4x8-neonfma-lane-ld64.c",
456 "src/f32-igemm/gen/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800457 "src/f32-igemm/gen/6x8-neonfma-lane-ld128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700458 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
459 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
460 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
461 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
462 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
463 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
464 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800465 "src/f32-spmm/gen/12x1-neonfma.c",
466 "src/f32-spmm/gen/12x2-neonfma.c",
467 "src/f32-spmm/gen/12x4-neonfma.c",
468 "src/f32-spmm/gen/16x1-neonfma-pipelined.c",
469 "src/f32-spmm/gen/16x1-neonfma-unroll2.c",
470 "src/f32-spmm/gen/16x1-neonfma.c",
471 "src/f32-spmm/gen/16x2-neonfma.c",
472 "src/f32-spmm/gen/16x4-neonfma.c",
473 "src/f32-spmm/gen/4x1-neonfma-pipelined.c",
474 "src/f32-spmm/gen/4x1-neonfma-unroll2.c",
475 "src/f32-spmm/gen/4x1-neonfma.c",
476 "src/f32-spmm/gen/4x2-neonfma.c",
477 "src/f32-spmm/gen/4x4-neonfma.c",
478 "src/f32-spmm/gen/8x1-neonfma-pipelined.c",
479 "src/f32-spmm/gen/8x1-neonfma-unroll2.c",
480 "src/f32-spmm/gen/8x1-neonfma.c",
481 "src/f32-spmm/gen/8x2-neonfma.c",
482 "src/f32-spmm/gen/8x4-neonfma.c",
Marat Dukhan22aae132019-11-22 17:10:29 -0800483 "src/math/sigmoid-neonfma-p5-div.c",
Marat Dukhan91f8d862019-11-27 12:25:42 -0800484 "src/math/sigmoid-neonfma-lut2048-p1-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700485]
486
487AARCH64_NEONFP16ARITH_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800488 "src/f16-gemm/gen/4x8-neonfp16arith-ld64.c",
489 "src/f16-gemm/gen/6x8-neonfp16arith-ld64.c",
490 "src/f16-gemm/gen/8x8-neonfp16arith-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700491]
492
493SSE_UKERNELS = [
494 "src/f32-avgpool/mp9p8q-sse.c",
495 "src/f32-avgpool/up9-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800496 "src/f32-bilinear/gen/sse-c4.c",
497 "src/f32-bilinear/gen/sse-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700498 "src/f32-clamp/sse.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800499 "src/f32-dwconv-spchw/3x3p1-sse.c",
500 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800501 "src/f32-dwconv/gen/up4x25-sse-acc2.c",
502 "src/f32-dwconv/gen/up4x25-sse.c",
503 "src/f32-dwconv/gen/up4x4-sse-acc2.c",
504 "src/f32-dwconv/gen/up4x4-sse.c",
505 "src/f32-dwconv/gen/up4x9-sse-acc2.c",
506 "src/f32-dwconv/gen/up4x9-sse.c",
507 "src/f32-dwconv/gen/up8x25-sse-acc2.c",
508 "src/f32-dwconv/gen/up8x25-sse.c",
509 "src/f32-dwconv/gen/up8x4-sse-acc2.c",
510 "src/f32-dwconv/gen/up8x4-sse.c",
511 "src/f32-dwconv/gen/up8x9-sse-acc2.c",
512 "src/f32-dwconv/gen/up8x9-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700513 "src/f32-gavgpool-spchw/sse-x4.c",
514 "src/f32-gavgpool/mp7p7q-sse.c",
515 "src/f32-gavgpool/up7-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800516 "src/f32-gemm/gen/1x8-sse-dup.c",
517 "src/f32-gemm/gen/1x8-sse-load1.c",
518 "src/f32-gemm/gen/1x8s4-sse.c",
519 "src/f32-gemm/gen/4x8-sse-dup.c",
520 "src/f32-gemm/gen/4x8-sse-load1.c",
521 "src/f32-gemm/gen/4x8s4-sse.c",
522 "src/f32-gemm/gen-inc/1x8-sse-dup.c",
523 "src/f32-gemm/gen-inc/1x8-sse-load1.c",
524 "src/f32-gemm/gen-inc/1x8s4-sse.c",
525 "src/f32-gemm/gen-inc/4x8-sse-dup.c",
526 "src/f32-gemm/gen-inc/4x8-sse-load1.c",
527 "src/f32-gemm/gen-inc/4x8s4-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700528 "src/f32-hswish/sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800529 "src/f32-igemm/gen/1x8-sse-dup.c",
530 "src/f32-igemm/gen/1x8-sse-load1.c",
531 "src/f32-igemm/gen/1x8s4-sse.c",
532 "src/f32-igemm/gen/4x2c4-sse.c",
533 "src/f32-igemm/gen/4x8-sse-dup.c",
534 "src/f32-igemm/gen/4x8-sse-load1.c",
535 "src/f32-igemm/gen/4x8s4-sse.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800536 "src/f32-maxpool/9p8x-sse-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700537 "src/f32-pavgpool/mp9p8q-sse.c",
538 "src/f32-pavgpool/up9-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800539 "src/f32-ppmm/gen/4x8-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700540 "src/f32-rmax/sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800541 "src/f32-spmm/gen/4x1-sse.c",
542 "src/f32-spmm/gen/8x1-sse.c",
543 "src/f32-vbinary/gen/vadd-sse-x4.c",
544 "src/f32-vbinary/gen/vadd-sse-x8.c",
545 "src/f32-vbinary/gen/vaddc-sse-x4.c",
546 "src/f32-vbinary/gen/vaddc-sse-x8.c",
547 "src/f32-vbinary/gen/vmul-sse-x4.c",
548 "src/f32-vbinary/gen/vmul-sse-x8.c",
549 "src/f32-vbinary/gen/vmulc-sse-x4.c",
550 "src/f32-vbinary/gen/vmulc-sse-x8.c",
551 "src/f32-vbinary/gen/vrsubc-sse-x4.c",
552 "src/f32-vbinary/gen/vrsubc-sse-x8.c",
553 "src/f32-vbinary/gen/vsub-sse-x4.c",
554 "src/f32-vbinary/gen/vsub-sse-x8.c",
555 "src/f32-vbinary/gen/vsubc-sse-x4.c",
556 "src/f32-vbinary/gen/vsubc-sse-x8.c",
557 "src/f32-vmulcaddc/gen/c4-sse-2x.c",
558 "src/f32-vmulcaddc/gen/c8-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700559 "src/x32-packx/x4-sse.c",
560]
561
562SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800563 "src/f32-argmaxpool/9p8x-sse2-c4.c",
564 "src/f32-argmaxpool/4x-sse2-c4.c",
565 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800566 "src/f32-prelu/gen/sse2-2x4.c",
567 "src/f32-prelu/gen/sse2-2x8.c",
568 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
569 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700570 "src/q8-avgpool/mp9p8q-sse2.c",
571 "src/q8-avgpool/up9-sse2.c",
572 "src/q8-igemm/4x4c2-sse2.c",
573 "src/q8-dwconv/up8x9-sse2.c",
574 "src/q8-gavgpool/mp7p7q-sse2.c",
575 "src/q8-gavgpool/up7-sse2.c",
576 "src/q8-gemm/2x4c8-sse2.c",
577 "src/q8-gemm/4x4c2-sse2.c",
578 "src/q8-vadd/sse2.c",
579 "src/u8-clamp/sse2.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800580 "src/u8-maxpool/9p8x-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700581 "src/u8-rmax/sse2.c",
582 "src/x32-pad/x2-sse2.c",
583 "src/x32-zip/x2-sse2.c",
584 "src/x32-zip/x3-sse2.c",
585 "src/x32-zip/x4-sse2.c",
586 "src/x32-zip/xm-sse2.c",
587 "src/x8-zip/x2-sse2.c",
588 "src/x8-zip/x3-sse2.c",
589 "src/x8-zip/x4-sse2.c",
590 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -0800591 "src/math/exp-sse2-p5.c",
592 "src/math/expminus-sse2-p5.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -0800593 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700594]
595
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800596SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800597 "src/f32-prelu/gen/sse41-2x4.c",
598 "src/f32-prelu/gen/sse41-2x8.c",
599 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
600 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800601]
602
Marat Dukhan08c4a432019-10-03 09:29:21 -0700603AVX_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800604 "src/f32-dwconv/gen/up16x4-avx-acc2.c",
605 "src/f32-dwconv/gen/up16x4-avx.c",
606 "src/f32-dwconv/gen/up8x4-avx-acc2.c",
607 "src/f32-dwconv/gen/up8x4-avx.c",
608 "src/f32-dwconv/gen/up16x9-avx-acc2.c",
609 "src/f32-dwconv/gen/up16x9-avx.c",
610 "src/f32-dwconv/gen/up8x9-avx-acc2.c",
611 "src/f32-dwconv/gen/up8x9-avx.c",
612 "src/f32-dwconv/gen/up16x25-avx-acc2.c",
613 "src/f32-dwconv/gen/up16x25-avx.c",
614 "src/f32-dwconv/gen/up8x25-avx-acc2.c",
615 "src/f32-dwconv/gen/up8x25-avx.c",
616 "src/f32-gemm/gen/1x8-avx-broadcast.c",
617 "src/f32-gemm/gen/4x8-avx-broadcast.c",
618 "src/f32-gemm/gen/5x8-avx-broadcast.c",
619 "src/f32-gemm/gen/6x8-avx-broadcast.c",
620 "src/f32-gemm/gen/7x8-avx-broadcast.c",
621 "src/f32-gemm/gen-inc/1x8-avx-broadcast.c",
622 "src/f32-gemm/gen-inc/4x8-avx-broadcast.c",
623 "src/f32-gemm/gen-inc/5x8-avx-broadcast.c",
