blob: 4e40b4f5b477077ccec015fd09867df99025f3a7 [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 Dukhane2c3f292019-11-27 15:40:54 -0800604 "src/f32-clamp/avx.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800605 "src/f32-dwconv/gen/up16x4-avx-acc2.c",
606 "src/f32-dwconv/gen/up16x4-avx.c",
607 "src/f32-dwconv/gen/up8x4-avx-acc2.c",
608 "src/f32-dwconv/gen/up8x4-avx.c",
609 "src/f32-dwconv/gen/up16x9-avx-acc2.c",
610 "src/f32-dwconv/gen/up16x9-avx.c",
611 "src/f32-dwconv/gen/up8x9-avx-acc2.c",
612 "src/f32-dwconv/gen/up8x9-avx.c",
613 "src/f32-dwconv/gen/up16x25-avx-acc2.c",
614 "src/f32-dwconv/gen/up16x25-avx.c",
615 "src/f32-dwconv/gen/up8x25-avx-acc2.c",
616 "src/f32-dwconv/gen/up8x25-avx.c",
617 "src/f32-gemm/gen/1x8-avx-broadcast.c",
618 "src/f32-gemm/gen/4x8-avx-broadcast.c",
619 "src/f32-gemm/gen/5x8-avx-broadcast.c",
620 "src/f32-gemm/gen/6x8-avx-broadcast.c",
621 "src/f32-gemm/gen/7x8-avx-broadcast.c",
622 "src/f32-gemm/gen-inc/1x8-avx-broadcast.c",
623 "src/f32-gemm/gen-inc/4x8-avx-broadcast.c",
624 "src/f32-gemm/gen-inc/5x8-avx-broadcast.c",
625 "src/f32-gemm/gen-inc/6x8-avx-broadcast.c",
626 "src/f32-gemm/gen-inc/7x8-avx-broadcast.c",
627 "src/f32-igemm/gen/1x8-avx-broadcast.c",
628 "src/f32-igemm/gen/4x8-avx-broadcast.c",
629 "src/f32-igemm/gen/5x8-avx-broadcast.c",
630 "src/f32-igemm/gen/6x8-avx-broadcast.c",
631 "src/f32-igemm/gen/7x8-avx-broadcast.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700632 "src/f32-rmax/avx.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700633 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700634]
635
Marat Dukhanfda12b82019-11-21 12:27:59 -0800636FMA3_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800637 "src/f32-dwconv/gen/up16x4-fma3-acc2.c",
638 "src/f32-dwconv/gen/up16x4-fma3.c",
639 "src/f32-dwconv/gen/up8x4-fma3-acc2.c",
640 "src/f32-dwconv/gen/up8x4-fma3.c",
641 "src/f32-dwconv/gen/up16x9-fma3-acc2.c",
642 "src/f32-dwconv/gen/up16x9-fma3.c",
643 "src/f32-dwconv/gen/up8x9-fma3-acc2.c",
644 "src/f32-dwconv/gen/up8x9-fma3.c",
645 "src/f32-dwconv/gen/up16x25-fma3-acc2.c",
646 "src/f32-dwconv/gen/up16x25-fma3.c",
647 "src/f32-dwconv/gen/up8x25-fma3-acc2.c",
648 "src/f32-dwconv/gen/up8x25-fma3.c",
649 "src/f32-gemm/gen/1x8-fma3-broadcast.c",
650 "src/f32-gemm/gen/4x8-fma3-broadcast.c",
651 "src/f32-gemm/gen/5x8-fma3-broadcast.c",
652 "src/f32-gemm/gen/6x8-fma3-broadcast.c",
653 "src/f32-gemm/gen/7x8-fma3-broadcast.c",
654 "src/f32-gemm/gen/8x8-fma3-broadcast.c",
655 "src/f32-gemm/gen-inc/1x8-fma3-broadcast.c",
656 "src/f32-gemm/gen-inc/4x8-fma3-broadcast.c",
657 "src/f32-gemm/gen-inc/5x8-fma3-broadcast.c",
658 "src/f32-gemm/gen-inc/6x8-fma3-broadcast.c",
659 "src/f32-gemm/gen-inc/7x8-fma3-broadcast.c",
660 "src/f32-gemm/gen-inc/8x8-fma3-broadcast.c",
661 "src/f32-igemm/gen/1x8-fma3-broadcast.c",
662 "src/f32-igemm/gen/4x8-fma3-broadcast.c",
663 "src/f32-igemm/gen/5x8-fma3-broadcast.c",
664 "src/f32-igemm/gen/6x8-fma3-broadcast.c",
665 "src/f32-igemm/gen/7x8-fma3-broadcast.c",
666 "src/f32-igemm/gen/8x8-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -0800667]
668
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700669AVX2_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700670 "src/f32-raddexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700671 "src/f32-raddextexp/avx2-p5-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700672 "src/f32-raddstoreexpminusmax/avx2-p5-unroll64.c",
673 "src/f32-vscaleexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700674 "src/f32-vscaleextexp/avx2-p5-unroll64.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700675 "src/math/exp-avx2-p5.c",
676 "src/math/exp-avx2-perm-p3.c",
677 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -0700678 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700679 "src/math/extexp-avx2-p5.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700680]
681
Marat Dukhan08c4a432019-10-03 09:29:21 -0700682AVX512F_UKERNELS = [
Marat Dukhane2c3f292019-11-27 15:40:54 -0800683 "src/f32-clamp/avx512f.c",
Marat Dukhan479f87e2019-11-27 15:17:06 -0800684 "src/f32-dwconv/gen/up32x4-avx512f-acc2.c",
685 "src/f32-dwconv/gen/up32x4-avx512f.c",
686 "src/f32-dwconv/gen/up16x4-avx512f-acc2.c",
687 "src/f32-dwconv/gen/up16x4-avx512f.c",
688 "src/f32-dwconv/gen/up32x9-avx512f-acc2.c",
689 "src/f32-dwconv/gen/up32x9-avx512f.c",
690 "src/f32-dwconv/gen/up16x9-avx512f-acc2.c",
691 "src/f32-dwconv/gen/up16x9-avx512f.c",
692 "src/f32-dwconv/gen/up32x25-avx512f-acc2.c",
693 "src/f32-dwconv/gen/up32x25-avx512f.c",
694 "src/f32-dwconv/gen/up16x25-avx512f-acc2.c",
695 "src/f32-dwconv/gen/up16x25-avx512f.c",
Marat Dukhan0f349c42019-11-27 11:58:54 -0800696 "src/f32-gemm/gen/1x16-avx512f-broadcast.c",
697 "src/f32-gemm/gen/4x16-avx512f-broadcast.c",
698 "src/f32-gemm/gen/5x16-avx512f-broadcast.c",
699 "src/f32-gemm/gen/6x16-avx512f-broadcast.c",
700 "src/f32-gemm/gen/7x16-avx512f-broadcast.c",
701 "src/f32-gemm/gen/8x16-avx512f-broadcast.c",
702 "src/f32-gemm/gen-inc/1x16-avx512f-broadcast.c",
703 "src/f32-gemm/gen-inc/4x16-avx512f-broadcast.c",
704 "src/f32-gemm/gen-inc/5x16-avx512f-broadcast.c",
705 "src/f32-gemm/gen-inc/6x16-avx512f-broadcast.c",
706 "src/f32-gemm/gen-inc/7x16-avx512f-broadcast.c",
707 "src/f32-gemm/gen-inc/8x16-avx512f-broadcast.c",
708 "src/f32-igemm/gen/1x16-avx512f-broadcast.c",
709 "src/f32-igemm/gen/4x16-avx512f-broadcast.c",
710 "src/f32-igemm/gen/5x16-avx512f-broadcast.c",
711 "src/f32-igemm/gen/6x16-avx512f-broadcast.c",
712 "src/f32-igemm/gen/7x16-avx512f-broadcast.c",
713 "src/f32-igemm/gen/8x16-avx512f-broadcast.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700714 "src/f32-raddexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700715 "src/f32-raddextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700716 "src/f32-raddstoreexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700717 "src/f32-rmax/avx512f.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700718 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700719 "src/f32-vscaleexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700720 "src/f32-vscaleextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700721 "src/math/exp-avx512f-p5-scalef.c",
722 "src/math/exp-avx512f-p5.c",
723 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -0700724 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700725 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700726]
727
728AARCH32_ASM_UKERNELS = [
729 "src/q8-dwconv/up8x9-aarch32-neon.S",
730]
731
732AARCH64_ASM_UKERNELS = [
733 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
734 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800735 "src/f32-gemm/gen/1x12-aarch64-neonfma-cortex-a53.S",
736 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a53.S",
737 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
738 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
739 "src/f32-gemm/gen/4x12-aarch64-neonfma-cortex-a53.S",
