blob: 3c41cdebe93a5468a4fb297379190bd73a818fbf [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",
Marat Dukhan441e2212019-12-04 18:30:49 -080086 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -080087 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen563df5f2019-10-23 08:02:21 -070088 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-scalar-1x1.c",
Erich Elsen0cc2c532019-10-15 04:44:18 -070089 "src/f32-dwconv-spchw/3x3p1-scalar.c",
Erich Elsenac4de802019-10-16 04:35:30 -070090 "src/f32-dwconv-spchw/3x3s2p1-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -080091 "src/f32-dwconv-spchw/5x5p2-scalar.c",
Erich Elsen38709a62019-11-08 11:58:45 -080092 "src/f32-dwconv-spchw/5x5s2p2-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -080093 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
94 "src/f32-dwconv/gen/up1x25-scalar.c",
95 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
96 "src/f32-dwconv/gen/up1x4-scalar.c",
97 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
98 "src/f32-dwconv/gen/up1x9-scalar.c",
99 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
100 "src/f32-dwconv/gen/up2x25-scalar.c",
101 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
102 "src/f32-dwconv/gen/up2x4-scalar.c",
103 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
104 "src/f32-dwconv/gen/up2x9-scalar.c",
Erich Elsen34dc2c02019-10-16 05:11:41 -0700105 "src/f32-gavgpool-spchw/scalar-x1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700106 "src/f32-gavgpool/mp7p7q-scalar.c",
107 "src/f32-gavgpool/up7-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800108 "src/f32-gemm/gen-inc/1x4-scalar.c",
109 "src/f32-gemm/gen-inc/2x4-scalar.c",
110 "src/f32-gemm/gen-inc/4x4-scalar.c",
111 "src/f32-gemm/gen/1x4-scalar.c",
112 "src/f32-gemm/gen/2x4-scalar.c",
113 "src/f32-gemm/gen/4x2-scalar.c",
114 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700115 "src/f32-hswish/scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800116 "src/f32-igemm/gen/1x4-scalar.c",
117 "src/f32-igemm/gen/2x4-scalar.c",
118 "src/f32-igemm/gen/4x2-scalar.c",
119 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800120 "src/f32-maxpool/9p8x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700121 "src/f32-pavgpool/mp9p8q-scalar.c",
122 "src/f32-pavgpool/up9-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800123 "src/f32-ppmm/gen/2x4-scalar.c",
124 "src/f32-ppmm/gen/3x3-scalar.c",
125 "src/f32-ppmm/gen/4x2-scalar.c",
126 "src/f32-ppmm/gen/4x4-scalar.c",
127 "src/f32-prelu/gen/scalar-2x1.c",
128 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700129 "src/f32-rmax/scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800130 "src/f32-spmm/gen/1x1-scalar-pipelined.c",
131 "src/f32-spmm/gen/1x1-scalar.c",
132 "src/f32-spmm/gen/2x1-scalar-pipelined.c",
133 "src/f32-spmm/gen/2x1-scalar.c",
134 "src/f32-spmm/gen/4x1-scalar-pipelined.c",
135 "src/f32-spmm/gen/4x1-scalar.c",
136 "src/f32-spmm/gen/8x1-scalar-pipelined.c",
137 "src/f32-spmm/gen/8x1-scalar.c",
138 "src/f32-spmm/gen/8x2-scalar.c",
139 "src/f32-spmm/gen/8x4-scalar.c",
140 "src/f32-vbinary/gen/vadd-scalar-x1.c",
141 "src/f32-vbinary/gen/vadd-scalar-x2.c",
142 "src/f32-vbinary/gen/vadd-scalar-x4.c",