624 "src/f32-gemm/gen-inc/6x8-avx-broadcast.c",
625 "src/f32-gemm/gen-inc/7x8-avx-broadcast.c",
626 "src/f32-igemm/gen/1x8-avx-broadcast.c",
627 "src/f32-igemm/gen/4x8-avx-broadcast.c",
628 "src/f32-igemm/gen/5x8-avx-broadcast.c",
629 "src/f32-igemm/gen/6x8-avx-broadcast.c",
630 "src/f32-igemm/gen/7x8-avx-broadcast.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700631 "src/f32-rmax/avx.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700632 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700633]
634
Marat Dukhanfda12b82019-11-21 12:27:59 -0800635FMA3_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800636 "src/f32-dwconv/gen/up16x4-fma3-acc2.c",
637 "src/f32-dwconv/gen/up16x4-fma3.c",
638 "src/f32-dwconv/gen/up8x4-fma3-acc2.c",
639 "src/f32-dwconv/gen/up8x4-fma3.c",
640 "src/f32-dwconv/gen/up16x9-fma3-acc2.c",
641 "src/f32-dwconv/gen/up16x9-fma3.c",
642 "src/f32-dwconv/gen/up8x9-fma3-acc2.c",
643 "src/f32-dwconv/gen/up8x9-fma3.c",
644 "src/f32-dwconv/gen/up16x25-fma3-acc2.c",
645 "src/f32-dwconv/gen/up16x25-fma3.c",
646 "src/f32-dwconv/gen/up8x25-fma3-acc2.c",
647 "src/f32-dwconv/gen/up8x25-fma3.c",
648 "src/f32-gemm/gen/1x8-fma3-broadcast.c",
649 "src/f32-gemm/gen/4x8-fma3-broadcast.c",
650 "src/f32-gemm/gen/5x8-fma3-broadcast.c",
651 "src/f32-gemm/gen/6x8-fma3-broadcast.c",
652 "src/f32-gemm/gen/7x8-fma3-broadcast.c",
653 "src/f32-gemm/gen/8x8-fma3-broadcast.c",
654 "src/f32-gemm/gen-inc/1x8-fma3-broadcast.c",
655 "src/f32-gemm/gen-inc/4x8-fma3-broadcast.c",
656 "src/f32-gemm/gen-inc/5x8-fma3-broadcast.c",
657 "src/f32-gemm/gen-inc/6x8-fma3-broadcast.c",
658 "src/f32-gemm/gen-inc/7x8-fma3-broadcast.c",
659 "src/f32-gemm/gen-inc/8x8-fma3-broadcast.c",
660 "src/f32-igemm/gen/1x8-fma3-broadcast.c",
661 "src/f32-igemm/gen/4x8-fma3-broadcast.c",
662 "src/f32-igemm/gen/5x8-fma3-broadcast.c",
663 "src/f32-igemm/gen/6x8-fma3-broadcast.c",
664 "src/f32-igemm/gen/7x8-fma3-broadcast.c",
665 "src/f32-igemm/gen/8x8-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -0800666]
667
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700668AVX2_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700669 "src/f32-raddexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700670 "src/f32-raddextexp/avx2-p5-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700671 "src/f32-raddstoreexpminusmax/avx2-p5-unroll64.c",
672 "src/f32-vscaleexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700673 "src/f32-vscaleextexp/avx2-p5-unroll64.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700674 "src/math/exp-avx2-p5.c",
675 "src/math/exp-avx2-perm-p3.c",
676 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -0700677 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700678 "src/math/extexp-avx2-p5.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700679]
680
Marat Dukhan08c4a432019-10-03 09:29:21 -0700681AVX512F_UKERNELS = [
Marat Dukhan479f87e2019-11-27 15:17:06 -0800682 "src/f32-dwconv/gen/up32x4-avx512f-acc2.c",
683 "src/f32-dwconv/gen/up32x4-avx512f.c",
684 "src/f32-dwconv/gen/up16x4-avx512f-acc2.c",
685 "src/f32-dwconv/gen/up16x4-avx512f.c",
686 "src/f32-dwconv/gen/up32x9-avx512f-acc2.c",
687 "src/f32-dwconv/gen/up32x9-avx512f.c",
688 "src/f32-dwconv/gen/up16x9-avx512f-acc2.c",
689 "src/f32-dwconv/gen/up16x9-avx512f.c",
690 "src/f32-dwconv/gen/up32x25-avx512f-acc2.c",
691 "src/f32-dwconv/gen/up32x25-avx512f.c",
692 "src/f32-dwconv/gen/up16x25-avx512f-acc2.c",
693 "src/f32-dwconv/gen/up16x25-avx512f.c",
Marat Dukhan0f349c42019-11-27 11:58:54 -0800694 "src/f32-gemm/gen/1x16-avx512f-broadcast.c",
695 "src/f32-gemm/gen/4x16-avx512f-broadcast.c",
696 "src/f32-gemm/gen/5x16-avx512f-broadcast.c",
697 "src/f32-gemm/gen/6x16-avx512f-broadcast.c",
698 "src/f32-gemm/gen/7x16-avx512f-broadcast.c",
699 "src/f32-gemm/gen/8x16-avx512f-broadcast.c",
700 "src/f32-gemm/gen-inc/1x16-avx512f-broadcast.c",
701 "src/f32-gemm/gen-inc/4x16-avx512f-broadcast.c",
702 "src/f32-gemm/gen-inc/5x16-avx512f-broadcast.c",
703 "src/f32-gemm/gen-inc/6x16-avx512f-broadcast.c",
704 "src/f32-gemm/gen-inc/7x16-avx512f-broadcast.c",
705 "src/f32-gemm/gen-inc/8x16-avx512f-broadcast.c",
706 "src/f32-igemm/gen/1x16-avx512f-broadcast.c",
707 "src/f32-igemm/gen/4x16-avx512f-broadcast.c",
708 "src/f32-igemm/gen/5x16-avx512f-broadcast.c",
709 "src/f32-igemm/gen/6x16-avx512f-broadcast.c",
710 "src/f32-igemm/gen/7x16-avx512f-broadcast.c",
711 "src/f32-igemm/gen/8x16-avx512f-broadcast.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700712 "src/f32-raddexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700713 "src/f32-raddextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700714 "src/f32-raddstoreexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700715 "src/f32-rmax/avx512f.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700716 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700717 "src/f32-vscaleexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700718 "src/f32-vscaleextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700719 "src/math/exp-avx512f-p5-scalef.c",
720 "src/math/exp-avx512f-p5.c",
721 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -0700722 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700723 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700724]
725
726AARCH32_ASM_UKERNELS = [
727 "src/q8-dwconv/up8x9-aarch32-neon.S",
728]
729
730AARCH64_ASM_UKERNELS = [
731 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
732 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800733 "src/f32-gemm/gen/1x12-aarch64-neonfma-cortex-a53.S",
734 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a53.S",
735 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
736 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
737 "src/f32-gemm/gen/4x12-aarch64-neonfma-cortex-a53.S",
738 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a53.S",
739 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
740 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
741 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld128.S",
742 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld64.S",
743 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
744 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a53.S",
745 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
746 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a73.S",
747 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
748 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld128.S",
749 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld64.S",
750 "src/f32-gemm/gen-inc/1x12-aarch64-neonfma-cortex-a53.S",
751 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a53.S",
752 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a57.S",
753 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a75.S",
754 "src/f32-gemm/gen-inc/4x12-aarch64-neonfma-cortex-a53.S",
755 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a53.S",
756 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a57.S",
757 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a75.S",
758 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld128.S",
759 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld64.S",
760 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a75.S",
761 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a53.S",
762 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a57.S",
763 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a73.S",
764 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a75.S",
765 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld128.S",
766 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700767 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700768 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700769 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S",