740 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a53.S",
741 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
742 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
743 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld128.S",
744 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld64.S",
745 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
746 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a53.S",
747 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
748 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a73.S",
749 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
750 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld128.S",
751 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld64.S",
752 "src/f32-gemm/gen-inc/1x12-aarch64-neonfma-cortex-a53.S",
753 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a53.S",
754 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a57.S",
755 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a75.S",
756 "src/f32-gemm/gen-inc/4x12-aarch64-neonfma-cortex-a53.S",
757 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a53.S",
758 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a57.S",
759 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a75.S",
760 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld128.S",
761 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld64.S",
762 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a75.S",
763 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a53.S",
764 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a57.S",
765 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a73.S",
766 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a75.S",
767 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld128.S",
768 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700769 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700770 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700771 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S",
772 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S",
773 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700774 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700775 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
776 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700777 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700778 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
779 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
780 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
781]
782
783INTERNAL_MICROKERNEL_HDRS = [
784 "src/xnnpack/argmaxpool.h",
785 "src/xnnpack/avgpool.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800786 "src/xnnpack/bilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700787 "src/xnnpack/clamp.h",
788 "src/xnnpack/common.h",
789 "src/xnnpack/conv.h",
790 "src/xnnpack/dwconv.h",
791 "src/xnnpack/gavgpool.h",
792 "src/xnnpack/gemm.h",
793 "src/xnnpack/hswish.h",
794 "src/xnnpack/igemm.h",
795 "src/xnnpack/lut.h",
796 "src/xnnpack/math.h",
797 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -0800798 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700799 "src/xnnpack/packx.h",
800 "src/xnnpack/pad.h",
801 "src/xnnpack/params.h",
802 "src/xnnpack/pavgpool.h",
803 "src/xnnpack/ppmm.h",
804 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700805 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700806 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700807 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700808 "src/xnnpack/rmax.h",
809 "src/xnnpack/scalar-utils.h",
810 "src/xnnpack/spmm.h",
811 "src/xnnpack/unpool.h",
812 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800813 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700814 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700815 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700816 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700817 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800818 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700819 "src/xnnpack/zip.h",
820]
821
822INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
823 "include/xnnpack.h",
824 "src/xnnpack/allocator.h",
825 "src/xnnpack/compute.h",
826 "src/xnnpack/im2col.h",
827 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700828 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700829 "src/xnnpack/operator.h",
830 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700831 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700832 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700833 "src/xnnpack/requantization.h",
834]
835
836ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
837 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700838]
839
840MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700841 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700842 "include/xnnpack.h",
843]
844
845MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
846 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700847 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700848 "src/xnnpack/requantization.h",
849 "include/xnnpack.h",
850]
851
852OPERATOR_TEST_PARAMS_HDRS = [
853 "src/xnnpack/params.h",
854 "src/xnnpack/common.h",
855]
856
857WEIGHTS_PACK_HDRS = [
858 "src/xnnpack/pack.h",
859 "src/xnnpack/operator.h",
860 "src/xnnpack/compute.h",
861]
862
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700863LOGGING_COPTS = select({
864 # No logging in optimized mode
865 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
866 # Full logging in debug mode
867 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
868 # Error-only logging in default (fastbuild) mode
869 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
870})
871
872LOGGING_HDRS = [
873 "src/xnnpack/log.h",
874]
875
Marat Dukhan08c4a432019-10-03 09:29:21 -0700876xnnpack_cc_library(
877 name = "scalar_ukernels",
878 srcs = SCALAR_UKERNELS,
879 hdrs = INTERNAL_HDRS,
880 aarch32_copts = ["-marm"],
881 copts = xnnpack_std_copts(),
882 deps = [
883 "@FP16",
884 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -0800885 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700886 ],
887)
888
889xnnpack_cc_library(
890 name = "psimd_ukernels",
891 srcs = PSIMD_UKERNELS,
892 hdrs = INTERNAL_HDRS,
893 aarch32_copts = [
894 "-marm",
895 "-mfpu=neon",
896 ],
897 copts = xnnpack_std_copts(),
898 optimized_copts = [
899 "-O3",
900 "-ffast-math",
901 ],
902 deps = [
903 "@FP16",
904 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -0800905 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700906 ],
907)
908
909xnnpack_cc_library(
910 name = "neon_ukernels",
911 hdrs = INTERNAL_HDRS,
912 aarch32_copts = [
913 "-marm",
914 "-mfpu=neon",
915 ],
916 aarch32_srcs = NEON_UKERNELS,
917 aarch64_srcs = NEON_UKERNELS,
918 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -0800919 deps = [
