blob: 4c58e0cfa2572c411ecfe33eef026cfcf31b9aa3 [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",
Marat Dukhanb1a0fc32019-12-02 19:32:02 -080056 "src/binary-elementwise-nd.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080057 "src/channel-pad-nc.c",
58 "src/channel-shuffle-nc.c",
59 "src/clamp-nc.c",
60 "src/convolution-nchw.c",
61 "src/convolution-nhwc.c",
62 "src/deconvolution-nhwc.c",
63 "src/fully-connected-nc.c",
64 "src/global-average-pooling-ncw.c",
65 "src/global-average-pooling-nwc.c",
66 "src/hardswish-nc.c",
67 "src/leaky-relu-nc.c",
68 "src/max-pooling-nhwc.c",
Marat Dukhanefc47b82019-11-18 09:25:38 -080069 "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",
Frank Barchard8b0f0262019-11-27 23:18:40 -0800730 "src/f32-gemm/4x8-aarch32-neon-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700731]
732
733AARCH64_ASM_UKERNELS = [
734 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
735 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800736 "src/f32-gemm/gen/1x12-aarch64-neonfma-cortex-a53.S",
737 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a53.S",
738 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
739 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
740 "src/f32-gemm/gen/4x12-aarch64-neonfma-cortex-a53.S",
741 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a53.S",
742 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
743 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
744 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld128.S",
745 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld64.S",
746 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
747 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a53.S",
748 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
749 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a73.S",
750 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
751 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld128.S",
752 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld64.S",
753 "src/f32-gemm/gen-inc/1x12-aarch64-neonfma-cortex-a53.S",
754 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a53.S",
755 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a57.S",
756 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a75.S",
757 "src/f32-gemm/gen-inc/4x12-aarch64-neonfma-cortex-a53.S",
758 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a53.S",
759 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a57.S",
760 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a75.S",
761 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld128.S",
762 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld64.S",
763 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a75.S",
764 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a53.S",
765 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a57.S",
766 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a73.S",
767 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a75.S",
768 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld128.S",
769 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700770 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700771 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700772 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S",
773 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S",
774 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700775 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700776 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
777 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700778 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700779 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
780 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
781 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
782]
783
784INTERNAL_MICROKERNEL_HDRS = [
785 "src/xnnpack/argmaxpool.h",
786 "src/xnnpack/avgpool.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800787 "src/xnnpack/bilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700788 "src/xnnpack/clamp.h",
789 "src/xnnpack/common.h",
790 "src/xnnpack/conv.h",
791 "src/xnnpack/dwconv.h",
792 "src/xnnpack/gavgpool.h",
793 "src/xnnpack/gemm.h",
794 "src/xnnpack/hswish.h",
795 "src/xnnpack/igemm.h",
796 "src/xnnpack/lut.h",
797 "src/xnnpack/math.h",
798 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -0800799 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700800 "src/xnnpack/packx.h",
801 "src/xnnpack/pad.h",
802 "src/xnnpack/params.h",
803 "src/xnnpack/pavgpool.h",
804 "src/xnnpack/ppmm.h",
805 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700806 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700807 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700808 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700809 "src/xnnpack/rmax.h",
810 "src/xnnpack/scalar-utils.h",
811 "src/xnnpack/spmm.h",
812 "src/xnnpack/unpool.h",
813 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800814 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700815 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700816 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700817 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700818 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800819 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700820 "src/xnnpack/zip.h",
821]
822
823INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
824 "include/xnnpack.h",
825 "src/xnnpack/allocator.h",
826 "src/xnnpack/compute.h",
827 "src/xnnpack/im2col.h",
828 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700829 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700830 "src/xnnpack/operator.h",
831 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700832 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700833 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700834 "src/xnnpack/requantization.h",
835]
836
837ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
838 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700839]
840
841MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700842 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700843 "include/xnnpack.h",
844]
845
846MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