770 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S",
771 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700772 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700773 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
774 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700775 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700776 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
777 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
778 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
779]
780
781INTERNAL_MICROKERNEL_HDRS = [
782 "src/xnnpack/argmaxpool.h",
783 "src/xnnpack/avgpool.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800784 "src/xnnpack/bilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700785 "src/xnnpack/clamp.h",
786 "src/xnnpack/common.h",
787 "src/xnnpack/conv.h",
788 "src/xnnpack/dwconv.h",
789 "src/xnnpack/gavgpool.h",
790 "src/xnnpack/gemm.h",
791 "src/xnnpack/hswish.h",
792 "src/xnnpack/igemm.h",
793 "src/xnnpack/lut.h",
794 "src/xnnpack/math.h",
795 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -0800796 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700797 "src/xnnpack/packx.h",
798 "src/xnnpack/pad.h",
799 "src/xnnpack/params.h",
800 "src/xnnpack/pavgpool.h",
801 "src/xnnpack/ppmm.h",
802 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700803 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700804 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700805 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700806 "src/xnnpack/rmax.h",
807 "src/xnnpack/scalar-utils.h",
808 "src/xnnpack/spmm.h",
809 "src/xnnpack/unpool.h",
810 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800811 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700812 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700813 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700814 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700815 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800816 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700817 "src/xnnpack/zip.h",
818]
819
820INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
821 "include/xnnpack.h",
822 "src/xnnpack/allocator.h",
823 "src/xnnpack/compute.h",
824 "src/xnnpack/im2col.h",
825 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700826 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700827 "src/xnnpack/operator.h",
828 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700829 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700830 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700831 "src/xnnpack/requantization.h",
832]
833
834ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
835 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700836]
837
838MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700839 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700840 "include/xnnpack.h",
841]
842
843MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
844 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700845 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700846 "src/xnnpack/requantization.h",
847 "include/xnnpack.h",
848]
849
850OPERATOR_TEST_PARAMS_HDRS = [
851 "src/xnnpack/params.h",
852 "src/xnnpack/common.h",
853]
854
855WEIGHTS_PACK_HDRS = [
856 "src/xnnpack/pack.h",
857 "src/xnnpack/operator.h",
858 "src/xnnpack/compute.h",
859]
860
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700861LOGGING_COPTS = select({
862 # No logging in optimized mode
863 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
864 # Full logging in debug mode
865 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
866 # Error-only logging in default (fastbuild) mode
867 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
868})
869
870LOGGING_HDRS = [
871 "src/xnnpack/log.h",
872]
873
Marat Dukhan08c4a432019-10-03 09:29:21 -0700874xnnpack_cc_library(
875 name = "scalar_ukernels",
876 srcs = SCALAR_UKERNELS,
877 hdrs = INTERNAL_HDRS,
878 aarch32_copts = ["-marm"],
879 copts = xnnpack_std_copts(),
880 deps = [
881 "@FP16",
882 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -0800883 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700884 ],
885)
886
887xnnpack_cc_library(
888 name = "psimd_ukernels",
889 srcs = PSIMD_UKERNELS,
890 hdrs = INTERNAL_HDRS,
891 aarch32_copts = [
892 "-marm",
893 "-mfpu=neon",
894 ],
895 copts = xnnpack_std_copts(),
896 optimized_copts = [
897 "-O3",
898 "-ffast-math",
899 ],
900 deps = [
901 "@FP16",
902 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -0800903 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700904 ],
905)
906
907xnnpack_cc_library(
908 name = "neon_ukernels",
909 hdrs = INTERNAL_HDRS,
910 aarch32_copts = [
911 "-marm",
912 "-mfpu=neon",
913 ],
914 aarch32_srcs = NEON_UKERNELS,
915 aarch64_srcs = NEON_UKERNELS,
916 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -0800917 deps = [
918 "@FP16",
919 "@pthreadpool",
920 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700921)
922
923xnnpack_cc_library(
924 name = "neonfma_ukernels",
925 hdrs = INTERNAL_HDRS,
926 aarch32_copts = [
927 "-marm",
928 "-mfpu=neon-vfpv4",
929 ],
930 aarch32_srcs = NEONFMA_UKERNELS,
931 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
932 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -0800933 deps = [
934 "@FP16",
935 "@pthreadpool",
936 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700937)
938
939xnnpack_cc_library(
940 name = "neonfp16arith_ukernels",
941 hdrs = INTERNAL_HDRS,
942 aarch64_copts = ["-march=armv8.2-a+fp16"],
943 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
944 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -0800945 deps = [
946 "@FP16",
947 "@pthreadpool",
948 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700949)
950
951xnnpack_cc_library(
952 name = "sse2_ukernels",
953 hdrs = INTERNAL_HDRS,
954 copts = xnnpack_std_copts(),
955 x86_copts = ["-msse2"],
956 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -0800957 deps = [
958 "@FP16",
959 "@pthreadpool",
960 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700961)
962
963xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800964 name = "sse41_ukernels",
965 hdrs = INTERNAL_HDRS,
966 copts = xnnpack_std_copts(),
967 x86_copts = ["-msse4.1"],
968 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -0800969 deps = [
970 "@FP16",
971 "@pthreadpool",
972 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800973)
974
975xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700976 name = "avx_ukernels",
977 hdrs = INTERNAL_HDRS,
978 copts = xnnpack_std_copts(),
979 x86_copts = ["-mavx"],
980 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -0800981 deps = [
982 "@FP16",
983 "@pthreadpool",
984 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700985)
986
987xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -0800988 name = "fma3_ukernels",
989 hdrs = INTERNAL_HDRS,
990 copts = xnnpack_std_copts(),
991 x86_copts = [
992 "-mfma",
993 ],
994 x86_srcs = FMA3_UKERNELS,
995 deps = [
996 "@FP16",
997 "@pthreadpool",
998 ],
999)
1000
1001xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001002 name = "avx2_ukernels",
1003 hdrs = INTERNAL_HDRS,
1004 copts = xnnpack_std_copts(),
1005 x86_copts = [
1006 "-mfma",
1007 "-mavx2",
1008 ],
1009 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001010 deps = [
1011 "@FP16",
1012 "@pthreadpool",
1013 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001014)
1015
1016xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001017 name = "avx512f_ukernels",
1018 hdrs = INTERNAL_HDRS,
1019 copts = xnnpack_std_copts(),
1020 x86_copts = ["-mavx512f"],
1021 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001022 deps = [