920 "@FP16",
921 "@pthreadpool",
922 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700923)
924
925xnnpack_cc_library(
926 name = "neonfma_ukernels",
927 hdrs = INTERNAL_HDRS,
928 aarch32_copts = [
929 "-marm",
930 "-mfpu=neon-vfpv4",
931 ],
932 aarch32_srcs = NEONFMA_UKERNELS,
933 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
934 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -0800935 deps = [
936 "@FP16",
937 "@pthreadpool",
938 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700939)
940
941xnnpack_cc_library(
942 name = "neonfp16arith_ukernels",
943 hdrs = INTERNAL_HDRS,
944 aarch64_copts = ["-march=armv8.2-a+fp16"],
945 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
946 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -0800947 deps = [
948 "@FP16",
949 "@pthreadpool",
950 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700951)
952
953xnnpack_cc_library(
954 name = "sse2_ukernels",
955 hdrs = INTERNAL_HDRS,
956 copts = xnnpack_std_copts(),
957 x86_copts = ["-msse2"],
958 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -0800959 deps = [
960 "@FP16",
961 "@pthreadpool",
962 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700963)
964
965xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800966 name = "sse41_ukernels",
967 hdrs = INTERNAL_HDRS,
968 copts = xnnpack_std_copts(),
969 x86_copts = ["-msse4.1"],
970 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -0800971 deps = [
972 "@FP16",
973 "@pthreadpool",
974 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800975)
976
977xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700978 name = "avx_ukernels",
979 hdrs = INTERNAL_HDRS,
980 copts = xnnpack_std_copts(),
981 x86_copts = ["-mavx"],
982 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -0800983 deps = [
984 "@FP16",
985 "@pthreadpool",
986 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700987)
988
989xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -0800990 name = "fma3_ukernels",
991 hdrs = INTERNAL_HDRS,
992 copts = xnnpack_std_copts(),
993 x86_copts = [
994 "-mfma",
995 ],
996 x86_srcs = FMA3_UKERNELS,
997 deps = [
998 "@FP16",
999 "@pthreadpool",
1000 ],
1001)
1002
1003xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001004 name = "avx2_ukernels",
1005 hdrs = INTERNAL_HDRS,
1006 copts = xnnpack_std_copts(),
1007 x86_copts = [
1008 "-mfma",
1009 "-mavx2",
1010 ],
1011 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001012 deps = [
1013 "@FP16",
1014 "@pthreadpool",
1015 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001016)
1017
1018xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001019 name = "avx512f_ukernels",
1020 hdrs = INTERNAL_HDRS,
1021 copts = xnnpack_std_copts(),
1022 x86_copts = ["-mavx512f"],
1023 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001024 deps = [
1025 "@FP16",
1026 "@pthreadpool",
1027 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001028)
1029
1030xnnpack_cc_library(
1031 name = "asm_ukernels",
1032 hdrs = ["src/xnnpack/assembly.h"],
1033 aarch32_srcs = AARCH32_ASM_UKERNELS,
1034 aarch64_srcs = AARCH64_ASM_UKERNELS,
1035)
1036
1037xnnpack_aggregate_library(
1038 name = "ukernels",
1039 aarch32_deps = [
1040 ":psimd_ukernels",
1041 ":neon_ukernels",
1042 ":neonfma_ukernels",
1043 ":asm_ukernels",
1044 ],
1045 aarch64_deps = [
1046 ":psimd_ukernels",
1047 ":neon_ukernels",
1048 ":neonfma_ukernels",
1049 ":neonfp16arith_ukernels",
1050 ":asm_ukernels",
1051 ],
1052 generic_deps = [":scalar_ukernels"],
1053 wasmsimd_deps = [
1054 ":psimd_ukernels",
1055 ],
1056 x86_deps = [
1057 ":psimd_ukernels",
1058 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001059 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001060 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001061 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001062 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001063 ":avx512f_ukernels",
1064 ],
1065)
1066
1067xnnpack_cc_library(
1068 name = "im2col",
1069 srcs = ["src/im2col.c"],
1070 hdrs = [
1071 "src/xnnpack/common.h",
1072 "src/xnnpack/im2col.h",
1073 ],
1074 copts = xnnpack_std_copts(),
1075)
1076
1077xnnpack_cc_library(
1078 name = "indirection",
1079 srcs = ["src/indirection.c"],
1080 hdrs = INTERNAL_HDRS,
1081 copts = xnnpack_std_copts(),
1082 deps = [
1083 "@FP16",
1084 "@FXdiv",
1085 "@pthreadpool",
1086 ],
1087)
1088
1089xnnpack_cc_library(
1090 name = "operator_run",
1091 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001092 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
1093 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001094 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
1095 "-Wno-vla",
1096 ],
1097 deps = [
1098 "@FP16",
1099 "@FXdiv",
1100 "@clog",
1101 "@pthreadpool",
1102 ],
1103)
1104
1105cc_library(
1106 name = "enable_assembly",
1107 defines = select({
1108 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1109 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001110 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001111 }),
1112)
1113
Marat Dukhancf056b22019-10-07 10:26:29 -07001114xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001115 name = "operators",
1116 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001117 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001118 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07001119 ],
1120 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001121 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001122 "-Isrc",
1123 "-Iinclude",
1124 ] + select({
1125 ":debug_build": [],
1126 "//conditions:default": xnnpack_min_size_copts(),
1127 }),
Marat Dukhancf056b22019-10-07 10:26:29 -07001128 wasm_srcs = ["src/wasm-stubs.c"],
1129 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001130 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001131 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001132 "@FP16",
1133 "@FXdiv",
1134 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001135 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001136 ],
1137)
1138
1139cc_library(
1140 name = "XNNPACK",
1141 srcs = [
1142 "src/init.c",
1143 ],
1144 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1145 "-Isrc",
1146 "-Iinclude",
1147 ] + select({
1148 ":debug_build": [],
1149 "//conditions:default": xnnpack_min_size_copts(),
1150 }),
1151 includes = ["include"],
1152 linkstatic = True,
1153 textual_hdrs = ["include/xnnpack.h"],
1154 visibility = xnnpack_visibility(),
1155 deps = [
1156 ":enable_assembly",
1157 ":ukernels",
1158 ":operator_run",
1159 ":operators",
1160 "@clog",
1161 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07001162 ] + select({
1163 ":emscripten": [],
1164 "//conditions:default": ["@cpuinfo"],
1165 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001166)
1167
1168cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001169 name = "xnnpack_operators_nhwc_f32",
1170 srcs = [
1171 "src/init.c",
1172 ],