847 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700848 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700849 "src/xnnpack/requantization.h",
850 "include/xnnpack.h",
851]
852
853OPERATOR_TEST_PARAMS_HDRS = [
854 "src/xnnpack/params.h",
855 "src/xnnpack/common.h",
856]
857
858WEIGHTS_PACK_HDRS = [
859 "src/xnnpack/pack.h",
860 "src/xnnpack/operator.h",
861 "src/xnnpack/compute.h",
862]
863
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700864LOGGING_COPTS = select({
865 # No logging in optimized mode
866 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
867 # Full logging in debug mode
868 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
869 # Error-only logging in default (fastbuild) mode
870 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
871})
872
873LOGGING_HDRS = [
874 "src/xnnpack/log.h",
875]
876
Marat Dukhan08c4a432019-10-03 09:29:21 -0700877xnnpack_cc_library(
878 name = "scalar_ukernels",
879 srcs = SCALAR_UKERNELS,
880 hdrs = INTERNAL_HDRS,
881 aarch32_copts = ["-marm"],
882 copts = xnnpack_std_copts(),
883 deps = [
884 "@FP16",
885 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -0800886 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700887 ],
888)
889
890xnnpack_cc_library(
891 name = "psimd_ukernels",
892 srcs = PSIMD_UKERNELS,
893 hdrs = INTERNAL_HDRS,
894 aarch32_copts = [
895 "-marm",
896 "-mfpu=neon",
897 ],
898 copts = xnnpack_std_copts(),
899 optimized_copts = [
900 "-O3",
901 "-ffast-math",
902 ],
903 deps = [
904 "@FP16",
905 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -0800906 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700907 ],
908)
909
910xnnpack_cc_library(
911 name = "neon_ukernels",
912 hdrs = INTERNAL_HDRS,
913 aarch32_copts = [
914 "-marm",
915 "-mfpu=neon",
916 ],
917 aarch32_srcs = NEON_UKERNELS,
918 aarch64_srcs = NEON_UKERNELS,
919 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -0800920 deps = [
921 "@FP16",
922 "@pthreadpool",
923 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700924)
925
926xnnpack_cc_library(
927 name = "neonfma_ukernels",
928 hdrs = INTERNAL_HDRS,
929 aarch32_copts = [
930 "-marm",
931 "-mfpu=neon-vfpv4",
932 ],
933 aarch32_srcs = NEONFMA_UKERNELS,
934 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
935 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -0800936 deps = [
937 "@FP16",
938 "@pthreadpool",
939 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700940)
941
942xnnpack_cc_library(
943 name = "neonfp16arith_ukernels",
944 hdrs = INTERNAL_HDRS,
945 aarch64_copts = ["-march=armv8.2-a+fp16"],
946 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
947 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -0800948 deps = [
949 "@FP16",
950 "@pthreadpool",
951 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700952)
953
954xnnpack_cc_library(
955 name = "sse2_ukernels",
956 hdrs = INTERNAL_HDRS,
957 copts = xnnpack_std_copts(),
958 x86_copts = ["-msse2"],
959 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -0800960 deps = [
961 "@FP16",
962 "@pthreadpool",
963 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700964)
965
966xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800967 name = "sse41_ukernels",
968 hdrs = INTERNAL_HDRS,
969 copts = xnnpack_std_copts(),
970 x86_copts = ["-msse4.1"],
971 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -0800972 deps = [
973 "@FP16",
974 "@pthreadpool",
975 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800976)
977
978xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700979 name = "avx_ukernels",
980 hdrs = INTERNAL_HDRS,
981 copts = xnnpack_std_copts(),
982 x86_copts = ["-mavx"],
983 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -0800984 deps = [
985 "@FP16",
986 "@pthreadpool",
987 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700988)
989
990xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -0800991 name = "fma3_ukernels",
992 hdrs = INTERNAL_HDRS,
993 copts = xnnpack_std_copts(),
994 x86_copts = [
995 "-mfma",
996 ],
997 x86_srcs = FMA3_UKERNELS,
998 deps = [
999 "@FP16",
1000 "@pthreadpool",
1001 ],
1002)
1003
1004xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001005 name = "avx2_ukernels",
1006 hdrs = INTERNAL_HDRS,
1007 copts = xnnpack_std_copts(),
1008 x86_copts = [
1009 "-mfma",
1010 "-mavx2",
1011 ],
1012 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001013 deps = [
1014 "@FP16",
1015 "@pthreadpool",
1016 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001017)
1018
1019xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001020 name = "avx512f_ukernels",
1021 hdrs = INTERNAL_HDRS,
1022 copts = xnnpack_std_copts(),
1023 x86_copts = ["-mavx512f"],
1024 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001025 deps = [
1026 "@FP16",
1027 "@pthreadpool",
1028 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001029)
1030
1031xnnpack_cc_library(
1032 name = "asm_ukernels",
1033 hdrs = ["src/xnnpack/assembly.h"],
1034 aarch32_srcs = AARCH32_ASM_UKERNELS,
1035 aarch64_srcs = AARCH64_ASM_UKERNELS,
1036)
1037
1038xnnpack_aggregate_library(
1039 name = "ukernels",
1040 aarch32_deps = [
1041 ":psimd_ukernels",
1042 ":neon_ukernels",
1043 ":neonfma_ukernels",
1044 ":asm_ukernels",
1045 ],
1046 aarch64_deps = [
1047 ":psimd_ukernels",
1048 ":neon_ukernels",
1049 ":neonfma_ukernels",
1050 ":neonfp16arith_ukernels",
1051 ":asm_ukernels",
1052 ],
1053 generic_deps = [":scalar_ukernels"],
1054 wasmsimd_deps = [
1055 ":psimd_ukernels",
1056 ],
1057 x86_deps = [
1058 ":psimd_ukernels",
1059 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001060 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001061 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001062 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001063 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001064 ":avx512f_ukernels",
1065 ],
1066)
1067
1068xnnpack_cc_library(
1069 name = "im2col",
1070 srcs = ["src/im2col.c"],
1071 hdrs = [
1072 "src/xnnpack/common.h",
1073 "src/xnnpack/im2col.h",
1074 ],
1075 copts = xnnpack_std_copts(),
1076)
1077
1078xnnpack_cc_library(
1079 name = "indirection",
1080 srcs = ["src/indirection.c"],
1081 hdrs = INTERNAL_HDRS,
1082 copts = xnnpack_std_copts(),
1083 deps = [
1084 "@FP16",