1023 "@FP16",
1024 "@pthreadpool",
1025 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001026)
1027
1028xnnpack_cc_library(
1029 name = "asm_ukernels",
1030 hdrs = ["src/xnnpack/assembly.h"],
1031 aarch32_srcs = AARCH32_ASM_UKERNELS,
1032 aarch64_srcs = AARCH64_ASM_UKERNELS,
1033)
1034
1035xnnpack_aggregate_library(
1036 name = "ukernels",
1037 aarch32_deps = [
1038 ":psimd_ukernels",
1039 ":neon_ukernels",
1040 ":neonfma_ukernels",
1041 ":asm_ukernels",
1042 ],
1043 aarch64_deps = [
1044 ":psimd_ukernels",
1045 ":neon_ukernels",
1046 ":neonfma_ukernels",
1047 ":neonfp16arith_ukernels",
1048 ":asm_ukernels",
1049 ],
1050 generic_deps = [":scalar_ukernels"],
1051 wasmsimd_deps = [
1052 ":psimd_ukernels",
1053 ],
1054 x86_deps = [
1055 ":psimd_ukernels",
1056 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001057 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001058 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001059 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001060 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001061 ":avx512f_ukernels",
1062 ],
1063)
1064
1065xnnpack_cc_library(
1066 name = "im2col",
1067 srcs = ["src/im2col.c"],
1068 hdrs = [
1069 "src/xnnpack/common.h",
1070 "src/xnnpack/im2col.h",
1071 ],
1072 copts = xnnpack_std_copts(),
1073)
1074
1075xnnpack_cc_library(
1076 name = "indirection",
1077 srcs = ["src/indirection.c"],
1078 hdrs = INTERNAL_HDRS,
1079 copts = xnnpack_std_copts(),
1080 deps = [
1081 "@FP16",
1082 "@FXdiv",
1083 "@pthreadpool",
1084 ],
1085)
1086
1087xnnpack_cc_library(
1088 name = "operator_run",
1089 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001090 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
1091 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001092 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
1093 "-Wno-vla",
1094 ],
1095 deps = [
1096 "@FP16",
1097 "@FXdiv",
1098 "@clog",
1099 "@pthreadpool",
1100 ],
1101)
1102
1103cc_library(
1104 name = "enable_assembly",
1105 defines = select({
1106 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1107 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001108 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001109 }),
1110)
1111
Marat Dukhancf056b22019-10-07 10:26:29 -07001112xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001113 name = "operators",
1114 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001115 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001116 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07001117 ],
1118 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001119 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001120 "-Isrc",
1121 "-Iinclude",
1122 ] + select({
1123 ":debug_build": [],
1124 "//conditions:default": xnnpack_min_size_copts(),
1125 }),
Marat Dukhancf056b22019-10-07 10:26:29 -07001126 wasm_srcs = ["src/wasm-stubs.c"],
1127 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001128 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001129 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001130 "@FP16",
1131 "@FXdiv",
1132 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001133 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001134 ],
1135)
1136
1137cc_library(
1138 name = "XNNPACK",
1139 srcs = [
1140 "src/init.c",
1141 ],
1142 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1143 "-Isrc",
1144 "-Iinclude",
1145 ] + select({
1146 ":debug_build": [],
1147 "//conditions:default": xnnpack_min_size_copts(),
1148 }),
1149 includes = ["include"],
1150 linkstatic = True,
1151 textual_hdrs = ["include/xnnpack.h"],
1152 visibility = xnnpack_visibility(),
1153 deps = [
1154 ":enable_assembly",
1155 ":ukernels",
1156 ":operator_run",
1157 ":operators",
1158 "@clog",
1159 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07001160 ] + select({
1161 ":emscripten": [],
1162 "//conditions:default": ["@cpuinfo"],
1163 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001164)
1165
1166cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001167 name = "xnnpack_operators_nhwc_f32",
1168 srcs = [
1169 "src/init.c",
1170 ],
1171 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1172 "-Isrc",
1173 "-Iinclude",
1174 ] + select({
1175 ":debug_build": [],
1176 "//conditions:default": xnnpack_min_size_copts(),
1177 }),
1178 defines = [
1179 "XNN_NO_Q8_OPERATORS",
1180 "XNN_NO_U8_OPERATORS",
1181 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08001182 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001183 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001184 includes = ["include"],
1185 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001186 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001187 visibility = xnnpack_visibility(),
1188 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001189 ":enable_assembly",
1190 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001191 ":operator_run",
1192 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001193 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001194 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001195 ] + select({
1196 ":emscripten": [],
1197 "//conditions:default": ["@cpuinfo"],
1198 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001199)
1200
Marat Dukhancf056b22019-10-07 10:26:29 -07001201xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001202 name = "bench_utils",
1203 srcs = ["bench/utils.cc"],
1204 hdrs = ["bench/utils.h"],
1205 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08001206 deps = [
1207 "@com_google_benchmark//:benchmark",
1208 "@cpuinfo",
1209 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001210)
1211
Frank Barchard7e955972019-10-11 10:34:25 -07001212######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07001213
1214xnnpack_benchmark(
1215 name = "q8_gemm_bench",
1216 srcs = [
1217 "bench/gemm.h",
1218 "bench/q8-gemm.cc",
1219 "src/xnnpack/AlignedAllocator.h",
1220 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1221 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
1222 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
1223)
1224
1225xnnpack_benchmark(
1226 name = "f16_gemm_bench",
1227 srcs = [
1228 "bench/f16-gemm.cc",
1229 "bench/gemm.h",
1230 "src/xnnpack/AlignedAllocator.h",
1231 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1232 copts = ["-Wno-unused-function"],
1233 deps = MICROKERNEL_BENCHMARK_DEPS,
1234)
1235
1236xnnpack_benchmark(
1237 name = "f32_igemm_bench",
1238 srcs = [
1239 "bench/f32-igemm.cc",
1240 "bench/conv.h",
1241 "src/xnnpack/AlignedAllocator.h",
1242 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07001243 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001244)
1245
1246xnnpack_benchmark(
1247 name = "f32_conv_hwc_bench",
1248 srcs = [
1249 "bench/f32-conv-hwc.cc",
1250 "bench/dconv.h",
1251 "src/xnnpack/AlignedAllocator.h",
1252 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1253 copts = ["-Wno-unused-function"],
1254 deps = MICROKERNEL_BENCHMARK_DEPS,
1255)
1256
1257xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07001258 name = "f32_conv_hwc2spchw_bench",
1259 srcs = [
1260 "bench/f32-conv-hwc2spchw.cc",
1261 "bench/dconv.h",
1262 "src/xnnpack/AlignedAllocator.h",
1263 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1264 copts = ["-Wno-unused-function"],
1265 deps = MICROKERNEL_BENCHMARK_DEPS,
1266)
1267
1268xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001269 name = "f32_dwconv_bench",
1270 srcs = [
1271 "bench/f32-dwconv.cc",
1272 "bench/dwconv.h",
1273 "src/xnnpack/AlignedAllocator.h",
1274 ] + MICROKERNEL_BENCHMARK_HDRS,
1275 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1276)
1277
1278xnnpack_benchmark(
1279 name = "f32_dwconv_spchw_bench",
1280 srcs = [
1281 "bench/f32-dwconv-spchw.cc",