143 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
144 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
145 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
146 "src/f32-vbinary/gen/vmul-scalar-x1.c",
147 "src/f32-vbinary/gen/vmul-scalar-x2.c",
148 "src/f32-vbinary/gen/vmul-scalar-x4.c",
149 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
150 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
151 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
152 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
153 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
154 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
155 "src/f32-vbinary/gen/vsub-scalar-x1.c",
156 "src/f32-vbinary/gen/vsub-scalar-x2.c",
157 "src/f32-vbinary/gen/vsub-scalar-x4.c",
158 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
159 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
160 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
161 "src/f32-vmulcaddc/gen/c1-scalar-2x.c",
162 "src/f32-vmulcaddc/gen/c2-scalar-2x.c",
163 "src/f32-vmulcaddc/gen/c4-scalar-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700164 "src/q8-avgpool/mp9p8q-scalar.c",
165 "src/q8-avgpool/up9-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700166 "src/q8-dwconv/up1x9-scalar.c",
167 "src/q8-gavgpool/mp7p7q-scalar.c",
168 "src/q8-gavgpool/up7-scalar.c",
169 "src/q8-gemm/2x2-scalar.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800170 "src/q8-igemm/2x2-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700171 "src/q8-vadd/scalar.c",
172 "src/u8-clamp/scalar.c",
173 "src/u8-lut32norm/scalar.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800174 "src/u8-maxpool/9p8x-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700175 "src/u8-rmax/scalar.c",
176 "src/x32-packx/x2-scalar.c",
177 "src/x32-packx/x3-scalar.c",
178 "src/x32-packx/x4-scalar.c",
179 "src/x32-pad/x2-scalar.c",
180 "src/x32-unpool/scalar.c",
181 "src/x32-zip/x2-scalar.c",
182 "src/x32-zip/x3-scalar.c",
183 "src/x32-zip/x4-scalar.c",
184 "src/x32-zip/xm-scalar.c",
185 "src/x8-lut/scalar.c",
186 "src/x8-zip/x2-scalar.c",
187 "src/x8-zip/x3-scalar.c",
188 "src/x8-zip/x4-scalar.c",
189 "src/x8-zip/xm-scalar.c",
190]
191
Marat Dukhan436ebe62019-12-04 15:10:12 -0800192WASM_UKERNELS = [
193 "src/f32-avgpool/mp9p8q-wasm.c",
194 "src/f32-avgpool/up9-wasm.c",
195 "src/f32-clamp/wasm.c",
196 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
197 "src/f32-dwconv/gen/up1x25-wasm.c",
198 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
199 "src/f32-dwconv/gen/up1x4-wasm.c",
200 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
201 "src/f32-dwconv/gen/up1x9-wasm.c",
202 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
203 "src/f32-dwconv/gen/up2x25-wasm.c",
204 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
205 "src/f32-dwconv/gen/up2x4-wasm.c",
206 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
207 "src/f32-dwconv/gen/up2x9-wasm.c",
208 "src/f32-gavgpool/mp7p7q-wasm.c",
209 "src/f32-gavgpool/up7-wasm.c",
210 "src/f32-gemm/gen-inc/1x4-wasm.c",
211 "src/f32-gemm/gen-inc/2x4-wasm.c",
212 "src/f32-gemm/gen-inc/4x4-wasm.c",
213 "src/f32-gemm/gen/1x4-wasm.c",
214 "src/f32-gemm/gen/2x4-wasm.c",
215 "src/f32-gemm/gen/4x2-wasm.c",