1173 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1174 "-Isrc",
1175 "-Iinclude",
1176 ] + select({
1177 ":debug_build": [],
1178 "//conditions:default": xnnpack_min_size_copts(),
1179 }),
1180 defines = [
1181 "XNN_NO_Q8_OPERATORS",
1182 "XNN_NO_U8_OPERATORS",
1183 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08001184 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001185 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001186 includes = ["include"],
1187 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001188 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001189 visibility = xnnpack_visibility(),
1190 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001191 ":enable_assembly",
1192 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001193 ":operator_run",
1194 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001195 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001196 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001197 ] + select({
1198 ":emscripten": [],
1199 "//conditions:default": ["@cpuinfo"],
1200 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001201)
1202
Marat Dukhancf056b22019-10-07 10:26:29 -07001203xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001204 name = "bench_utils",
1205 srcs = ["bench/utils.cc"],
1206 hdrs = ["bench/utils.h"],
1207 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08001208 deps = [
1209 "@com_google_benchmark//:benchmark",
1210 "@cpuinfo",
1211 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001212)
1213
Frank Barchard7e955972019-10-11 10:34:25 -07001214######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07001215
1216xnnpack_benchmark(
1217 name = "q8_gemm_bench",
1218 srcs = [
1219 "bench/gemm.h",
1220 "bench/q8-gemm.cc",
1221 "src/xnnpack/AlignedAllocator.h",
1222 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1223 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
1224 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
1225)
1226
1227xnnpack_benchmark(
1228 name = "f16_gemm_bench",
1229 srcs = [
1230 "bench/f16-gemm.cc",
1231 "bench/gemm.h",
1232 "src/xnnpack/AlignedAllocator.h",
1233 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1234 copts = ["-Wno-unused-function"],
1235 deps = MICROKERNEL_BENCHMARK_DEPS,
1236)
1237
1238xnnpack_benchmark(
1239 name = "f32_igemm_bench",
1240 srcs = [
1241 "bench/f32-igemm.cc",
1242 "bench/conv.h",
1243 "src/xnnpack/AlignedAllocator.h",
1244 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07001245 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001246)
1247
1248xnnpack_benchmark(
1249 name = "f32_conv_hwc_bench",
1250 srcs = [
1251 "bench/f32-conv-hwc.cc",
1252 "bench/dconv.h",
1253 "src/xnnpack/AlignedAllocator.h",
1254 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1255 copts = ["-Wno-unused-function"],
1256 deps = MICROKERNEL_BENCHMARK_DEPS,
1257)
1258
1259xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07001260 name = "f32_conv_hwc2spchw_bench",
1261 srcs = [
1262 "bench/f32-conv-hwc2spchw.cc",
1263 "bench/dconv.h",
1264 "src/xnnpack/AlignedAllocator.h",
1265 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1266 copts = ["-Wno-unused-function"],
1267 deps = MICROKERNEL_BENCHMARK_DEPS,
1268)
1269
1270xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001271 name = "f32_dwconv_bench",
1272 srcs = [
1273 "bench/f32-dwconv.cc",
1274 "bench/dwconv.h",
1275 "src/xnnpack/AlignedAllocator.h",
1276 ] + MICROKERNEL_BENCHMARK_HDRS,
1277 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1278)
1279
1280xnnpack_benchmark(
1281 name = "f32_dwconv_spchw_bench",
1282 srcs = [
1283 "bench/f32-dwconv-spchw.cc",
1284 "bench/dwconv.h",
1285 "src/xnnpack/AlignedAllocator.h",
1286 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1287 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1288)
1289
1290xnnpack_benchmark(
1291 name = "f32_gemm_bench",
1292 srcs = [
1293 "bench/f32-gemm.cc",
1294 "bench/gemm.h",
1295 "src/xnnpack/AlignedAllocator.h",
1296 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1297 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07001298 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001299)
1300
1301xnnpack_benchmark(
1302 name = "f32_rmax_bench",
1303 srcs = [
1304 "bench/f32-rmax.cc",
1305 "src/xnnpack/AlignedAllocator.h",
1306 ] + MICROKERNEL_BENCHMARK_HDRS,
1307 deps = MICROKERNEL_BENCHMARK_DEPS,
1308)
1309
1310xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08001311 name = "f32_sigmoid_bench",
1312 srcs = [
1313 "bench/f32-sigmoid.cc",
1314 "src/xnnpack/AlignedAllocator.h",
1315 ] + MICROKERNEL_BENCHMARK_HDRS,
1316 copts = ["-Wno-unused-function"],
1317 deps = MICROKERNEL_BENCHMARK_DEPS,
1318)
1319
1320xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001321 name = "f32_spmm_bench",
1322 srcs = [
1323 "bench/f32-spmm.cc",
1324 "bench/gemm.h",
1325 "src/xnnpack/AlignedAllocator.h",
1326 ] + MICROKERNEL_BENCHMARK_HDRS,
1327 copts = ["-Wno-unused-function"],
1328 deps = MICROKERNEL_BENCHMARK_DEPS,
1329)
1330
1331xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07001332 name = "f32_softargmax_bench",
1333 srcs = [
1334 "bench/f32-softargmax.cc",
1335 ] + MICROKERNEL_BENCHMARK_HDRS,
1336 copts = ["-Wno-unused-function"],
1337 deps = MICROKERNEL_BENCHMARK_DEPS,
1338)
1339
1340xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001341 name = "f32_im2col_gemm_bench",
1342 srcs = [
1343 "bench/f32-im2col-gemm.cc",
1344 "bench/conv.h",
1345 "src/xnnpack/AlignedAllocator.h",
1346 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1347 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
1348)
1349
1350########################### Benchmarks for operators ###########################
1351
1352xnnpack_benchmark(
1353 name = "add_bench",
1354 srcs = ["bench/add.cc"],
1355 deps = OPERATOR_BENCHMARK_DEPS,
1356)
1357
1358xnnpack_benchmark(
1359 name = "average_pooling_bench",
1360 srcs = ["bench/average-pooling.cc"],
1361 deps = OPERATOR_BENCHMARK_DEPS,
1362)
1363
1364xnnpack_benchmark(
1365 name = "channel_shuffle_bench",
1366 srcs = ["bench/channel-shuffle.cc"],
1367 deps = OPERATOR_BENCHMARK_DEPS,
1368)
1369
1370xnnpack_benchmark(
1371 name = "convolution_bench",
1372 srcs = ["bench/convolution.cc"],
1373 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1374 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1375)
1376
1377xnnpack_benchmark(
1378 name = "deconvolution_bench",
1379 srcs = ["bench/deconvolution.cc"],
1380 copts = xnnpack_optional_tflite_copts(),
1381 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1382)
1383
1384xnnpack_benchmark(
1385 name = "global_average_pooling_bench",
1386 srcs = ["bench/global-average-pooling.cc"],
1387 deps = OPERATOR_BENCHMARK_DEPS,
1388)
1389
1390xnnpack_benchmark(
1391 name = "max_pooling_bench",
1392 srcs = ["bench/max-pooling.cc"],
1393 deps = OPERATOR_BENCHMARK_DEPS,
1394)
1395
1396xnnpack_benchmark(
1397 name = "sigmoid_bench",