1282 "bench/dwconv.h",
1283 "src/xnnpack/AlignedAllocator.h",
1284 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1285 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1286)
1287
1288xnnpack_benchmark(
1289 name = "f32_gemm_bench",
1290 srcs = [
1291 "bench/f32-gemm.cc",
1292 "bench/gemm.h",
1293 "src/xnnpack/AlignedAllocator.h",
1294 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1295 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07001296 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001297)
1298
1299xnnpack_benchmark(
1300 name = "f32_rmax_bench",
1301 srcs = [
1302 "bench/f32-rmax.cc",
1303 "src/xnnpack/AlignedAllocator.h",
1304 ] + MICROKERNEL_BENCHMARK_HDRS,
1305 deps = MICROKERNEL_BENCHMARK_DEPS,
1306)
1307
1308xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08001309 name = "f32_sigmoid_bench",
1310 srcs = [
1311 "bench/f32-sigmoid.cc",
1312 "src/xnnpack/AlignedAllocator.h",
1313 ] + MICROKERNEL_BENCHMARK_HDRS,
1314 copts = ["-Wno-unused-function"],
1315 deps = MICROKERNEL_BENCHMARK_DEPS,
1316)
1317
1318xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001319 name = "f32_spmm_bench",
1320 srcs = [
1321 "bench/f32-spmm.cc",
1322 "bench/gemm.h",
1323 "src/xnnpack/AlignedAllocator.h",
1324 ] + MICROKERNEL_BENCHMARK_HDRS,
1325 copts = ["-Wno-unused-function"],
1326 deps = MICROKERNEL_BENCHMARK_DEPS,
1327)
1328
1329xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07001330 name = "f32_softargmax_bench",
1331 srcs = [
1332 "bench/f32-softargmax.cc",
1333 ] + MICROKERNEL_BENCHMARK_HDRS,
1334 copts = ["-Wno-unused-function"],
1335 deps = MICROKERNEL_BENCHMARK_DEPS,
1336)
1337
1338xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001339 name = "f32_im2col_gemm_bench",
1340 srcs = [
1341 "bench/f32-im2col-gemm.cc",
1342 "bench/conv.h",
1343 "src/xnnpack/AlignedAllocator.h",
1344 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1345 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
1346)
1347
1348########################### Benchmarks for operators ###########################
1349
1350xnnpack_benchmark(
1351 name = "add_bench",
1352 srcs = ["bench/add.cc"],
1353 deps = OPERATOR_BENCHMARK_DEPS,
1354)
1355
1356xnnpack_benchmark(
1357 name = "average_pooling_bench",
1358 srcs = ["bench/average-pooling.cc"],
1359 deps = OPERATOR_BENCHMARK_DEPS,
1360)
1361
1362xnnpack_benchmark(
1363 name = "channel_shuffle_bench",
1364 srcs = ["bench/channel-shuffle.cc"],
1365 deps = OPERATOR_BENCHMARK_DEPS,
1366)
1367
1368xnnpack_benchmark(
1369 name = "convolution_bench",
1370 srcs = ["bench/convolution.cc"],
1371 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1372 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1373)
1374
1375xnnpack_benchmark(
1376 name = "deconvolution_bench",
1377 srcs = ["bench/deconvolution.cc"],
1378 copts = xnnpack_optional_tflite_copts(),
1379 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1380)
1381
1382xnnpack_benchmark(
1383 name = "global_average_pooling_bench",
1384 srcs = ["bench/global-average-pooling.cc"],
1385 deps = OPERATOR_BENCHMARK_DEPS,
1386)
1387
1388xnnpack_benchmark(
1389 name = "max_pooling_bench",
1390 srcs = ["bench/max-pooling.cc"],
1391 deps = OPERATOR_BENCHMARK_DEPS,
1392)
1393
1394xnnpack_benchmark(
1395 name = "sigmoid_bench",
1396 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08001397 copts = xnnpack_optional_tflite_copts(),
1398 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001399)
1400
1401xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07001402 name = "prelu_bench",
1403 srcs = ["bench/prelu.cc"],
1404 copts = xnnpack_optional_tflite_copts(),
1405 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1406)
1407
1408xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001409 name = "softargmax_bench",
1410 srcs = ["bench/softargmax.cc"],
1411 deps = OPERATOR_BENCHMARK_DEPS,
1412)
1413
Marat Dukhanc068bb62019-10-04 13:24:39 -07001414############################# End-to-end benchmarks ############################
1415
1416cc_library(
1417 name = "mobilenet_v1",
1418 srcs = ["models/mobilenet-v1.cc"],
1419 hdrs = ["models/models.h"],
1420 linkstatic = True,
1421 deps = [
1422 ":XNNPACK",
1423 "@pthreadpool",
1424 ],
1425)
1426
1427cc_library(
1428 name = "mobilenet_v2",
1429 srcs = ["models/mobilenet-v2.cc"],
1430 hdrs = ["models/models.h"],
1431 linkstatic = True,
1432 deps = [
1433 ":XNNPACK",
1434 "@pthreadpool",
1435 ],
1436)
1437
1438xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07001439 name = "f32_dwconv_e2e_bench",
1440 srcs = ["bench/f32-dwconv-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1441 copts = ["-Wno-unused-function"],
1442 deps = MICROKERNEL_BENCHMARK_DEPS + [
1443 ":XNNPACK",
1444 ":mobilenet_v1",
1445 ":mobilenet_v2",
1446 ],
1447)
1448
1449xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07001450 name = "f32_gemm_e2e_bench",
1451 srcs = ["bench/f32-gemm-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1452 copts = ["-Wno-unused-function"],
1453 deps = MICROKERNEL_BENCHMARK_DEPS + [
1454 ":XNNPACK",
1455 ":mobilenet_v1",
1456 ":mobilenet_v2",
1457 ],
1458)
1459
1460xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07001461 name = "end2end_bench",
1462 srcs = ["bench/end2end.cc"],
1463 deps = [
1464 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07001465 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07001466 ":mobilenet_v1",
1467 ":mobilenet_v2",
1468 "@pthreadpool",
1469 ],
1470)
1471
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001472#################### Accuracy evaluation for math functions ####################
1473
1474xnnpack_benchmark(
1475 name = "f32_exp_eval",
1476 srcs = [
1477 "eval/f32-exp.cc",
1478 "src/xnnpack/AlignedAllocator.h",
1479 ] + ACCURACY_EVAL_HDRS,
1480 deps = ACCURACY_EVAL_DEPS,
1481)
1482
Marat Dukhan515c9772019-10-17 18:07:57 -07001483xnnpack_benchmark(
1484 name = "f32_expminus_eval",
1485 srcs = [
1486 "eval/f32-expminus.cc",
1487 "src/xnnpack/AlignedAllocator.h",
1488 ] + ACCURACY_EVAL_HDRS,
1489 deps = ACCURACY_EVAL_DEPS,
1490)
1491
Marat Dukhan98ba4412019-10-23 02:14:28 -07001492xnnpack_benchmark(
1493 name = "f32_extexp_eval",
1494 srcs = [
1495 "eval/f32-extexp.cc",
1496 "src/xnnpack/AlignedAllocator.h",
1497 ] + ACCURACY_EVAL_HDRS,
1498 deps = ACCURACY_EVAL_DEPS,
1499)
1500
Marat Dukhan346a9e52019-11-15 09:06:30 -08001501xnnpack_benchmark(
1502 name = "f32_sigmoid_eval",
1503 srcs = [
1504 "eval/f32-sigmoid.cc",
1505 "src/xnnpack/AlignedAllocator.h",
1506 ] + ACCURACY_EVAL_HDRS,
1507 deps = ACCURACY_EVAL_DEPS,
1508)
1509
Marat Dukhan08c4a432019-10-03 09:29:21 -07001510######################### Unit tests for micro-kernels #########################
1511
1512xnnpack_unit_test(
1513 name = "f16_gemm_test",
1514 srcs = [
1515 "test/f16-gemm.cc",
1516 "test/gemm-microkernel-tester.h",
1517 "src/xnnpack/AlignedAllocator.h",
1518 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1519 deps = MICROKERNEL_TEST_DEPS,
1520)
1521
1522xnnpack_unit_test(
1523 name = "f32_argmaxpool_test",
1524 srcs = [
1525 "test/f32-argmaxpool.cc",
1526 "test/argmaxpool-microkernel-tester.h",
1527 "src/xnnpack/AlignedAllocator.h",
1528 ] + MICROKERNEL_TEST_HDRS,
1529 deps = MICROKERNEL_TEST_DEPS,
1530)
1531
1532xnnpack_unit_test(
1533 name = "f32_avgpool_test",
1534 srcs = [
1535 "test/f32-avgpool.cc",
1536 "test/avgpool-microkernel-tester.h",
1537 "src/xnnpack/AlignedAllocator.h",
1538 ] + MICROKERNEL_TEST_HDRS,
1539 deps = MICROKERNEL_TEST_DEPS,
1540)
1541
1542xnnpack_unit_test(
Marat Dukhan35dacfb2019-11-07 19:18:16 -08001543 name = "f32_bilinear_test",
1544 srcs = [
1545 "test/f32-bilinear.cc",
1546 "test/bilinear-microkernel-tester.h",
1547 "src/xnnpack/AlignedAllocator.h",
1548 ] + MICROKERNEL_TEST_HDRS,