216 "src/f32-gemm/gen/4x4-wasm.c",
217 "src/f32-hswish/wasm.c",
218 "src/f32-igemm/gen/1x4-wasm.c",
219 "src/f32-igemm/gen/2x4-wasm.c",
220 "src/f32-igemm/gen/4x2-wasm.c",
221 "src/f32-igemm/gen/4x4-wasm.c",
222 "src/f32-maxpool/9p8x-wasm-c1.c",
223 "src/f32-pavgpool/mp9p8q-wasm.c",
224 "src/f32-pavgpool/up9-wasm.c",
225 "src/f32-prelu/gen/wasm-2x1.c",
226 "src/f32-prelu/gen/wasm-2x4.c",
227 "src/f32-vbinary/gen/vadd-wasm-x1.c",
228 "src/f32-vbinary/gen/vadd-wasm-x2.c",
229 "src/f32-vbinary/gen/vadd-wasm-x4.c",
230 "src/f32-vbinary/gen/vaddc-wasm-x1.c",
231 "src/f32-vbinary/gen/vaddc-wasm-x2.c",
232 "src/f32-vbinary/gen/vaddc-wasm-x4.c",
233 "src/f32-vbinary/gen/vmul-wasm-x1.c",
234 "src/f32-vbinary/gen/vmul-wasm-x2.c",
235 "src/f32-vbinary/gen/vmul-wasm-x4.c",
236 "src/f32-vbinary/gen/vmulc-wasm-x1.c",
237 "src/f32-vbinary/gen/vmulc-wasm-x2.c",
238 "src/f32-vbinary/gen/vmulc-wasm-x4.c",
239 "src/f32-vbinary/gen/vrsubc-wasm-x1.c",
240 "src/f32-vbinary/gen/vrsubc-wasm-x2.c",
241 "src/f32-vbinary/gen/vrsubc-wasm-x4.c",
242 "src/f32-vbinary/gen/vsub-wasm-x1.c",
243 "src/f32-vbinary/gen/vsub-wasm-x2.c",
244 "src/f32-vbinary/gen/vsub-wasm-x4.c",
245 "src/f32-vbinary/gen/vsubc-wasm-x1.c",
246 "src/f32-vbinary/gen/vsubc-wasm-x2.c",
247 "src/f32-vbinary/gen/vsubc-wasm-x4.c",
248 "src/f32-vmulcaddc/gen/c1-wasm-2x.c",
249 "src/f32-vmulcaddc/gen/c2-wasm-2x.c",
250 "src/f32-vmulcaddc/gen/c4-wasm-2x.c",
251]
252
Marat Dukhan08c4a432019-10-03 09:29:21 -0700253PSIMD_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800254 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800255 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800256 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700257 "src/f32-avgpool/mp9p8q-psimd.c",
258 "src/f32-avgpool/up9-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800259 "src/f32-bilinear/gen/psimd-c4.c",
260 "src/f32-bilinear/gen/psimd-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700261 "src/f32-clamp/psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800262 "src/f32-dwconv/gen/up4x25-psimd-acc2.c",
263 "src/f32-dwconv/gen/up4x25-psimd.c",
264 "src/f32-dwconv/gen/up4x4-psimd-acc2.c",
265 "src/f32-dwconv/gen/up4x4-psimd.c",
266 "src/f32-dwconv/gen/up4x9-psimd-acc2.c",
267 "src/f32-dwconv/gen/up4x9-psimd.c",
268 "src/f32-dwconv/gen/up8x25-psimd-acc2.c",
269 "src/f32-dwconv/gen/up8x25-psimd.c",
270 "src/f32-dwconv/gen/up8x4-psimd-acc2.c",
271 "src/f32-dwconv/gen/up8x4-psimd.c",
272 "src/f32-dwconv/gen/up8x9-psimd-acc2.c",
273 "src/f32-dwconv/gen/up8x9-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700274 "src/f32-gavgpool/mp7p7q-psimd.c",
275 "src/f32-gavgpool/up7-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800276 "src/f32-gemm/gen/1x8-psimd-loadsplat.c",
277 "src/f32-gemm/gen/1x8-psimd-splat.c",
278 "src/f32-gemm/gen/1x8s4-psimd.c",
279 "src/f32-gemm/gen/4x8-psimd-loadsplat.c",
280 "src/f32-gemm/gen/4x8-psimd-splat.c",
281 "src/f32-gemm/gen/4x8s4-psimd.c",
282 "src/f32-gemm/gen/6x8-psimd-loadsplat.c",
283 "src/f32-gemm/gen/6x8-psimd-splat.c",
284 "src/f32-gemm/gen/6x8s4-psimd.c",
285 "src/f32-gemm/gen-inc/1x8-psimd-loadsplat.c",
286 "src/f32-gemm/gen-inc/1x8-psimd-splat.c",
287 "src/f32-gemm/gen-inc/1x8s4-psimd.c",