1398 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08001399 copts = xnnpack_optional_tflite_copts(),
1400 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001401)
1402
1403xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07001404 name = "prelu_bench",
1405 srcs = ["bench/prelu.cc"],
1406 copts = xnnpack_optional_tflite_copts(),
1407 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1408)
1409
1410xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001411 name = "softargmax_bench",
1412 srcs = ["bench/softargmax.cc"],
1413 deps = OPERATOR_BENCHMARK_DEPS,
1414)
1415
Marat Dukhanc068bb62019-10-04 13:24:39 -07001416############################# End-to-end benchmarks ############################
1417
1418cc_library(
1419 name = "mobilenet_v1",
1420 srcs = ["models/mobilenet-v1.cc"],
1421 hdrs = ["models/models.h"],
1422 linkstatic = True,
1423 deps = [
1424 ":XNNPACK",
1425 "@pthreadpool",
1426 ],
1427)
1428
1429cc_library(
1430 name = "mobilenet_v2",
1431 srcs = ["models/mobilenet-v2.cc"],
1432 hdrs = ["models/models.h"],
1433 linkstatic = True,
1434 deps = [
1435 ":XNNPACK",
1436 "@pthreadpool",
1437 ],
1438)
1439
1440xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07001441 name = "f32_dwconv_e2e_bench",
1442 srcs = ["bench/f32-dwconv-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1443 copts = ["-Wno-unused-function"],
1444 deps = MICROKERNEL_BENCHMARK_DEPS + [
1445 ":XNNPACK",
1446 ":mobilenet_v1",
1447 ":mobilenet_v2",
1448 ],
1449)
1450
1451xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07001452 name = "f32_gemm_e2e_bench",
1453 srcs = ["bench/f32-gemm-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1454 copts = ["-Wno-unused-function"],
1455 deps = MICROKERNEL_BENCHMARK_DEPS + [
1456 ":XNNPACK",
1457 ":mobilenet_v1",
1458 ":mobilenet_v2",
1459 ],
1460)
1461
1462xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07001463 name = "end2end_bench",
1464 srcs = ["bench/end2end.cc"],
1465 deps = [
1466 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07001467 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07001468 ":mobilenet_v1",
1469 ":mobilenet_v2",
1470 "@pthreadpool",
1471 ],
1472)
1473
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001474#################### Accuracy evaluation for math functions ####################
1475
1476xnnpack_benchmark(
1477 name = "f32_exp_eval",
1478 srcs = [
1479 "eval/f32-exp.cc",
1480 "src/xnnpack/AlignedAllocator.h",
1481 ] + ACCURACY_EVAL_HDRS,
1482 deps = ACCURACY_EVAL_DEPS,
1483)
1484
Marat Dukhan515c9772019-10-17 18:07:57 -07001485xnnpack_benchmark(
1486 name = "f32_expminus_eval",
1487 srcs = [
1488 "eval/f32-expminus.cc",
1489 "src/xnnpack/AlignedAllocator.h",
1490 ] + ACCURACY_EVAL_HDRS,
1491 deps = ACCURACY_EVAL_DEPS,
1492)
1493
Marat Dukhan98ba4412019-10-23 02:14:28 -07001494xnnpack_benchmark(
1495 name = "f32_extexp_eval",
1496 srcs = [
1497 "eval/f32-extexp.cc",
1498 "src/xnnpack/AlignedAllocator.h",
1499 ] + ACCURACY_EVAL_HDRS,
1500 deps = ACCURACY_EVAL_DEPS,
1501)
1502
Marat Dukhan346a9e52019-11-15 09:06:30 -08001503xnnpack_benchmark(
1504 name = "f32_sigmoid_eval",
1505 srcs = [
1506 "eval/f32-sigmoid.cc",
1507 "src/xnnpack/AlignedAllocator.h",
1508 ] + ACCURACY_EVAL_HDRS,
1509 deps = ACCURACY_EVAL_DEPS,
1510)
1511
Marat Dukhan08c4a432019-10-03 09:29:21 -07001512######################### Unit tests for micro-kernels #########################
1513
1514xnnpack_unit_test(
1515 name = "f16_gemm_test",
1516 srcs = [
1517 "test/f16-gemm.cc",
1518 "test/gemm-microkernel-tester.h",
1519 "src/xnnpack/AlignedAllocator.h",
1520 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1521 deps = MICROKERNEL_TEST_DEPS,
1522)
1523
1524xnnpack_unit_test(
1525 name = "f32_argmaxpool_test",
1526 srcs = [
1527 "test/f32-argmaxpool.cc",
1528 "test/argmaxpool-microkernel-tester.h",
1529 "src/xnnpack/AlignedAllocator.h",
1530 ] + MICROKERNEL_TEST_HDRS,
1531 deps = MICROKERNEL_TEST_DEPS,
1532)
1533
1534xnnpack_unit_test(
1535 name = "f32_avgpool_test",
1536 srcs = [
1537 "test/f32-avgpool.cc",
1538 "test/avgpool-microkernel-tester.h",
1539 "src/xnnpack/AlignedAllocator.h",
1540 ] + MICROKERNEL_TEST_HDRS,
1541 deps = MICROKERNEL_TEST_DEPS,
1542)
1543
1544xnnpack_unit_test(
Marat Dukhan35dacfb2019-11-07 19:18:16 -08001545 name = "f32_bilinear_test",
1546 srcs = [
1547 "test/f32-bilinear.cc",
1548 "test/bilinear-microkernel-tester.h",
1549 "src/xnnpack/AlignedAllocator.h",
1550 ] + MICROKERNEL_TEST_HDRS,
1551 deps = MICROKERNEL_TEST_DEPS,
1552)
1553
1554xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001555 name = "f32_clamp_test",
1556 srcs = [
1557 "test/f32-clamp.cc",
1558 "test/clamp-microkernel-tester.h",
1559 ] + MICROKERNEL_TEST_HDRS,
1560 deps = MICROKERNEL_TEST_DEPS,
1561)
1562
1563xnnpack_unit_test(
1564 name = "f32_igemm_test",
1565 srcs = [
1566 "test/f32-igemm.cc",
1567 "test/gemm-microkernel-tester.h",
1568 "src/xnnpack/AlignedAllocator.h",
1569 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1570 deps = MICROKERNEL_TEST_DEPS,
1571)
1572
1573xnnpack_unit_test(
1574 name = "f32_conv_hwc_test",
1575 srcs = [
1576 "test/f32-conv-hwc.cc",
1577 "test/conv-hwc-microkernel-tester.h",
1578 "src/xnnpack/AlignedAllocator.h",
1579 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1580 deps = MICROKERNEL_TEST_DEPS,
1581)
1582
1583xnnpack_unit_test(
1584 name = "f32_conv_hwc2spchw_test",
1585 srcs = [
1586 "test/f32-conv-hwc2spchw.cc",
1587 "test/conv-hwc2spchw-microkernel-tester.h",
1588 "src/xnnpack/AlignedAllocator.h",
1589 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1590 deps = MICROKERNEL_TEST_DEPS,
1591)
1592
1593xnnpack_unit_test(
1594 name = "f32_dwconv_test",
1595 srcs = [
1596 "test/f32-dwconv.cc",
1597 "test/dwconv-microkernel-tester.h",
1598 "src/xnnpack/AlignedAllocator.h",
1599 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1600 deps = MICROKERNEL_TEST_DEPS,
1601)
1602
1603xnnpack_unit_test(
1604 name = "f32_dwconv_spchw_test",
1605 srcs = [
1606 "test/f32-dwconv-spchw.cc",
1607 "test/dwconv-spchw-microkernel-tester.h",
1608 "src/xnnpack/AlignedAllocator.h",
1609 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1610 deps = MICROKERNEL_TEST_DEPS,
1611)
1612
1613xnnpack_unit_test(
1614 name = "f32_gavgpool_test",
1615 srcs = [
1616 "test/f32-gavgpool.cc",
1617 "test/gavgpool-microkernel-tester.h",
1618 "src/xnnpack/AlignedAllocator.h",
1619 ] + MICROKERNEL_TEST_HDRS,
1620 deps = MICROKERNEL_TEST_DEPS,
1621)
1622
1623xnnpack_unit_test(
1624 name = "f32_gavgpool_spchw_test",
1625 srcs = [
1626 "test/f32-gavgpool-spchw.cc",
1627 "test/gavgpool-spchw-microkernel-tester.h",
1628 "src/xnnpack/AlignedAllocator.h",
1629 ] + MICROKERNEL_TEST_HDRS,
1630 deps = MICROKERNEL_TEST_DEPS,
1631)
1632
1633xnnpack_unit_test(
1634 name = "f32_gemm_test",
1635 srcs = [
1636 "test/f32-gemm.cc",
1637 "test/gemm-microkernel-tester.h",
1638 "src/xnnpack/AlignedAllocator.h",
1639 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1640 deps = MICROKERNEL_TEST_DEPS,