1085 "@FXdiv",
1086 "@pthreadpool",
1087 ],
1088)
1089
1090xnnpack_cc_library(
1091 name = "operator_run",
1092 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001093 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
1094 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001095 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
1096 "-Wno-vla",
1097 ],
1098 deps = [
1099 "@FP16",
1100 "@FXdiv",
1101 "@clog",
1102 "@pthreadpool",
1103 ],
1104)
1105
1106cc_library(
1107 name = "enable_assembly",
1108 defines = select({
1109 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1110 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001111 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001112 }),
1113)
1114
Marat Dukhancf056b22019-10-07 10:26:29 -07001115xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001116 name = "operators",
1117 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001118 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001119 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07001120 ],
1121 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001122 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001123 "-Isrc",
1124 "-Iinclude",
1125 ] + select({
1126 ":debug_build": [],
1127 "//conditions:default": xnnpack_min_size_copts(),
1128 }),
Marat Dukhancf056b22019-10-07 10:26:29 -07001129 wasm_srcs = ["src/wasm-stubs.c"],
1130 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001131 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001132 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001133 "@FP16",
1134 "@FXdiv",
1135 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001136 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001137 ],
1138)
1139
1140cc_library(
1141 name = "XNNPACK",
1142 srcs = [
1143 "src/init.c",
1144 ],
1145 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1146 "-Isrc",
1147 "-Iinclude",
1148 ] + select({
1149 ":debug_build": [],
1150 "//conditions:default": xnnpack_min_size_copts(),
1151 }),
1152 includes = ["include"],
1153 linkstatic = True,
1154 textual_hdrs = ["include/xnnpack.h"],
1155 visibility = xnnpack_visibility(),
1156 deps = [
1157 ":enable_assembly",
1158 ":ukernels",
1159 ":operator_run",
1160 ":operators",
1161 "@clog",
1162 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07001163 ] + select({
1164 ":emscripten": [],
1165 "//conditions:default": ["@cpuinfo"],
1166 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001167)
1168
1169cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001170 name = "xnnpack_operators_nhwc_f32",
1171 srcs = [
1172 "src/init.c",
1173 ],
1174 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1175 "-Isrc",
1176 "-Iinclude",
1177 ] + select({
1178 ":debug_build": [],
1179 "//conditions:default": xnnpack_min_size_copts(),
1180 }),
1181 defines = [
1182 "XNN_NO_Q8_OPERATORS",
1183 "XNN_NO_U8_OPERATORS",
1184 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08001185 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001186 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001187 includes = ["include"],
1188 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001189 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001190 visibility = xnnpack_visibility(),
1191 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001192 ":enable_assembly",
1193 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001194 ":operator_run",
1195 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001196 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001197 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001198 ] + select({
1199 ":emscripten": [],
1200 "//conditions:default": ["@cpuinfo"],
1201 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001202)
1203
Marat Dukhancf056b22019-10-07 10:26:29 -07001204xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001205 name = "bench_utils",
1206 srcs = ["bench/utils.cc"],
1207 hdrs = ["bench/utils.h"],
1208 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08001209 deps = [
1210 "@com_google_benchmark//:benchmark",
1211 "@cpuinfo",
1212 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001213)
1214
Frank Barchard7e955972019-10-11 10:34:25 -07001215######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07001216
1217xnnpack_benchmark(
1218 name = "q8_gemm_bench",
1219 srcs = [
1220 "bench/gemm.h",
1221 "bench/q8-gemm.cc",
1222 "src/xnnpack/AlignedAllocator.h",
1223 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1224 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
1225 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
1226)
1227
1228xnnpack_benchmark(
1229 name = "f16_gemm_bench",
1230 srcs = [
1231 "bench/f16-gemm.cc",
1232 "bench/gemm.h",
1233 "src/xnnpack/AlignedAllocator.h",
1234 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1235 copts = ["-Wno-unused-function"],
1236 deps = MICROKERNEL_BENCHMARK_DEPS,
1237)
1238
1239xnnpack_benchmark(
1240 name = "f32_igemm_bench",
1241 srcs = [
1242 "bench/f32-igemm.cc",
1243 "bench/conv.h",
1244 "src/xnnpack/AlignedAllocator.h",
1245 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07001246 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001247)
1248
1249xnnpack_benchmark(
1250 name = "f32_conv_hwc_bench",
1251 srcs = [
1252 "bench/f32-conv-hwc.cc",
1253 "bench/dconv.h",
1254 "src/xnnpack/AlignedAllocator.h",
1255 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1256 copts = ["-Wno-unused-function"],
1257 deps = MICROKERNEL_BENCHMARK_DEPS,
1258)
1259
1260xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07001261 name = "f32_conv_hwc2spchw_bench",
1262 srcs = [
1263 "bench/f32-conv-hwc2spchw.cc",
1264 "bench/dconv.h",
1265 "src/xnnpack/AlignedAllocator.h",
1266 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1267 copts = ["-Wno-unused-function"],
1268 deps = MICROKERNEL_BENCHMARK_DEPS,
1269)
1270
1271xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001272 name = "f32_dwconv_bench",
1273 srcs = [
1274 "bench/f32-dwconv.cc",
1275 "bench/dwconv.h",
1276 "src/xnnpack/AlignedAllocator.h",
1277 ] + MICROKERNEL_BENCHMARK_HDRS,
1278 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1279)
1280
1281xnnpack_benchmark(
1282 name = "f32_dwconv_spchw_bench",
1283 srcs = [