288 "src/f32-gemm/gen-inc/4x8-psimd-loadsplat.c",
289 "src/f32-gemm/gen-inc/4x8-psimd-splat.c",
290 "src/f32-gemm/gen-inc/4x8s4-psimd.c",
291 "src/f32-gemm/gen-inc/6x8-psimd-loadsplat.c",
292 "src/f32-gemm/gen-inc/6x8-psimd-splat.c",
293 "src/f32-gemm/gen-inc/6x8s4-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700294 "src/f32-hswish/psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800295 "src/f32-igemm/gen/1x8-psimd-loadsplat.c",
296 "src/f32-igemm/gen/1x8-psimd-splat.c",
297 "src/f32-igemm/gen/1x8s4-psimd.c",
298 "src/f32-igemm/gen/4x2c4-psimd.c",
299 "src/f32-igemm/gen/4x8-psimd-loadsplat.c",
300 "src/f32-igemm/gen/4x8-psimd-splat.c",
301 "src/f32-igemm/gen/4x8s4-psimd.c",
302 "src/f32-igemm/gen/6x8-psimd-loadsplat.c",
303 "src/f32-igemm/gen/6x8-psimd-splat.c",
304 "src/f32-igemm/gen/6x8s4-psimd.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800305 "src/f32-maxpool/9p8x-psimd-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700306 "src/f32-pavgpool/mp9p8q-psimd.c",
307 "src/f32-pavgpool/up9-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800308 "src/f32-ppmm/gen/4x8-psimd.c",
309 "src/f32-prelu/gen/psimd-2x4.c",
310 "src/f32-prelu/gen/psimd-2x8.c",
311 "src/f32-vbinary/gen/vadd-psimd-x4.c",
312 "src/f32-vbinary/gen/vadd-psimd-x8.c",
313 "src/f32-vbinary/gen/vaddc-psimd-x4.c",
314 "src/f32-vbinary/gen/vaddc-psimd-x8.c",
315 "src/f32-vbinary/gen/vmul-psimd-x4.c",
316 "src/f32-vbinary/gen/vmul-psimd-x8.c",
317 "src/f32-vbinary/gen/vmulc-psimd-x4.c",
318 "src/f32-vbinary/gen/vmulc-psimd-x8.c",
319 "src/f32-vbinary/gen/vrsubc-psimd-x4.c",
320 "src/f32-vbinary/gen/vrsubc-psimd-x8.c",
321 "src/f32-vbinary/gen/vsub-psimd-x4.c",
322 "src/f32-vbinary/gen/vsub-psimd-x8.c",
323 "src/f32-vbinary/gen/vsubc-psimd-x4.c",
324 "src/f32-vbinary/gen/vsubc-psimd-x8.c",
325 "src/f32-vmulcaddc/gen/c4-psimd-2x.c",
326 "src/f32-vmulcaddc/gen/c8-psimd-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700327 "src/x32-packx/x4-psimd.c",
328 "src/x32-pad/x2-psimd.c",
329 "src/x32-unpool/psimd.c",
330 "src/x32-zip/x2-psimd.c",
331 "src/x32-zip/x3-psimd.c",
332 "src/x32-zip/x4-psimd.c",
333 "src/x32-zip/xm-psimd.c",
334]
335
336# ISA-specific micro-kernels
337NEON_UKERNELS = [
338 "src/f32-avgpool/mp9p8q-neon.c",
339 "src/f32-avgpool/up9-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800340 "src/f32-bilinear/gen/neon-c4.c",
341 "src/f32-bilinear/gen/neon-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700342 "src/f32-clamp/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800343 "src/f32-dwconv/gen/up4x9-neon.c",
344 "src/f32-dwconv/gen/up4x9-neon-acc2.c",
345 "src/f32-dwconv/gen/up8x9-neon.c",
346 "src/f32-dwconv/gen/up8x9-neon-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700347 "src/f32-gavgpool-spchw/neon-x4.c",
348 "src/f32-gavgpool/mp7p7q-neon.c",
349 "src/f32-gavgpool/up7-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800350 "src/f32-gemm/gen/1x8-neon-lane-ld64.c",
351 "src/f32-gemm/gen/4x2-neon-lane-ld64.c",
352 "src/f32-gemm/gen/4x8-neon-lane-ld128.c",
353 "src/f32-gemm/gen/4x8-neon-lane-ld64.c",
354 "src/f32-gemm/gen/5x8-neon-lane-ld64.c",
355 "src/f32-gemm/gen/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800356 "src/f32-gemm/gen/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800357 "src/f32-gemm/gen/1x8-neon-dup-ld64.c",