1549 deps = MICROKERNEL_TEST_DEPS,
1550)
1551
1552xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001553 name = "f32_clamp_test",
1554 srcs = [
1555 "test/f32-clamp.cc",
1556 "test/clamp-microkernel-tester.h",
1557 ] + MICROKERNEL_TEST_HDRS,
1558 deps = MICROKERNEL_TEST_DEPS,
1559)
1560
1561xnnpack_unit_test(
1562 name = "f32_igemm_test",
1563 srcs = [
1564 "test/f32-igemm.cc",
1565 "test/gemm-microkernel-tester.h",
1566 "src/xnnpack/AlignedAllocator.h",
1567 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1568 deps = MICROKERNEL_TEST_DEPS,
1569)
1570
1571xnnpack_unit_test(
1572 name = "f32_conv_hwc_test",
1573 srcs = [
1574 "test/f32-conv-hwc.cc",
1575 "test/conv-hwc-microkernel-tester.h",
1576 "src/xnnpack/AlignedAllocator.h",
1577 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1578 deps = MICROKERNEL_TEST_DEPS,
1579)
1580
1581xnnpack_unit_test(
1582 name = "f32_conv_hwc2spchw_test",
1583 srcs = [
1584 "test/f32-conv-hwc2spchw.cc",
1585 "test/conv-hwc2spchw-microkernel-tester.h",
1586 "src/xnnpack/AlignedAllocator.h",
1587 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1588 deps = MICROKERNEL_TEST_DEPS,
1589)
1590
1591xnnpack_unit_test(
1592 name = "f32_dwconv_test",
1593 srcs = [
1594 "test/f32-dwconv.cc",
1595 "test/dwconv-microkernel-tester.h",
1596 "src/xnnpack/AlignedAllocator.h",
1597 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1598 deps = MICROKERNEL_TEST_DEPS,
1599)
1600
1601xnnpack_unit_test(
1602 name = "f32_dwconv_spchw_test",
1603 srcs = [
1604 "test/f32-dwconv-spchw.cc",
1605 "test/dwconv-spchw-microkernel-tester.h",
1606 "src/xnnpack/AlignedAllocator.h",
1607 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1608 deps = MICROKERNEL_TEST_DEPS,
1609)
1610
1611xnnpack_unit_test(
1612 name = "f32_gavgpool_test",
1613 srcs = [
1614 "test/f32-gavgpool.cc",
1615 "test/gavgpool-microkernel-tester.h",
1616 "src/xnnpack/AlignedAllocator.h",
1617 ] + MICROKERNEL_TEST_HDRS,
1618 deps = MICROKERNEL_TEST_DEPS,
1619)
1620
1621xnnpack_unit_test(
1622 name = "f32_gavgpool_spchw_test",
1623 srcs = [
1624 "test/f32-gavgpool-spchw.cc",
1625 "test/gavgpool-spchw-microkernel-tester.h",
1626 "src/xnnpack/AlignedAllocator.h",
1627 ] + MICROKERNEL_TEST_HDRS,
1628 deps = MICROKERNEL_TEST_DEPS,
1629)
1630
1631xnnpack_unit_test(
1632 name = "f32_gemm_test",
1633 srcs = [
1634 "test/f32-gemm.cc",
1635 "test/gemm-microkernel-tester.h",
1636 "src/xnnpack/AlignedAllocator.h",
1637 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1638 deps = MICROKERNEL_TEST_DEPS,
1639)
1640
1641xnnpack_unit_test(
1642 name = "f32_gemminc_test",
1643 srcs = [
1644 "test/f32-gemminc.cc",
1645 "test/gemm-microkernel-tester.h",
1646 "src/xnnpack/AlignedAllocator.h",
1647 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1648 deps = MICROKERNEL_TEST_DEPS,
1649)
1650
1651xnnpack_unit_test(
1652 name = "f32_hswish_test",
1653 srcs = [
1654 "test/f32-hswish.cc",
1655 "test/hswish-microkernel-tester.h",
1656 ] + MICROKERNEL_TEST_HDRS,
1657 deps = MICROKERNEL_TEST_DEPS,
1658)
1659
1660xnnpack_unit_test(
1661 name = "f32_maxpool_test",
1662 srcs = [
1663 "test/f32-maxpool.cc",
1664 "test/maxpool-microkernel-tester.h",
1665 ] + MICROKERNEL_TEST_HDRS,
1666 deps = MICROKERNEL_TEST_DEPS,
1667)
1668
1669xnnpack_unit_test(
1670 name = "f32_pavgpool_test",
1671 srcs = [
1672 "test/f32-pavgpool.cc",
1673 "test/avgpool-microkernel-tester.h",
1674 "src/xnnpack/AlignedAllocator.h",
1675 ] + MICROKERNEL_TEST_HDRS,
1676 deps = MICROKERNEL_TEST_DEPS,
1677)
1678
1679xnnpack_unit_test(
1680 name = "f32_ppmm_test",
1681 srcs = [
1682 "test/f32-ppmm.cc",
1683 "test/gemm-microkernel-tester.h",
1684 "src/xnnpack/AlignedAllocator.h",
1685 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1686 deps = MICROKERNEL_TEST_DEPS,
1687)
1688
1689xnnpack_unit_test(
1690 name = "f32_prelu_test",
1691 srcs = [
1692 "test/f32-prelu.cc",
1693 "test/prelu-microkernel-tester.h",
1694 "src/xnnpack/AlignedAllocator.h",
1695 ] + MICROKERNEL_TEST_HDRS,
1696 deps = MICROKERNEL_TEST_DEPS,
1697)
1698
1699xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001700 name = "f32_raddexpminusmax_test",
1701 srcs = [
1702 "test/f32-raddexpminusmax.cc",
1703 "test/raddexpminusmax-microkernel-tester.h",
1704 ] + MICROKERNEL_TEST_HDRS,
1705 deps = MICROKERNEL_TEST_DEPS,
1706)
1707
1708xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001709 name = "f32_raddextexp_test",
1710 srcs = [
1711 "test/f32-raddextexp.cc",
1712 "test/raddextexp-microkernel-tester.h",
1713 ] + MICROKERNEL_TEST_HDRS,
1714 deps = MICROKERNEL_TEST_DEPS,
1715)
1716
1717xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001718 name = "f32_raddstoreexpminusmax_test",
1719 srcs = [
1720 "test/f32-raddstoreexpminusmax.cc",
1721 "test/raddstoreexpminusmax-microkernel-tester.h",
1722 ] + MICROKERNEL_TEST_HDRS,
1723 deps = MICROKERNEL_TEST_DEPS,
1724)
1725
1726xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001727 name = "f32_rmax_test",
1728 srcs = [
1729 "test/f32-rmax.cc",
1730 "test/rmax-microkernel-tester.h",
1731 ] + MICROKERNEL_TEST_HDRS,
1732 deps = MICROKERNEL_TEST_DEPS,
1733)
1734
1735xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08001736 name = "f32_sigmoid_test",
1737 srcs = [
1738 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001739 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001740 ] + MICROKERNEL_TEST_HDRS,
1741 deps = MICROKERNEL_TEST_DEPS,
1742)
1743
1744xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001745 name = "f32_spmm_test",
1746 srcs = [
1747 "test/f32-spmm.cc",
1748 "test/spmm-microkernel-tester.h",
1749 "src/xnnpack/AlignedAllocator.h",
1750 ] + MICROKERNEL_TEST_HDRS,
1751 deps = MICROKERNEL_TEST_DEPS,
1752)
1753
1754xnnpack_unit_test(
1755 name = "f32_vadd_test",
1756 srcs = [
1757 "test/f32-vadd.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001758 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001759 ] + MICROKERNEL_TEST_HDRS,
1760 deps = MICROKERNEL_TEST_DEPS,
1761)
1762
1763xnnpack_unit_test(
1764 name = "f32_vaddc_test",
1765 srcs = [
1766 "test/f32-vaddc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001767 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001768 ] + MICROKERNEL_TEST_HDRS,
1769 deps = MICROKERNEL_TEST_DEPS,
1770)
1771
1772xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001773 name = "f32_vmul_test",
1774 srcs = [
1775 "test/f32-vmul.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001776 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001777 ] + MICROKERNEL_TEST_HDRS,
1778 deps = MICROKERNEL_TEST_DEPS,
1779)
1780
1781xnnpack_unit_test(
1782 name = "f32_vmulc_test",
1783 srcs = [
1784 "test/f32-vmulc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001785 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001786 ] + MICROKERNEL_TEST_HDRS,
1787 deps = MICROKERNEL_TEST_DEPS,
1788)
1789
1790xnnpack_unit_test(
1791 name = "f32_vmulcaddc_test",
1792 srcs = [
1793 "test/f32-vmulcaddc.cc",
1794 "test/vmulcaddc-microkernel-tester.h",
1795 "src/xnnpack/AlignedAllocator.h",
1796 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1797 deps = MICROKERNEL_TEST_DEPS,
1798)
1799
1800xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001801 name = "f32_vscale_test",
1802 srcs = [
1803 "test/f32-vscale.cc",
1804 "test/vscale-microkernel-tester.h",
1805 ] + MICROKERNEL_TEST_HDRS,
1806 deps = MICROKERNEL_TEST_DEPS,
1807)
1808
1809xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001810 name = "f32_vscaleexpminusmax_test",
1811 srcs = [
1812 "test/f32-vscaleexpminusmax.cc",
1813 "test/vscaleexpminusmax-microkernel-tester.h",
1814 ] + MICROKERNEL_TEST_HDRS,
1815 deps = MICROKERNEL_TEST_DEPS,
1816)
1817
1818xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001819 name = "f32_vscaleextexp_test",
1820 srcs = [