1284 "bench/f32-dwconv-spchw.cc",
1285 "bench/dwconv.h",
1286 "src/xnnpack/AlignedAllocator.h",
1287 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1288 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1289)
1290
1291xnnpack_benchmark(
1292 name = "f32_gemm_bench",
1293 srcs = [
1294 "bench/f32-gemm.cc",
1295 "bench/gemm.h",
1296 "src/xnnpack/AlignedAllocator.h",
1297 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1298 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07001299 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001300)
1301
1302xnnpack_benchmark(
1303 name = "f32_rmax_bench",
1304 srcs = [
1305 "bench/f32-rmax.cc",
1306 "src/xnnpack/AlignedAllocator.h",
1307 ] + MICROKERNEL_BENCHMARK_HDRS,
1308 deps = MICROKERNEL_BENCHMARK_DEPS,
1309)
1310
1311xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08001312 name = "f32_sigmoid_bench",
1313 srcs = [
1314 "bench/f32-sigmoid.cc",
1315 "src/xnnpack/AlignedAllocator.h",
1316 ] + MICROKERNEL_BENCHMARK_HDRS,
1317 copts = ["-Wno-unused-function"],
1318 deps = MICROKERNEL_BENCHMARK_DEPS,
1319)
1320
1321xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001322 name = "f32_spmm_bench",
1323 srcs = [
1324 "bench/f32-spmm.cc",
1325 "bench/gemm.h",
1326 "src/xnnpack/AlignedAllocator.h",
1327 ] + MICROKERNEL_BENCHMARK_HDRS,
1328 copts = ["-Wno-unused-function"],
1329 deps = MICROKERNEL_BENCHMARK_DEPS,
1330)
1331
1332xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07001333 name = "f32_softargmax_bench",
1334 srcs = [
1335 "bench/f32-softargmax.cc",
1336 ] + MICROKERNEL_BENCHMARK_HDRS,
1337 copts = ["-Wno-unused-function"],
1338 deps = MICROKERNEL_BENCHMARK_DEPS,
1339)
1340
1341xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001342 name = "f32_im2col_gemm_bench",
1343 srcs = [
1344 "bench/f32-im2col-gemm.cc",
1345 "bench/conv.h",
1346 "src/xnnpack/AlignedAllocator.h",
1347 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1348 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
1349)
1350
1351########################### Benchmarks for operators ###########################
1352
1353xnnpack_benchmark(
1354 name = "add_bench",
1355 srcs = ["bench/add.cc"],
1356 deps = OPERATOR_BENCHMARK_DEPS,
1357)
1358
1359xnnpack_benchmark(
1360 name = "average_pooling_bench",
1361 srcs = ["bench/average-pooling.cc"],
1362 deps = OPERATOR_BENCHMARK_DEPS,
1363)
1364
1365xnnpack_benchmark(
1366 name = "channel_shuffle_bench",
1367 srcs = ["bench/channel-shuffle.cc"],
1368 deps = OPERATOR_BENCHMARK_DEPS,
1369)
1370
1371xnnpack_benchmark(
1372 name = "convolution_bench",
1373 srcs = ["bench/convolution.cc"],
1374 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1375 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1376)
1377
1378xnnpack_benchmark(
1379 name = "deconvolution_bench",
1380 srcs = ["bench/deconvolution.cc"],
1381 copts = xnnpack_optional_tflite_copts(),
1382 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1383)
1384
1385xnnpack_benchmark(
1386 name = "global_average_pooling_bench",
1387 srcs = ["bench/global-average-pooling.cc"],
1388 deps = OPERATOR_BENCHMARK_DEPS,
1389)
1390
1391xnnpack_benchmark(
1392 name = "max_pooling_bench",
1393 srcs = ["bench/max-pooling.cc"],
1394 deps = OPERATOR_BENCHMARK_DEPS,
1395)
1396
1397xnnpack_benchmark(
1398 name = "sigmoid_bench",
1399 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08001400 copts = xnnpack_optional_tflite_copts(),
1401 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001402)
1403
1404xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07001405 name = "prelu_bench",
1406 srcs = ["bench/prelu.cc"],
1407 copts = xnnpack_optional_tflite_copts(),
1408 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1409)
1410
1411xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001412 name = "softargmax_bench",
1413 srcs = ["bench/softargmax.cc"],
1414 deps = OPERATOR_BENCHMARK_DEPS,
1415)
1416
Marat Dukhanc068bb62019-10-04 13:24:39 -07001417############################# End-to-end benchmarks ############################
1418
1419cc_library(
1420 name = "mobilenet_v1",
1421 srcs = ["models/mobilenet-v1.cc"],
1422 hdrs = ["models/models.h"],
1423 linkstatic = True,
1424 deps = [
1425 ":XNNPACK",
1426 "@pthreadpool",
1427 ],
1428)
1429
1430cc_library(
1431 name = "mobilenet_v2",
1432 srcs = ["models/mobilenet-v2.cc"],
1433 hdrs = ["models/models.h"],
1434 linkstatic = True,
1435 deps = [
1436 ":XNNPACK",
1437 "@pthreadpool",
1438 ],
1439)
1440
1441xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07001442 name = "f32_dwconv_e2e_bench",
1443 srcs = ["bench/f32-dwconv-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1444 copts = ["-Wno-unused-function"],
1445 deps = MICROKERNEL_BENCHMARK_DEPS + [
1446 ":XNNPACK",
1447 ":mobilenet_v1",
1448 ":mobilenet_v2",
1449 ],
1450)
1451
1452xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07001453 name = "f32_gemm_e2e_bench",
1454 srcs = ["bench/f32-gemm-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1455 copts = ["-Wno-unused-function"],
1456 deps = MICROKERNEL_BENCHMARK_DEPS + [
1457 ":XNNPACK",
1458 ":mobilenet_v1",
1459 ":mobilenet_v2",
1460 ],
1461)
1462
1463xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07001464 name = "end2end_bench",
1465 srcs = ["bench/end2end.cc"],
1466 deps = [
1467 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07001468 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07001469 ":mobilenet_v1",
1470 ":mobilenet_v2",
1471 "@pthreadpool",
1472 ],
1473)
1474
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001475#################### Accuracy evaluation for math functions ####################
1476
1477xnnpack_benchmark(
1478 name = "f32_exp_eval",
1479 srcs = [
1480 "eval/f32-exp.cc",
1481 "src/xnnpack/AlignedAllocator.h",
1482 ] + ACCURACY_EVAL_HDRS,
1483 deps = ACCURACY_EVAL_DEPS,
1484)
1485
Marat Dukhan515c9772019-10-17 18:07:57 -07001486xnnpack_benchmark(
1487 name = "f32_expminus_eval",
1488 srcs = [
1489 "eval/f32-expminus.cc",
1490 "src/xnnpack/AlignedAllocator.h",
1491 ] + ACCURACY_EVAL_HDRS,
1492 deps = ACCURACY_EVAL_DEPS,
1493)
1494
Marat Dukhan98ba4412019-10-23 02:14:28 -07001495xnnpack_benchmark(
1496 name = "f32_extexp_eval",
1497 srcs = [
1498 "eval/f32-extexp.cc",
1499 "src/xnnpack/AlignedAllocator.h",