358 "src/f32-gemm/gen/4x8-neon-dup-ld128.c",
359 "src/f32-gemm/gen/4x8-neon-dup-ld64.c",
360 "src/f32-gemm/gen/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800361 "src/f32-gemm/gen/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800362 "src/f32-gemm/gen/1x8s4-neon.c",
363 "src/f32-gemm/gen/4x8s4-neon.c",
364 "src/f32-gemm/gen/6x8s4-neon.c",
365 "src/f32-gemm/gen/8x8s4-neon.c",
366 "src/f32-gemm/gen-inc/1x8-neon-lane-ld64.c",
367 "src/f32-gemm/gen-inc/4x8-neon-lane-ld128.c",
368 "src/f32-gemm/gen-inc/4x8-neon-lane-ld64.c",
369 "src/f32-gemm/gen-inc/5x8-neon-lane-ld64.c",
370 "src/f32-gemm/gen-inc/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800371 "src/f32-gemm/gen-inc/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800372 "src/f32-gemm/gen-inc/1x8-neon-dup-ld64.c",
373 "src/f32-gemm/gen-inc/4x8-neon-dup-ld128.c",
374 "src/f32-gemm/gen-inc/4x8-neon-dup-ld64.c",
375 "src/f32-gemm/gen-inc/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800376 "src/f32-gemm/gen-inc/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800377 "src/f32-gemm/gen-inc/1x8s4-neon.c",
378 "src/f32-gemm/gen-inc/4x8s4-neon.c",
379 "src/f32-gemm/gen-inc/6x8s4-neon.c",
380 "src/f32-gemm/gen-inc/8x8s4-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700381 "src/f32-hswish/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800382 "src/f32-igemm/gen/1x8-neon-lane-ld64.c",
383 "src/f32-igemm/gen/4x2-neon-lane-ld64.c",
384 "src/f32-igemm/gen/4x4-neon-lane-ld64.c",
385 "src/f32-igemm/gen/4x8-neon-lane-ld128.c",
386 "src/f32-igemm/gen/4x8-neon-lane-ld64.c",
387 "src/f32-igemm/gen/6x8-neon-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800388 "src/f32-igemm/gen/6x8-neon-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800389 "src/f32-igemm/gen/1x8-neon-dup-ld64.c",
390 "src/f32-igemm/gen/4x8-neon-dup-ld128.c",
391 "src/f32-igemm/gen/4x8-neon-dup-ld64.c",
392 "src/f32-igemm/gen/6x8-neon-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800393 "src/f32-igemm/gen/6x8-neon-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800394 "src/f32-igemm/gen/1x8s4-neon.c",
395 "src/f32-igemm/gen/4x8s4-neon.c",
396 "src/f32-igemm/gen/6x8s4-neon.c",
397 "src/f32-igemm/gen/8x8s4-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700398 "src/f32-pavgpool/mp9p8q-neon.c",
399 "src/f32-pavgpool/up9-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800400 "src/f32-ppmm/gen/4x8-neon.c",
401 "src/f32-ppmm/gen/8x8-neon.c",
402 "src/f32-prelu/gen/neon-2x4.c",
403 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700404 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800405 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
406 "src/f32-vbinary/gen/vadd-neon-x4.c",
407 "src/f32-vbinary/gen/vadd-neon-x8.c",
408 "src/f32-vbinary/gen/vaddc-neon-x4.c",
409 "src/f32-vbinary/gen/vaddc-neon-x8.c",
410 "src/f32-vbinary/gen/vmul-neon-x4.c",
411 "src/f32-vbinary/gen/vmul-neon-x8.c",
412 "src/f32-vbinary/gen/vmulc-neon-x4.c",
413 "src/f32-vbinary/gen/vmulc-neon-x8.c",
414 "src/f32-vbinary/gen/vrsubc-neon-x4.c",
415 "src/f32-vbinary/gen/vrsubc-neon-x8.c",
416 "src/f32-vbinary/gen/vsub-neon-x4.c",