1821 "test/f32-vscaleextexp.cc",
1822 "test/vscaleextexp-microkernel-tester.h",
1823 ] + MICROKERNEL_TEST_HDRS,
1824 deps = MICROKERNEL_TEST_DEPS,
1825)
1826
1827xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001828 name = "f32_vsub_test",
1829 srcs = [
1830 "test/f32-vsub.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001831 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001832 ] + MICROKERNEL_TEST_HDRS,
1833 deps = MICROKERNEL_TEST_DEPS,
1834)
1835
1836xnnpack_unit_test(
1837 name = "f32_vsubc_test",
1838 srcs = [
1839 "test/f32-vsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001840 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001841 ] + MICROKERNEL_TEST_HDRS,
1842 deps = MICROKERNEL_TEST_DEPS,
1843)
1844
1845xnnpack_unit_test(
1846 name = "f32_vrsubc_test",
1847 srcs = [
1848 "test/f32-vrsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001849 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001850 ] + MICROKERNEL_TEST_HDRS,
1851 deps = MICROKERNEL_TEST_DEPS,
1852)
1853
1854xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001855 name = "q8_avgpool_test",
1856 srcs = [
1857 "test/q8-avgpool.cc",
1858 "test/avgpool-microkernel-tester.h",
1859 "src/xnnpack/AlignedAllocator.h",
1860 ] + MICROKERNEL_TEST_HDRS,
1861 deps = MICROKERNEL_TEST_DEPS,
1862)
1863
1864xnnpack_unit_test(
1865 name = "q8_igemm_test",
1866 srcs = [
1867 "test/q8-igemm.cc",
1868 "test/gemm-microkernel-tester.h",
1869 "src/xnnpack/AlignedAllocator.h",
1870 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1871 deps = MICROKERNEL_TEST_DEPS,
1872)
1873
1874xnnpack_unit_test(
1875 name = "q8_dwconv_test",
1876 srcs = [
1877 "test/q8-dwconv.cc",
1878 "test/dwconv-microkernel-tester.h",
1879 "src/xnnpack/AlignedAllocator.h",
1880 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1881 deps = MICROKERNEL_TEST_DEPS,
1882)
1883
1884xnnpack_unit_test(
1885 name = "q8_gavgpool_test",
1886 srcs = [
1887 "test/q8-gavgpool.cc",
1888 "test/gavgpool-microkernel-tester.h",
1889 "src/xnnpack/AlignedAllocator.h",
1890 ] + MICROKERNEL_TEST_HDRS,
1891 deps = MICROKERNEL_TEST_DEPS,
1892)
1893
1894xnnpack_unit_test(
1895 name = "q8_gemm_test",
1896 srcs = [
1897 "test/q8-gemm.cc",
1898 "test/gemm-microkernel-tester.h",
1899 "src/xnnpack/AlignedAllocator.h",
1900 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1901 deps = MICROKERNEL_TEST_DEPS,
1902)
1903
1904xnnpack_unit_test(
1905 name = "q8_vadd_test",
1906 srcs = [
1907 "test/q8-vadd.cc",
1908 "test/vadd-microkernel-tester.h",
1909 ] + MICROKERNEL_TEST_HDRS,
1910 deps = MICROKERNEL_TEST_DEPS,
1911)
1912
1913xnnpack_unit_test(
1914 name = "u8_clamp_test",
1915 srcs = [
1916 "test/u8-clamp.cc",
1917 "test/clamp-microkernel-tester.h",
1918 ] + MICROKERNEL_TEST_HDRS,
1919 deps = MICROKERNEL_TEST_DEPS,
1920)
1921
1922xnnpack_unit_test(
1923 name = "u8_lut32norm_test",
1924 srcs = [
1925 "test/u8-lut32norm.cc",
1926 "test/lut-norm-microkernel-tester.h",
1927 ] + MICROKERNEL_TEST_HDRS,
1928 deps = MICROKERNEL_TEST_DEPS,
1929)
1930
1931xnnpack_unit_test(
1932 name = "u8_maxpool_test",
1933 srcs = [
1934 "test/u8-maxpool.cc",
1935 "test/maxpool-microkernel-tester.h",
1936 ] + MICROKERNEL_TEST_HDRS,
1937 deps = MICROKERNEL_TEST_DEPS,
1938)
1939
1940xnnpack_unit_test(
1941 name = "u8_rmax_test",
1942 srcs = [
1943 "test/u8-rmax.cc",
1944 "test/rmax-microkernel-tester.h",
1945 ] + MICROKERNEL_TEST_HDRS,
1946 deps = MICROKERNEL_TEST_DEPS,
1947)
1948
1949xnnpack_unit_test(
1950 name = "x32_packx_test",
1951 srcs = [
1952 "test/x32-packx.cc",
1953 "test/pack-microkernel-tester.h",
1954 "src/xnnpack/AlignedAllocator.h",
1955 ] + MICROKERNEL_TEST_HDRS,
1956 deps = MICROKERNEL_TEST_DEPS,
1957)
1958
1959xnnpack_unit_test(
1960 name = "x32_pad_test",
1961 srcs = [
1962 "test/x32-pad.cc",
1963 "test/pad-microkernel-tester.h",
1964 ] + MICROKERNEL_TEST_HDRS,
1965 deps = MICROKERNEL_TEST_DEPS,
1966)
1967
1968xnnpack_unit_test(
1969 name = "x32_unpool_test",
1970 srcs = [
1971 "test/x32-unpool.cc",
1972 "test/unpool-microkernel-tester.h",
1973 ] + MICROKERNEL_TEST_HDRS,
1974 deps = MICROKERNEL_TEST_DEPS,
1975)
1976
1977xnnpack_unit_test(
1978 name = "x32_zip_test",
1979 srcs = [
1980 "test/x32-zip.cc",
1981 "test/zip-microkernel-tester.h",
1982 ] + MICROKERNEL_TEST_HDRS,
1983 deps = MICROKERNEL_TEST_DEPS,
1984)
1985
1986xnnpack_unit_test(
1987 name = "x8_lut_test",
1988 srcs = [
1989 "test/x8-lut.cc",
1990 "test/lut-microkernel-tester.h",
1991 ] + MICROKERNEL_TEST_HDRS,
1992 deps = MICROKERNEL_TEST_DEPS,
1993)
1994
1995xnnpack_unit_test(
1996 name = "x8_zip_test",
1997 srcs = [
1998 "test/x8-zip.cc",
1999 "test/zip-microkernel-tester.h",
2000 ] + MICROKERNEL_TEST_HDRS,
2001 deps = MICROKERNEL_TEST_DEPS,
2002)
2003
2004########################### Size test for the library ##########################
2005
2006xnnpack_binary(
2007 name = "size_test",
2008 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002009 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002010)
2011
2012########################### Unit tests for operators ###########################
2013
2014xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002015 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002016 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002017 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002018 "test/add-operator-tester.h",
2019 ],
2020 deps = OPERATOR_TEST_DEPS,
2021)
2022
2023xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002024 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002025 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002026 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002027 "test/argmax-pooling-operator-tester.h",
2028 ] + OPERATOR_TEST_PARAMS_HDRS,
2029 deps = OPERATOR_TEST_DEPS,
2030)
2031
2032xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002033 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002034 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002035 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002036 "test/average-pooling-operator-tester.h",
2037 ] + OPERATOR_TEST_PARAMS_HDRS,
2038 deps = OPERATOR_TEST_DEPS,
2039)
2040
2041xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002042 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002043 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002044 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002045 "test/channel-pad-operator-tester.h",
2046 ] + OPERATOR_TEST_PARAMS_HDRS,
2047 deps = OPERATOR_TEST_DEPS,
2048)
2049
2050xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002051 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002052 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002053 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002054 "test/channel-shuffle-operator-tester.h",
2055 ],
2056 deps = OPERATOR_TEST_DEPS,
2057)
2058
2059xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002060 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002061 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002062 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002063 "test/clamp-operator-tester.h",
2064 ],
2065 deps = OPERATOR_TEST_DEPS,
2066)
2067
2068xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002069 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002070 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002071 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002072 "test/convolution-operator-tester.h",
2073 ],
2074 deps = OPERATOR_TEST_DEPS,
2075)
2076
2077xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002078 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002079 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002080 "test/convolution-nchw.cc",
2081 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002082 ],