1500 ] + ACCURACY_EVAL_HDRS,
1501 deps = ACCURACY_EVAL_DEPS,
1502)
1503
Marat Dukhan346a9e52019-11-15 09:06:30 -08001504xnnpack_benchmark(
1505 name = "f32_sigmoid_eval",
1506 srcs = [
1507 "eval/f32-sigmoid.cc",
1508 "src/xnnpack/AlignedAllocator.h",
1509 ] + ACCURACY_EVAL_HDRS,
1510 deps = ACCURACY_EVAL_DEPS,
1511)
1512
Marat Dukhan08c4a432019-10-03 09:29:21 -07001513######################### Unit tests for micro-kernels #########################
1514
1515xnnpack_unit_test(
1516 name = "f16_gemm_test",
1517 srcs = [
1518 "test/f16-gemm.cc",
1519 "test/gemm-microkernel-tester.h",
1520 "src/xnnpack/AlignedAllocator.h",
1521 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1522 deps = MICROKERNEL_TEST_DEPS,
1523)
1524
1525xnnpack_unit_test(
1526 name = "f32_argmaxpool_test",
1527 srcs = [
1528 "test/f32-argmaxpool.cc",
1529 "test/argmaxpool-microkernel-tester.h",
1530 "src/xnnpack/AlignedAllocator.h",
1531 ] + MICROKERNEL_TEST_HDRS,
1532 deps = MICROKERNEL_TEST_DEPS,
1533)
1534
1535xnnpack_unit_test(
1536 name = "f32_avgpool_test",
1537 srcs = [
1538 "test/f32-avgpool.cc",
1539 "test/avgpool-microkernel-tester.h",
1540 "src/xnnpack/AlignedAllocator.h",
1541 ] + MICROKERNEL_TEST_HDRS,
1542 deps = MICROKERNEL_TEST_DEPS,
1543)
1544
1545xnnpack_unit_test(
Marat Dukhan35dacfb2019-11-07 19:18:16 -08001546 name = "f32_bilinear_test",
1547 srcs = [
1548 "test/f32-bilinear.cc",
1549 "test/bilinear-microkernel-tester.h",
1550 "src/xnnpack/AlignedAllocator.h",
1551 ] + MICROKERNEL_TEST_HDRS,
1552 deps = MICROKERNEL_TEST_DEPS,
1553)
1554
1555xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001556 name = "f32_clamp_test",
1557 srcs = [
1558 "test/f32-clamp.cc",
1559 "test/clamp-microkernel-tester.h",
1560 ] + MICROKERNEL_TEST_HDRS,
1561 deps = MICROKERNEL_TEST_DEPS,
1562)
1563
1564xnnpack_unit_test(
1565 name = "f32_igemm_test",
1566 srcs = [
1567 "test/f32-igemm.cc",
1568 "test/gemm-microkernel-tester.h",
1569 "src/xnnpack/AlignedAllocator.h",
1570 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1571 deps = MICROKERNEL_TEST_DEPS,
1572)
1573
1574xnnpack_unit_test(
1575 name = "f32_conv_hwc_test",
1576 srcs = [
1577 "test/f32-conv-hwc.cc",
1578 "test/conv-hwc-microkernel-tester.h",
1579 "src/xnnpack/AlignedAllocator.h",
1580 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1581 deps = MICROKERNEL_TEST_DEPS,
1582)
1583
1584xnnpack_unit_test(
1585 name = "f32_conv_hwc2spchw_test",
1586 srcs = [
1587 "test/f32-conv-hwc2spchw.cc",
1588 "test/conv-hwc2spchw-microkernel-tester.h",
1589 "src/xnnpack/AlignedAllocator.h",
1590 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1591 deps = MICROKERNEL_TEST_DEPS,
1592)
1593
1594xnnpack_unit_test(
1595 name = "f32_dwconv_test",
1596 srcs = [
1597 "test/f32-dwconv.cc",
1598 "test/dwconv-microkernel-tester.h",
1599 "src/xnnpack/AlignedAllocator.h",
1600 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1601 deps = MICROKERNEL_TEST_DEPS,
1602)
1603
1604xnnpack_unit_test(
1605 name = "f32_dwconv_spchw_test",
1606 srcs = [
1607 "test/f32-dwconv-spchw.cc",
1608 "test/dwconv-spchw-microkernel-tester.h",
1609 "src/xnnpack/AlignedAllocator.h",
1610 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1611 deps = MICROKERNEL_TEST_DEPS,
1612)
1613
1614xnnpack_unit_test(
1615 name = "f32_gavgpool_test",
1616 srcs = [
1617 "test/f32-gavgpool.cc",
1618 "test/gavgpool-microkernel-tester.h",
1619 "src/xnnpack/AlignedAllocator.h",
1620 ] + MICROKERNEL_TEST_HDRS,
1621 deps = MICROKERNEL_TEST_DEPS,
1622)
1623
1624xnnpack_unit_test(
1625 name = "f32_gavgpool_spchw_test",
1626 srcs = [
1627 "test/f32-gavgpool-spchw.cc",
1628 "test/gavgpool-spchw-microkernel-tester.h",
1629 "src/xnnpack/AlignedAllocator.h",
1630 ] + MICROKERNEL_TEST_HDRS,
1631 deps = MICROKERNEL_TEST_DEPS,
1632)
1633
1634xnnpack_unit_test(
1635 name = "f32_gemm_test",
1636 srcs = [
1637 "test/f32-gemm.cc",
1638 "test/gemm-microkernel-tester.h",
1639 "src/xnnpack/AlignedAllocator.h",
1640 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1641 deps = MICROKERNEL_TEST_DEPS,
1642)
1643
1644xnnpack_unit_test(
1645 name = "f32_gemminc_test",
1646 srcs = [
1647 "test/f32-gemminc.cc",
1648 "test/gemm-microkernel-tester.h",
1649 "src/xnnpack/AlignedAllocator.h",
1650 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1651 deps = MICROKERNEL_TEST_DEPS,
1652)
1653
1654xnnpack_unit_test(
1655 name = "f32_hswish_test",
1656 srcs = [
1657 "test/f32-hswish.cc",
1658 "test/hswish-microkernel-tester.h",
1659 ] + MICROKERNEL_TEST_HDRS,
1660 deps = MICROKERNEL_TEST_DEPS,
1661)
1662
1663xnnpack_unit_test(
1664 name = "f32_maxpool_test",
1665 srcs = [
1666 "test/f32-maxpool.cc",
1667 "test/maxpool-microkernel-tester.h",
1668 ] + MICROKERNEL_TEST_HDRS,
1669 deps = MICROKERNEL_TEST_DEPS,
1670)
1671
1672xnnpack_unit_test(
1673 name = "f32_pavgpool_test",
1674 srcs = [
1675 "test/f32-pavgpool.cc",
1676 "test/avgpool-microkernel-tester.h",
1677 "src/xnnpack/AlignedAllocator.h",
1678 ] + MICROKERNEL_TEST_HDRS,
1679 deps = MICROKERNEL_TEST_DEPS,
1680)
1681
1682xnnpack_unit_test(
1683 name = "f32_ppmm_test",
1684 srcs = [
1685 "test/f32-ppmm.cc",
1686 "test/gemm-microkernel-tester.h",
1687 "src/xnnpack/AlignedAllocator.h",
1688 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1689 deps = MICROKERNEL_TEST_DEPS,
1690)
1691
1692xnnpack_unit_test(
1693 name = "f32_prelu_test",
1694 srcs = [
1695 "test/f32-prelu.cc",
1696 "test/prelu-microkernel-tester.h",
1697 "src/xnnpack/AlignedAllocator.h",
1698 ] + MICROKERNEL_TEST_HDRS,
1699 deps = MICROKERNEL_TEST_DEPS,
1700)
1701
1702xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001703 name = "f32_raddexpminusmax_test",
1704 srcs = [
1705 "test/f32-raddexpminusmax.cc",
1706 "test/raddexpminusmax-microkernel-tester.h",
1707 ] + MICROKERNEL_TEST_HDRS,
1708 deps = MICROKERNEL_TEST_DEPS,
1709)
1710
1711xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001712 name = "f32_raddextexp_test",
1713 srcs = [
1714 "test/f32-raddextexp.cc",
1715 "test/raddextexp-microkernel-tester.h",
1716 ] + MICROKERNEL_TEST_HDRS,
1717 deps = MICROKERNEL_TEST_DEPS,
1718)
1719
1720xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001721 name = "f32_raddstoreexpminusmax_test",
1722 srcs = [
1723 "test/f32-raddstoreexpminusmax.cc",
1724 "test/raddstoreexpminusmax-microkernel-tester.h",