417 "src/f32-vbinary/gen/vsub-neon-x8.c",
418 "src/f32-vbinary/gen/vsubc-neon-x4.c",
419 "src/f32-vbinary/gen/vsubc-neon-x8.c",
420 "src/f32-vmulcaddc/gen/c4-neon-2x.c",
421 "src/f32-vmulcaddc/gen/c8-neon-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700422 "src/q8-avgpool/mp9p8q-neon.c",
423 "src/q8-avgpool/up9-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700424 "src/q8-dwconv/up8x9-neon.c",
425 "src/q8-gavgpool/mp7p7q-neon.c",
426 "src/q8-gavgpool/up7-neon.c",
427 "src/q8-gemm/4x8-neon.c",
428 "src/q8-gemm/8x8-neon.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800429 "src/q8-igemm/4x8-neon.c",
430 "src/q8-igemm/8x8-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700431 "src/q8-vadd/neon.c",
432 "src/u8-clamp/neon.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800433 "src/u8-maxpool/9p8x-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700434 "src/u8-rmax/neon.c",
435 "src/x32-packx/x4-neon-st4.c",
436 "src/x32-pad/x2-neon.c",
437 "src/x32-zip/x2-neon.c",
438 "src/x32-zip/x3-neon.c",
439 "src/x32-zip/x4-neon.c",
440 "src/x32-zip/xm-neon.c",
441 "src/x8-zip/x2-neon.c",
442 "src/x8-zip/x3-neon.c",
443 "src/x8-zip/x4-neon.c",
444 "src/x8-zip/xm-neon.c",
445]
446
447NEONFMA_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800448 "src/f32-bilinear/gen/neonfma-c4.c",
449 "src/f32-bilinear/gen/neonfma-c8.c",
450 "src/f32-igemm/gen/1x8-neonfma-dup-ld64.c",
451 "src/f32-igemm/gen/4x8-neonfma-dup-ld128.c",
452 "src/f32-igemm/gen/4x8-neonfma-dup-ld64.c",
453 "src/f32-igemm/gen/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800454 "src/f32-igemm/gen/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800455 "src/f32-igemm/gen/1x8s4-neonfma.c",
456 "src/f32-igemm/gen/4x8s4-neonfma.c",
457 "src/f32-igemm/gen/6x8s4-neonfma.c",
458 "src/f32-igemm/gen/8x8s4-neonfma.c",
459 "src/f32-dwconv/gen/up4x9-neonfma.c",
460 "src/f32-dwconv/gen/up4x9-neonfma-acc2.c",
461 "src/f32-dwconv/gen/up8x9-neonfma.c",
462 "src/f32-dwconv/gen/up8x9-neonfma-acc2.c",
463 "src/f32-gemm/gen/1x8-neonfma-dup-ld64.c",
464 "src/f32-gemm/gen/4x8-neonfma-dup-ld128.c",
465 "src/f32-gemm/gen/4x8-neonfma-dup-ld64.c",
466 "src/f32-gemm/gen/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800467 "src/f32-gemm/gen/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800468 "src/f32-gemm/gen/1x8s4-neonfma.c",
469 "src/f32-gemm/gen/4x8s4-neonfma.c",
470 "src/f32-gemm/gen/6x8s4-neonfma.c",
471 "src/f32-gemm/gen/8x8s4-neonfma.c",
472 "src/f32-gemm/gen-inc/1x8-neonfma-dup-ld64.c",
473 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld128.c",
474 "src/f32-gemm/gen-inc/4x8-neonfma-dup-ld64.c",
475 "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800476 "src/f32-gemm/gen-inc/6x8-neonfma-dup-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800477 "src/f32-gemm/gen-inc/1x8s4-neonfma.c",
478 "src/f32-gemm/gen-inc/4x8s4-neonfma.c",
479 "src/f32-gemm/gen-inc/6x8s4-neonfma.c",
480 "src/f32-gemm/gen-inc/8x8s4-neonfma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700481 "src/f32-hswish/neonfma.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800482 "src/f32-ppmm/gen/4x8-neonfma.c",
483 "src/f32-ppmm/gen/8x8-neonfma.c",
484 "src/f32-sigmoid/gen/neonfma-p5-nr2fma-x16.c",