1641)
1642
1643xnnpack_unit_test(
1644 name = "f32_gemminc_test",
1645 srcs = [
1646 "test/f32-gemminc.cc",
1647 "test/gemm-microkernel-tester.h",
1648 "src/xnnpack/AlignedAllocator.h",
1649 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1650 deps = MICROKERNEL_TEST_DEPS,
1651)
1652
1653xnnpack_unit_test(
1654 name = "f32_hswish_test",
1655 srcs = [
1656 "test/f32-hswish.cc",
1657 "test/hswish-microkernel-tester.h",
1658 ] + MICROKERNEL_TEST_HDRS,
1659 deps = MICROKERNEL_TEST_DEPS,
1660)
1661
1662xnnpack_unit_test(
1663 name = "f32_maxpool_test",
1664 srcs = [
1665 "test/f32-maxpool.cc",
1666 "test/maxpool-microkernel-tester.h",
1667 ] + MICROKERNEL_TEST_HDRS,
1668 deps = MICROKERNEL_TEST_DEPS,
1669)
1670
1671xnnpack_unit_test(
1672 name = "f32_pavgpool_test",
1673 srcs = [
1674 "test/f32-pavgpool.cc",
1675 "test/avgpool-microkernel-tester.h",
1676 "src/xnnpack/AlignedAllocator.h",
1677 ] + MICROKERNEL_TEST_HDRS,
1678 deps = MICROKERNEL_TEST_DEPS,
1679)
1680
1681xnnpack_unit_test(
1682 name = "f32_ppmm_test",
1683 srcs = [
1684 "test/f32-ppmm.cc",
1685 "test/gemm-microkernel-tester.h",
1686 "src/xnnpack/AlignedAllocator.h",
1687 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1688 deps = MICROKERNEL_TEST_DEPS,
1689)
1690
1691xnnpack_unit_test(
1692 name = "f32_prelu_test",
1693 srcs = [
1694 "test/f32-prelu.cc",
1695 "test/prelu-microkernel-tester.h",
1696 "src/xnnpack/AlignedAllocator.h",
1697 ] + MICROKERNEL_TEST_HDRS,
1698 deps = MICROKERNEL_TEST_DEPS,
1699)
1700
1701xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001702 name = "f32_raddexpminusmax_test",
1703 srcs = [
1704 "test/f32-raddexpminusmax.cc",
1705 "test/raddexpminusmax-microkernel-tester.h",
1706 ] + MICROKERNEL_TEST_HDRS,
1707 deps = MICROKERNEL_TEST_DEPS,
1708)
1709
1710xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001711 name = "f32_raddextexp_test",
1712 srcs = [
1713 "test/f32-raddextexp.cc",
1714 "test/raddextexp-microkernel-tester.h",
1715 ] + MICROKERNEL_TEST_HDRS,
1716 deps = MICROKERNEL_TEST_DEPS,
1717)
1718
1719xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001720 name = "f32_raddstoreexpminusmax_test",
1721 srcs = [
1722 "test/f32-raddstoreexpminusmax.cc",
1723 "test/raddstoreexpminusmax-microkernel-tester.h",
1724 ] + MICROKERNEL_TEST_HDRS,
1725 deps = MICROKERNEL_TEST_DEPS,
1726)
1727
1728xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001729 name = "f32_rmax_test",
1730 srcs = [
1731 "test/f32-rmax.cc",
1732 "test/rmax-microkernel-tester.h",
1733 ] + MICROKERNEL_TEST_HDRS,
1734 deps = MICROKERNEL_TEST_DEPS,
1735)
1736
1737xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08001738 name = "f32_sigmoid_test",
1739 srcs = [
1740 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001741 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001742 ] + MICROKERNEL_TEST_HDRS,
1743 deps = MICROKERNEL_TEST_DEPS,
1744)
1745
1746xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001747 name = "f32_spmm_test",
1748 srcs = [
1749 "test/f32-spmm.cc",
1750 "test/spmm-microkernel-tester.h",
1751 "src/xnnpack/AlignedAllocator.h",
1752 ] + MICROKERNEL_TEST_HDRS,
1753 deps = MICROKERNEL_TEST_DEPS,
1754)
1755
1756xnnpack_unit_test(
1757 name = "f32_vadd_test",
1758 srcs = [
1759 "test/f32-vadd.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001760 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001761 ] + MICROKERNEL_TEST_HDRS,
1762 deps = MICROKERNEL_TEST_DEPS,
1763)
1764
1765xnnpack_unit_test(
1766 name = "f32_vaddc_test",
1767 srcs = [
1768 "test/f32-vaddc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001769 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001770 ] + MICROKERNEL_TEST_HDRS,
1771 deps = MICROKERNEL_TEST_DEPS,
1772)
1773
1774xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001775 name = "f32_vmul_test",
1776 srcs = [
1777 "test/f32-vmul.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001778 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001779 ] + MICROKERNEL_TEST_HDRS,
1780 deps = MICROKERNEL_TEST_DEPS,
1781)
1782
1783xnnpack_unit_test(
1784 name = "f32_vmulc_test",
1785 srcs = [
1786 "test/f32-vmulc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001787 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001788 ] + MICROKERNEL_TEST_HDRS,
1789 deps = MICROKERNEL_TEST_DEPS,
1790)
1791
1792xnnpack_unit_test(
1793 name = "f32_vmulcaddc_test",
1794 srcs = [
1795 "test/f32-vmulcaddc.cc",
1796 "test/vmulcaddc-microkernel-tester.h",
1797 "src/xnnpack/AlignedAllocator.h",
1798 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1799 deps = MICROKERNEL_TEST_DEPS,
1800)
1801
1802xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001803 name = "f32_vscale_test",
1804 srcs = [
1805 "test/f32-vscale.cc",
1806 "test/vscale-microkernel-tester.h",
1807 ] + MICROKERNEL_TEST_HDRS,
1808 deps = MICROKERNEL_TEST_DEPS,
1809)
1810
1811xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001812 name = "f32_vscaleexpminusmax_test",
1813 srcs = [
1814 "test/f32-vscaleexpminusmax.cc",
1815 "test/vscaleexpminusmax-microkernel-tester.h",
1816 ] + MICROKERNEL_TEST_HDRS,
1817 deps = MICROKERNEL_TEST_DEPS,
1818)
1819
1820xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001821 name = "f32_vscaleextexp_test",
1822 srcs = [
1823 "test/f32-vscaleextexp.cc",
1824 "test/vscaleextexp-microkernel-tester.h",
1825 ] + MICROKERNEL_TEST_HDRS,
1826 deps = MICROKERNEL_TEST_DEPS,
1827)
1828
1829xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001830 name = "f32_vsub_test",
1831 srcs = [
1832 "test/f32-vsub.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001833 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001834 ] + MICROKERNEL_TEST_HDRS,
1835 deps = MICROKERNEL_TEST_DEPS,
1836)
1837
1838xnnpack_unit_test(
1839 name = "f32_vsubc_test",
1840 srcs = [
1841 "test/f32-vsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001842 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001843 ] + MICROKERNEL_TEST_HDRS,
1844 deps = MICROKERNEL_TEST_DEPS,
1845)
1846
1847xnnpack_unit_test(
1848 name = "f32_vrsubc_test",
1849 srcs = [
1850 "test/f32-vrsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001851 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001852 ] + MICROKERNEL_TEST_HDRS,
1853 deps = MICROKERNEL_TEST_DEPS,
1854)
1855
1856xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001857 name = "q8_avgpool_test",
1858 srcs = [
1859 "test/q8-avgpool.cc",
1860 "test/avgpool-microkernel-tester.h",
1861 "src/xnnpack/AlignedAllocator.h",
1862 ] + MICROKERNEL_TEST_HDRS,
1863 deps = MICROKERNEL_TEST_DEPS,
1864)
1865
1866xnnpack_unit_test(
1867 name = "q8_igemm_test",
1868 srcs = [
1869 "test/q8-igemm.cc",
1870 "test/gemm-microkernel-tester.h",
1871 "src/xnnpack/AlignedAllocator.h",
1872 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1873 deps = MICROKERNEL_TEST_DEPS,
1874)
1875
1876xnnpack_unit_test(
1877 name = "q8_dwconv_test",