1725 ] + MICROKERNEL_TEST_HDRS,
1726 deps = MICROKERNEL_TEST_DEPS,
1727)
1728
1729xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001730 name = "f32_rmax_test",
1731 srcs = [
1732 "test/f32-rmax.cc",
1733 "test/rmax-microkernel-tester.h",
1734 ] + MICROKERNEL_TEST_HDRS,
1735 deps = MICROKERNEL_TEST_DEPS,
1736)
1737
1738xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08001739 name = "f32_sigmoid_test",
1740 srcs = [
1741 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001742 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001743 ] + MICROKERNEL_TEST_HDRS,
1744 deps = MICROKERNEL_TEST_DEPS,
1745)
1746
1747xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001748 name = "f32_spmm_test",
1749 srcs = [
1750 "test/f32-spmm.cc",
1751 "test/spmm-microkernel-tester.h",
1752 "src/xnnpack/AlignedAllocator.h",
1753 ] + MICROKERNEL_TEST_HDRS,
1754 deps = MICROKERNEL_TEST_DEPS,
1755)
1756
1757xnnpack_unit_test(
1758 name = "f32_vadd_test",
1759 srcs = [
1760 "test/f32-vadd.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001761 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001762 ] + MICROKERNEL_TEST_HDRS,
1763 deps = MICROKERNEL_TEST_DEPS,
1764)
1765
1766xnnpack_unit_test(
1767 name = "f32_vaddc_test",
1768 srcs = [
1769 "test/f32-vaddc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001770 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001771 ] + MICROKERNEL_TEST_HDRS,
1772 deps = MICROKERNEL_TEST_DEPS,
1773)
1774
1775xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001776 name = "f32_vmul_test",
1777 srcs = [
1778 "test/f32-vmul.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001779 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001780 ] + MICROKERNEL_TEST_HDRS,
1781 deps = MICROKERNEL_TEST_DEPS,
1782)
1783
1784xnnpack_unit_test(
1785 name = "f32_vmulc_test",
1786 srcs = [
1787 "test/f32-vmulc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001788 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001789 ] + MICROKERNEL_TEST_HDRS,
1790 deps = MICROKERNEL_TEST_DEPS,
1791)
1792
1793xnnpack_unit_test(
1794 name = "f32_vmulcaddc_test",
1795 srcs = [
1796 "test/f32-vmulcaddc.cc",
1797 "test/vmulcaddc-microkernel-tester.h",
1798 "src/xnnpack/AlignedAllocator.h",
1799 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1800 deps = MICROKERNEL_TEST_DEPS,
1801)
1802
1803xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001804 name = "f32_vscale_test",
1805 srcs = [
1806 "test/f32-vscale.cc",
1807 "test/vscale-microkernel-tester.h",
1808 ] + MICROKERNEL_TEST_HDRS,
1809 deps = MICROKERNEL_TEST_DEPS,
1810)
1811
1812xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001813 name = "f32_vscaleexpminusmax_test",
1814 srcs = [
1815 "test/f32-vscaleexpminusmax.cc",
1816 "test/vscaleexpminusmax-microkernel-tester.h",
1817 ] + MICROKERNEL_TEST_HDRS,
1818 deps = MICROKERNEL_TEST_DEPS,
1819)
1820
1821xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001822 name = "f32_vscaleextexp_test",
1823 srcs = [
1824 "test/f32-vscaleextexp.cc",
1825 "test/vscaleextexp-microkernel-tester.h",
1826 ] + MICROKERNEL_TEST_HDRS,
1827 deps = MICROKERNEL_TEST_DEPS,
1828)
1829
1830xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001831 name = "f32_vsub_test",
1832 srcs = [
1833 "test/f32-vsub.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001834 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001835 ] + MICROKERNEL_TEST_HDRS,
1836 deps = MICROKERNEL_TEST_DEPS,
1837)
1838
1839xnnpack_unit_test(
1840 name = "f32_vsubc_test",
1841 srcs = [
1842 "test/f32-vsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001843 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001844 ] + MICROKERNEL_TEST_HDRS,
1845 deps = MICROKERNEL_TEST_DEPS,
1846)
1847
1848xnnpack_unit_test(
1849 name = "f32_vrsubc_test",
1850 srcs = [
1851 "test/f32-vrsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001852 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001853 ] + MICROKERNEL_TEST_HDRS,
1854 deps = MICROKERNEL_TEST_DEPS,
1855)
1856
1857xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001858 name = "q8_avgpool_test",
1859 srcs = [
1860 "test/q8-avgpool.cc",
1861 "test/avgpool-microkernel-tester.h",
1862 "src/xnnpack/AlignedAllocator.h",
1863 ] + MICROKERNEL_TEST_HDRS,
1864 deps = MICROKERNEL_TEST_DEPS,
1865)
1866
1867xnnpack_unit_test(
1868 name = "q8_igemm_test",
1869 srcs = [
1870 "test/q8-igemm.cc",
1871 "test/gemm-microkernel-tester.h",
1872 "src/xnnpack/AlignedAllocator.h",
1873 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1874 deps = MICROKERNEL_TEST_DEPS,
1875)
1876
1877xnnpack_unit_test(
1878 name = "q8_dwconv_test",
1879 srcs = [
1880 "test/q8-dwconv.cc",
1881 "test/dwconv-microkernel-tester.h",
1882 "src/xnnpack/AlignedAllocator.h",
1883 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1884 deps = MICROKERNEL_TEST_DEPS,
1885)
1886
1887xnnpack_unit_test(
1888 name = "q8_gavgpool_test",
1889 srcs = [
1890 "test/q8-gavgpool.cc",
1891 "test/gavgpool-microkernel-tester.h",
1892 "src/xnnpack/AlignedAllocator.h",
1893 ] + MICROKERNEL_TEST_HDRS,
1894 deps = MICROKERNEL_TEST_DEPS,
1895)
1896
1897xnnpack_unit_test(
1898 name = "q8_gemm_test",
1899 srcs = [
1900 "test/q8-gemm.cc",
1901 "test/gemm-microkernel-tester.h",
1902 "src/xnnpack/AlignedAllocator.h",
1903 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1904 deps = MICROKERNEL_TEST_DEPS,
1905)
1906
1907xnnpack_unit_test(
1908 name = "q8_vadd_test",
1909 srcs = [
1910 "test/q8-vadd.cc",
1911 "test/vadd-microkernel-tester.h",
1912 ] + MICROKERNEL_TEST_HDRS,
1913 deps = MICROKERNEL_TEST_DEPS,
1914)
1915
1916xnnpack_unit_test(
1917 name = "u8_clamp_test",
1918 srcs = [
1919 "test/u8-clamp.cc",
1920 "test/clamp-microkernel-tester.h",
1921 ] + MICROKERNEL_TEST_HDRS,
1922 deps = MICROKERNEL_TEST_DEPS,
1923)
1924
1925xnnpack_unit_test(
1926 name = "u8_lut32norm_test",
1927 srcs = [
1928 "test/u8-lut32norm.cc",
1929 "test/lut-norm-microkernel-tester.h",
1930 ] + MICROKERNEL_TEST_HDRS,
1931 deps = MICROKERNEL_TEST_DEPS,
1932)
1933
1934xnnpack_unit_test(
1935 name = "u8_maxpool_test",
1936 srcs = [
1937 "test/u8-maxpool.cc",
1938 "test/maxpool-microkernel-tester.h",
1939 ] + MICROKERNEL_TEST_HDRS,
1940 deps = MICROKERNEL_TEST_DEPS,
1941)
1942