2083 deps = OPERATOR_TEST_DEPS,
2084)
2085
2086xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002087 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002088 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002089 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002090 "test/deconvolution-operator-tester.h",
2091 ] + OPERATOR_TEST_PARAMS_HDRS,
2092 deps = OPERATOR_TEST_DEPS,
2093)
2094
2095xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002096 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002097 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002098 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002099 "test/fully-connected-operator-tester.h",
2100 ],
2101 deps = OPERATOR_TEST_DEPS,
2102)
2103
2104xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002105 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002106 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002107 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002108 "test/global-average-pooling-operator-tester.h",
2109 ] + OPERATOR_TEST_PARAMS_HDRS,
2110 deps = OPERATOR_TEST_DEPS,
2111)
2112
2113xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002114 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002115 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002116 "test/global-average-pooling-ncw.cc",
2117 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002118 ],
2119 deps = OPERATOR_TEST_DEPS,
2120)
2121
2122xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002123 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002124 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002125 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002126 "test/hardswish-operator-tester.h",
2127 ],
2128 deps = OPERATOR_TEST_DEPS,
2129)
2130
2131xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002132 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002133 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002134 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002135 "test/leaky-relu-operator-tester.h",
2136 ],
2137 deps = OPERATOR_TEST_DEPS,
2138)
2139
2140xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002141 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002142 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002143 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002144 "test/max-pooling-operator-tester.h",
2145 ] + OPERATOR_TEST_PARAMS_HDRS,
2146 deps = OPERATOR_TEST_DEPS,
2147)
2148
2149xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002150 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08002151 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002152 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08002153 "test/multiply-operator-tester.h",
2154 ],
2155 deps = OPERATOR_TEST_DEPS,
2156)
2157
2158xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002159 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002160 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002161 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002162 "test/prelu-operator-tester.h",
2163 ] + OPERATOR_TEST_PARAMS_HDRS,
2164 deps = OPERATOR_TEST_DEPS,
2165)
2166
2167xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002168 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08002169 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002170 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08002171 "test/resize-bilinear-operator-tester.h",
2172 ] + OPERATOR_TEST_PARAMS_HDRS,
2173 deps = OPERATOR_TEST_DEPS,
2174)
2175
2176xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002177 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002178 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002179 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002180 "test/sigmoid-operator-tester.h",
2181 ],
2182 deps = OPERATOR_TEST_DEPS,
2183)
2184
2185xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002186 name = "softargmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002187 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002188 "test/softargmax-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002189 "test/softargmax-operator-tester.h",
2190 ],
2191 deps = OPERATOR_TEST_DEPS,
2192)
2193
2194xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002195 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002196 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002197 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002198 "test/unpooling-operator-tester.h",
2199 ],
2200 deps = OPERATOR_TEST_DEPS,
2201)
2202
2203############################# Build configurations #############################
2204
Marat Dukhanb8642352019-10-30 15:43:02 -07002205# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07002206config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07002207 name = "xnn_enable_assembly_explicit_true",
2208 define_values = {"xnn_enable_assembly": "true"},
2209)
2210
2211# Disables usage of assembly kernels.
2212config_setting(
2213 name = "xnn_enable_assembly_explicit_false",
2214 define_values = {"xnn_enable_assembly": "false"},
2215)
2216
2217# Builds with -c dbg
2218config_setting(
2219 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002220 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07002221 "compilation_mode": "dbg",
2222 },
2223)
2224
2225# Builds with -c opt
2226config_setting(
2227 name = "optimized_build",
2228 values = {
2229 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002230 },
2231)
2232
2233config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07002234 name = "linux_k8",
2235 values = {"cpu": "k8"},
2236)
2237
2238config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07002239 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07002240 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07002241)
2242
2243config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002244 name = "android",
2245 values = {"crosstool_top": "//external:android/crosstool"},
2246)
2247
2248config_setting(
2249 name = "android_armv7",
2250 values = {
2251 "crosstool_top": "//external:android/crosstool",
2252 "cpu": "armeabi-v7a",
2253 },
2254)
2255
2256config_setting(
2257 name = "android_arm64",
2258 values = {
2259 "crosstool_top": "//external:android/crosstool",
2260 "cpu": "arm64-v8a",
2261 },
2262)
2263
2264config_setting(
2265 name = "android_x86",
2266 values = {
2267 "crosstool_top": "//external:android/crosstool",
2268 "cpu": "x86",
2269 },
2270)
2271
2272config_setting(
2273 name = "android_x86_64",
2274 values = {
2275 "crosstool_top": "//external:android/crosstool",
2276 "cpu": "x86_64",
2277 },
2278)
2279
2280config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07002281 name = "macos_x86_64",
2282 values = {
2283 "apple_platform_type": "macos",
2284 "cpu": "darwin",
2285 },
2286)
2287
2288config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002289 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002290 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07002291)
2292
2293config_setting(
2294 name = "emscripten_wasm",
2295 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002296 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002297 "cpu": "wasm",
2298 },
2299)
2300
2301config_setting(
2302 name = "emscripten_wasmsimd",
2303 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002304 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002305 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07002306 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002307 },
2308)
2309
2310config_setting(
2311 name = "emscripten_asmjs",
2312 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002313 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002314 "cpu": "asmjs",
2315 },
2316)