1878 srcs = [
1879 "test/q8-dwconv.cc",
1880 "test/dwconv-microkernel-tester.h",
1881 "src/xnnpack/AlignedAllocator.h",
1882 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1883 deps = MICROKERNEL_TEST_DEPS,
1884)
1885
1886xnnpack_unit_test(
1887 name = "q8_gavgpool_test",
1888 srcs = [
1889 "test/q8-gavgpool.cc",
1890 "test/gavgpool-microkernel-tester.h",
1891 "src/xnnpack/AlignedAllocator.h",
1892 ] + MICROKERNEL_TEST_HDRS,
1893 deps = MICROKERNEL_TEST_DEPS,
1894)
1895
1896xnnpack_unit_test(
1897 name = "q8_gemm_test",
1898 srcs = [
1899 "test/q8-gemm.cc",
1900 "test/gemm-microkernel-tester.h",
1901 "src/xnnpack/AlignedAllocator.h",
1902 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1903 deps = MICROKERNEL_TEST_DEPS,
1904)
1905
1906xnnpack_unit_test(
1907 name = "q8_vadd_test",
1908 srcs = [
1909 "test/q8-vadd.cc",
1910 "test/vadd-microkernel-tester.h",
1911 ] + MICROKERNEL_TEST_HDRS,
1912 deps = MICROKERNEL_TEST_DEPS,
1913)
1914
1915xnnpack_unit_test(
1916 name = "u8_clamp_test",
1917 srcs = [
1918 "test/u8-clamp.cc",
1919 "test/clamp-microkernel-tester.h",
1920 ] + MICROKERNEL_TEST_HDRS,
1921 deps = MICROKERNEL_TEST_DEPS,
1922)
1923
1924xnnpack_unit_test(
1925 name = "u8_lut32norm_test",
1926 srcs = [
1927 "test/u8-lut32norm.cc",
1928 "test/lut-norm-microkernel-tester.h",
1929 ] + MICROKERNEL_TEST_HDRS,
1930 deps = MICROKERNEL_TEST_DEPS,
1931)
1932
1933xnnpack_unit_test(
1934 name = "u8_maxpool_test",
1935 srcs = [
1936 "test/u8-maxpool.cc",
1937 "test/maxpool-microkernel-tester.h",
1938 ] + MICROKERNEL_TEST_HDRS,
1939 deps = MICROKERNEL_TEST_DEPS,
1940)
1941
1942xnnpack_unit_test(
1943 name = "u8_rmax_test",
1944 srcs = [
1945 "test/u8-rmax.cc",
1946 "test/rmax-microkernel-tester.h",
1947 ] + MICROKERNEL_TEST_HDRS,
1948 deps = MICROKERNEL_TEST_DEPS,
1949)
1950
1951xnnpack_unit_test(
1952 name = "x32_packx_test",
1953 srcs = [
1954 "test/x32-packx.cc",
1955 "test/pack-microkernel-tester.h",
1956 "src/xnnpack/AlignedAllocator.h",
1957 ] + MICROKERNEL_TEST_HDRS,
1958 deps = MICROKERNEL_TEST_DEPS,
1959)
1960
1961xnnpack_unit_test(
1962 name = "x32_pad_test",
1963 srcs = [
1964 "test/x32-pad.cc",
1965 "test/pad-microkernel-tester.h",
1966 ] + MICROKERNEL_TEST_HDRS,
1967 deps = MICROKERNEL_TEST_DEPS,
1968)
1969
1970xnnpack_unit_test(
1971 name = "x32_unpool_test",
1972 srcs = [
1973 "test/x32-unpool.cc",
1974 "test/unpool-microkernel-tester.h",
1975 ] + MICROKERNEL_TEST_HDRS,
1976 deps = MICROKERNEL_TEST_DEPS,
1977)
1978
1979xnnpack_unit_test(
1980 name = "x32_zip_test",
1981 srcs = [
1982 "test/x32-zip.cc",
1983 "test/zip-microkernel-tester.h",
1984 ] + MICROKERNEL_TEST_HDRS,
1985 deps = MICROKERNEL_TEST_DEPS,
1986)
1987
1988xnnpack_unit_test(
1989 name = "x8_lut_test",
1990 srcs = [
1991 "test/x8-lut.cc",
1992 "test/lut-microkernel-tester.h",
1993 ] + MICROKERNEL_TEST_HDRS,
1994 deps = MICROKERNEL_TEST_DEPS,
1995)
1996
1997xnnpack_unit_test(
1998 name = "x8_zip_test",
1999 srcs = [
2000 "test/x8-zip.cc",
2001 "test/zip-microkernel-tester.h",
2002 ] + MICROKERNEL_TEST_HDRS,
2003 deps = MICROKERNEL_TEST_DEPS,
2004)
2005
2006########################### Size test for the library ##########################
2007
2008xnnpack_binary(
2009 name = "size_test",
2010 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002011 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002012)
2013
2014########################### Unit tests for operators ###########################
2015
2016xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002017 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002018 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002019 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002020 "test/add-operator-tester.h",
2021 ],
2022 deps = OPERATOR_TEST_DEPS,
2023)
2024
2025xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002026 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002027 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002028 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002029 "test/argmax-pooling-operator-tester.h",
2030 ] + OPERATOR_TEST_PARAMS_HDRS,
2031 deps = OPERATOR_TEST_DEPS,
2032)
2033
2034xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002035 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002036 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002037 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002038 "test/average-pooling-operator-tester.h",
2039 ] + OPERATOR_TEST_PARAMS_HDRS,
2040 deps = OPERATOR_TEST_DEPS,
2041)
2042
2043xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002044 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002045 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002046 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002047 "test/channel-pad-operator-tester.h",
2048 ] + OPERATOR_TEST_PARAMS_HDRS,
2049 deps = OPERATOR_TEST_DEPS,
2050)
2051
2052xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002053 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002054 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002055 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002056 "test/channel-shuffle-operator-tester.h",
2057 ],
2058 deps = OPERATOR_TEST_DEPS,
2059)
2060
2061xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002062 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002063 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002064 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002065 "test/clamp-operator-tester.h",
2066 ],
2067 deps = OPERATOR_TEST_DEPS,
2068)
2069
2070xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002071 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002072 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002073 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002074 "test/convolution-operator-tester.h",
2075 ],
2076 deps = OPERATOR_TEST_DEPS,
2077)
2078
2079xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002080 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002081 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002082 "test/convolution-nchw.cc",
2083 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002084 ],
2085 deps = OPERATOR_TEST_DEPS,
2086)
2087
2088xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002089 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002090 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002091 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002092 "test/deconvolution-operator-tester.h",
2093 ] + OPERATOR_TEST_PARAMS_HDRS,
2094 deps = OPERATOR_TEST_DEPS,
2095)
2096
2097xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002098 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002099 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002100 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002101 "test/fully-connected-operator-tester.h",
2102 ],
2103 deps = OPERATOR_TEST_DEPS,
2104)
2105