1943xnnpack_unit_test(
1944 name = "u8_rmax_test",
1945 srcs = [
1946 "test/u8-rmax.cc",
1947 "test/rmax-microkernel-tester.h",
1948 ] + MICROKERNEL_TEST_HDRS,
1949 deps = MICROKERNEL_TEST_DEPS,
1950)
1951
1952xnnpack_unit_test(
1953 name = "x32_packx_test",
1954 srcs = [
1955 "test/x32-packx.cc",
1956 "test/pack-microkernel-tester.h",
1957 "src/xnnpack/AlignedAllocator.h",
1958 ] + MICROKERNEL_TEST_HDRS,
1959 deps = MICROKERNEL_TEST_DEPS,
1960)
1961
1962xnnpack_unit_test(
1963 name = "x32_pad_test",
1964 srcs = [
1965 "test/x32-pad.cc",
1966 "test/pad-microkernel-tester.h",
1967 ] + MICROKERNEL_TEST_HDRS,
1968 deps = MICROKERNEL_TEST_DEPS,
1969)
1970
1971xnnpack_unit_test(
1972 name = "x32_unpool_test",
1973 srcs = [
1974 "test/x32-unpool.cc",
1975 "test/unpool-microkernel-tester.h",
1976 ] + MICROKERNEL_TEST_HDRS,
1977 deps = MICROKERNEL_TEST_DEPS,
1978)
1979
1980xnnpack_unit_test(
1981 name = "x32_zip_test",
1982 srcs = [
1983 "test/x32-zip.cc",
1984 "test/zip-microkernel-tester.h",
1985 ] + MICROKERNEL_TEST_HDRS,
1986 deps = MICROKERNEL_TEST_DEPS,
1987)
1988
1989xnnpack_unit_test(
1990 name = "x8_lut_test",
1991 srcs = [
1992 "test/x8-lut.cc",
1993 "test/lut-microkernel-tester.h",
1994 ] + MICROKERNEL_TEST_HDRS,
1995 deps = MICROKERNEL_TEST_DEPS,
1996)
1997
1998xnnpack_unit_test(
1999 name = "x8_zip_test",
2000 srcs = [
2001 "test/x8-zip.cc",
2002 "test/zip-microkernel-tester.h",
2003 ] + MICROKERNEL_TEST_HDRS,
2004 deps = MICROKERNEL_TEST_DEPS,
2005)
2006
2007########################### Size test for the library ##########################
2008
2009xnnpack_binary(
2010 name = "size_test",
2011 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002012 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002013)
2014
2015########################### Unit tests for operators ###########################
2016
2017xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002018 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002019 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002020 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002021 "test/add-operator-tester.h",
2022 ],
2023 deps = OPERATOR_TEST_DEPS,
2024)
2025
2026xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08002027 name = "add_nd_test",
2028 srcs = [
2029 "test/add-nd.cc",
2030 "test/binary-elementwise-operator-tester.h",
2031 ],
2032 deps = OPERATOR_TEST_DEPS,
2033)
2034
2035xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002036 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002037 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002038 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002039 "test/argmax-pooling-operator-tester.h",
2040 ] + OPERATOR_TEST_PARAMS_HDRS,
2041 deps = OPERATOR_TEST_DEPS,
2042)
2043
2044xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002045 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002046 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002047 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002048 "test/average-pooling-operator-tester.h",
2049 ] + OPERATOR_TEST_PARAMS_HDRS,
2050 deps = OPERATOR_TEST_DEPS,
2051)
2052
2053xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002054 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002055 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002056 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002057 "test/channel-pad-operator-tester.h",
2058 ] + OPERATOR_TEST_PARAMS_HDRS,
2059 deps = OPERATOR_TEST_DEPS,
2060)
2061
2062xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002063 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002064 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002065 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002066 "test/channel-shuffle-operator-tester.h",
2067 ],
2068 deps = OPERATOR_TEST_DEPS,
2069)
2070
2071xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002072 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002073 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002074 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002075 "test/clamp-operator-tester.h",
2076 ],
2077 deps = OPERATOR_TEST_DEPS,
2078)
2079
2080xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002081 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002082 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002083 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002084 "test/convolution-operator-tester.h",
2085 ],
2086 deps = OPERATOR_TEST_DEPS,
2087)
2088
2089xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002090 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002091 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002092 "test/convolution-nchw.cc",
2093 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002094 ],
2095 deps = OPERATOR_TEST_DEPS,
2096)
2097
2098xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002099 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002100 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002101 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002102 "test/deconvolution-operator-tester.h",
2103 ] + OPERATOR_TEST_PARAMS_HDRS,
2104 deps = OPERATOR_TEST_DEPS,
2105)
2106
2107xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002108 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002109 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002110 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002111 "test/fully-connected-operator-tester.h",
2112 ],
2113 deps = OPERATOR_TEST_DEPS,
2114)
2115
2116xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002117 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002118 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002119 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002120 "test/global-average-pooling-operator-tester.h",
2121 ] + OPERATOR_TEST_PARAMS_HDRS,
2122 deps = OPERATOR_TEST_DEPS,
2123)
2124
2125xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002126 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002127 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002128 "test/global-average-pooling-ncw.cc",
2129 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002130 ],
2131 deps = OPERATOR_TEST_DEPS,
2132)
2133
2134xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002135 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002136 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002137 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002138 "test/hardswish-operator-tester.h",
2139 ],
2140 deps = OPERATOR_TEST_DEPS,
2141)
2142
2143xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002144 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002145 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002146 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002147 "test/leaky-relu-operator-tester.h",
2148 ],
2149 deps = OPERATOR_TEST_DEPS,
2150)
2151
2152xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002153 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002154 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002155 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002156 "test/max-pooling-operator-tester.h",
2157 ] + OPERATOR_TEST_PARAMS_HDRS,
2158 deps = OPERATOR_TEST_DEPS,
2159)
2160
2161xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002162 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08002163 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08002164 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002165 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08002166 ],
2167 deps = OPERATOR_TEST_DEPS,
2168)
2169
2170xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002171 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002172 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002173 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002174 "test/prelu-operator-tester.h",
2175 ] + OPERATOR_TEST_PARAMS_HDRS,
2176 deps = OPERATOR_TEST_DEPS,
2177)
2178
2179xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002180 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08002181 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002182 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08002183 "test/resize-bilinear-operator-tester.h",
2184 ] + OPERATOR_TEST_PARAMS_HDRS,
2185 deps = OPERATOR_TEST_DEPS,
2186)
2187
2188xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002189 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002190 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002191 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002192 "test/sigmoid-operator-tester.h",
2193 ],
2194 deps = OPERATOR_TEST_DEPS,
2195)
2196
2197xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002198 name = "softargmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002199 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002200 "test/softargmax-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002201 "test/softargmax-operator-tester.h",
2202 ],
2203 deps = OPERATOR_TEST_DEPS,
2204)
2205
2206xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002207 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002208 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002209 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002210 "test/unpooling-operator-tester.h",
2211 ],
2212 deps = OPERATOR_TEST_DEPS,
2213)
2214
2215############################# Build configurations #############################
2216
Marat Dukhanb8642352019-10-30 15:43:02 -07002217# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07002218config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07002219 name = "xnn_enable_assembly_explicit_true",
2220 define_values = {"xnn_enable_assembly": "true"},
2221)
2222
2223# Disables usage of assembly kernels.
2224config_setting(
2225 name = "xnn_enable_assembly_explicit_false",
2226 define_values = {"xnn_enable_assembly": "false"},
2227)
2228
2229# Builds with -c dbg
2230config_setting(
2231 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002232 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07002233 "compilation_mode": "dbg",
2234 },
2235)
2236
2237# Builds with -c opt
2238config_setting(
2239 name = "optimized_build",
2240 values = {
2241 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002242 },
2243)
2244
2245config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07002246 name = "linux_k8",
2247 values = {"cpu": "k8"},
2248)
2249
2250config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07002251 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07002252 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07002253)
2254
2255config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002256 name = "android",
2257 values = {"crosstool_top": "//external:android/crosstool"},
2258)
2259
2260config_setting(
2261 name = "android_armv7",
2262 values = {
2263 "crosstool_top": "//external:android/crosstool",
2264 "cpu": "armeabi-v7a",
2265 },
2266)
2267
2268config_setting(
2269 name = "android_arm64",
2270 values = {
2271 "crosstool_top": "//external:android/crosstool",
2272 "cpu": "arm64-v8a",
2273 },
2274)
2275
2276config_setting(
2277 name = "android_x86",
2278 values = {
2279 "crosstool_top": "//external:android/crosstool",
2280 "cpu": "x86",
2281 },
2282)
2283
2284config_setting(
2285 name = "android_x86_64",
2286 values = {
2287 "crosstool_top": "//external:android/crosstool",
2288 "cpu": "x86_64",
2289 },
2290)
2291
2292config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07002293 name = "macos_x86_64",
2294 values = {
2295 "apple_platform_type": "macos",
2296 "cpu": "darwin",
2297 },
2298)
2299
2300config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002301 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002302 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07002303)
2304
2305config_setting(
2306 name = "emscripten_wasm",
2307 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002308 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002309 "cpu": "wasm",
2310 },
2311)
2312
2313config_setting(
2314 name = "emscripten_wasmsimd",
2315 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002316 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002317 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07002318 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002319 },
2320)
2321
2322config_setting(
2323 name = "emscripten_asmjs",
2324 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002325 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002326 "cpu": "asmjs",
2327 },
2328)