2106xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002107 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002108 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002109 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002110 "test/global-average-pooling-operator-tester.h",
2111 ] + OPERATOR_TEST_PARAMS_HDRS,
2112 deps = OPERATOR_TEST_DEPS,
2113)
2114
2115xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002116 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002117 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002118 "test/global-average-pooling-ncw.cc",
2119 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002120 ],
2121 deps = OPERATOR_TEST_DEPS,
2122)
2123
2124xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002125 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002126 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002127 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002128 "test/hardswish-operator-tester.h",
2129 ],
2130 deps = OPERATOR_TEST_DEPS,
2131)
2132
2133xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002134 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002135 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002136 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002137 "test/leaky-relu-operator-tester.h",
2138 ],
2139 deps = OPERATOR_TEST_DEPS,
2140)
2141
2142xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002143 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002144 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002145 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002146 "test/max-pooling-operator-tester.h",
2147 ] + OPERATOR_TEST_PARAMS_HDRS,
2148 deps = OPERATOR_TEST_DEPS,
2149)
2150
2151xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002152 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08002153 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002154 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08002155 "test/multiply-operator-tester.h",
2156 ],
2157 deps = OPERATOR_TEST_DEPS,
2158)
2159
2160xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002161 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002162 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002163 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002164 "test/prelu-operator-tester.h",
2165 ] + OPERATOR_TEST_PARAMS_HDRS,
2166 deps = OPERATOR_TEST_DEPS,
2167)
2168
2169xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002170 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08002171 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002172 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08002173 "test/resize-bilinear-operator-tester.h",
2174 ] + OPERATOR_TEST_PARAMS_HDRS,
2175 deps = OPERATOR_TEST_DEPS,
2176)
2177
2178xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002179 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002180 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002181 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002182 "test/sigmoid-operator-tester.h",
2183 ],
2184 deps = OPERATOR_TEST_DEPS,
2185)
2186
2187xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002188 name = "softargmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002189 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002190 "test/softargmax-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002191 "test/softargmax-operator-tester.h",
2192 ],
2193 deps = OPERATOR_TEST_DEPS,
2194)
2195
2196xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002197 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002198 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002199 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002200 "test/unpooling-operator-tester.h",
2201 ],
2202 deps = OPERATOR_TEST_DEPS,
2203)
2204
2205############################# Build configurations #############################
2206
Marat Dukhanb8642352019-10-30 15:43:02 -07002207# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07002208config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07002209 name = "xnn_enable_assembly_explicit_true",
2210 define_values = {"xnn_enable_assembly": "true"},
2211)
2212
2213# Disables usage of assembly kernels.
2214config_setting(
2215 name = "xnn_enable_assembly_explicit_false",
2216 define_values = {"xnn_enable_assembly": "false"},
2217)
2218
2219# Builds with -c dbg
2220config_setting(
2221 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002222 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07002223 "compilation_mode": "dbg",
2224 },
2225)
2226
2227# Builds with -c opt
2228config_setting(
2229 name = "optimized_build",
2230 values = {
2231 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002232 },
2233)
2234
2235config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07002236 name = "linux_k8",
2237 values = {"cpu": "k8"},
2238)
2239
2240config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07002241 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07002242 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07002243)
2244
2245config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002246 name = "android",
2247 values = {"crosstool_top": "//external:android/crosstool"},
2248)
2249
2250config_setting(
2251 name = "android_armv7",
2252 values = {
2253 "crosstool_top": "//external:android/crosstool",
2254 "cpu": "armeabi-v7a",
2255 },
2256)
2257
2258config_setting(
2259 name = "android_arm64",
2260 values = {
2261 "crosstool_top": "//external:android/crosstool",
2262 "cpu": "arm64-v8a",
2263 },
2264)
2265
2266config_setting(
2267 name = "android_x86",
2268 values = {
2269 "crosstool_top": "//external:android/crosstool",
2270 "cpu": "x86",
2271 },
2272)
2273
2274config_setting(
2275 name = "android_x86_64",
2276 values = {
2277 "crosstool_top": "//external:android/crosstool",
2278 "cpu": "x86_64",
2279 },
2280)
2281
2282config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07002283 name = "macos_x86_64",
2284 values = {
2285 "apple_platform_type": "macos",
2286 "cpu": "darwin",
2287 },
2288)
2289
2290config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002291 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002292 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07002293)
2294
2295config_setting(
2296 name = "emscripten_wasm",
2297 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002298 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002299 "cpu": "wasm",
2300 },
2301)
2302
2303config_setting(
2304 name = "emscripten_wasmsimd",
2305 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002306 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002307 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07002308 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002309 },
2310)
2311
2312config_setting(
2313 name = "emscripten_asmjs",
2314 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002315 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002316 "cpu": "asmjs",
2317 },
2318)