485 "src/f32-vmulcaddc/gen/c4-neonfma-2x.c",
486 "src/f32-vmulcaddc/gen/c8-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800487 "src/math/exp-neonfma-lut64-p2.c",
488 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800489 "src/math/expminus-neonfma-lut2048-p1.c",
490 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800491 "src/math/expminus-neonfma-p5.c",
Marat Dukhan91f8d862019-11-27 12:25:42 -0800492 "src/math/sigmoid-neonfma-lut2048-p1-nr1recps1fma.c",
493 "src/math/sigmoid-neonfma-lut2048-p1-nr2fma.c",
494 "src/math/sigmoid-neonfma-lut2048-p1-nr2recps.c",
Marat Dukhan22aae132019-11-22 17:10:29 -0800495 "src/math/sigmoid-neonfma-p5-nr1recps1fma.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -0800496 "src/math/sigmoid-neonfma-p5-nr2fma.c",
Marat Dukhan22aae132019-11-22 17:10:29 -0800497 "src/math/sigmoid-neonfma-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700498]
499
500AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800501 "src/f32-gemm/gen/1x8-neonfma-lane-ld64.c",
502 "src/f32-gemm/gen/4x2-neonfma-lane-ld64.c",
503 "src/f32-gemm/gen/4x8-neonfma-lane-ld128.c",
504 "src/f32-gemm/gen/4x8-neonfma-lane-ld64.c",
505 "src/f32-gemm/gen/5x8-neonfma-lane-ld64.c",
506 "src/f32-gemm/gen/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800507 "src/f32-gemm/gen/6x8-neonfma-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800508 "src/f32-gemm/gen-inc/1x8-neonfma-lane-ld64.c",
509 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld128.c",
510 "src/f32-gemm/gen-inc/4x8-neonfma-lane-ld64.c",
511 "src/f32-gemm/gen-inc/5x8-neonfma-lane-ld64.c",
512 "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800513 "src/f32-gemm/gen-inc/6x8-neonfma-lane-ld128.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800514 "src/f32-igemm/gen/1x8-neonfma-lane-ld64.c",
515 "src/f32-igemm/gen/4x2-neonfma-lane-ld64.c",
516 "src/f32-igemm/gen/4x4-neonfma-lane-ld64.c",
517 "src/f32-igemm/gen/4x8-neonfma-lane-ld128.c",
518 "src/f32-igemm/gen/4x8-neonfma-lane-ld64.c",
519 "src/f32-igemm/gen/6x8-neonfma-lane-ld64.c",
Frank Barchard69172d92019-11-26 16:22:39 -0800520 "src/f32-igemm/gen/6x8-neonfma-lane-ld128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700521 "src/f32-conv-hwc/3x3s2p1c3x4-neonfma-2x2.c",
522 "src/f32-conv-hwc/3x3s2p1c3x8-neonfma-2x2.c",
523 "src/f32-conv-hwc2spchw/3x3s2p1c3x4-neonfma-2x2.c",
524 "src/f32-dwconv-spchw/3x3p1-neonfma.c",
525 "src/f32-dwconv-spchw/5x5p2-neonfma.c",
526 "src/f32-dwconv-spchw/3x3s2p1-neonfma.c",
527 "src/f32-dwconv-spchw/5x5s2p2-neonfma.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800528 "src/f32-spmm/gen/12x1-neonfma.c",
529 "src/f32-spmm/gen/12x2-neonfma.c",
530 "src/f32-spmm/gen/12x4-neonfma.c",
531 "src/f32-spmm/gen/16x1-neonfma-pipelined.c",
532 "src/f32-spmm/gen/16x1-neonfma-unroll2.c",
533 "src/f32-spmm/gen/16x1-neonfma.c",
534 "src/f32-spmm/gen/16x2-neonfma.c",
535 "src/f32-spmm/gen/16x4-neonfma.c",
536 "src/f32-spmm/gen/4x1-neonfma-pipelined.c",
537 "src/f32-spmm/gen/4x1-neonfma-unroll2.c",
538 "src/f32-spmm/gen/4x1-neonfma.c",
539 "src/f32-spmm/gen/4x2-neonfma.c",
540 "src/f32-spmm/gen/4x4-neonfma.c",
541 "src/f32-spmm/gen/8x1-neonfma-pipelined.c",
542 "src/f32-spmm/gen/8x1-neonfma-unroll2.c",
543 "src/f32-spmm/gen/8x1-neonfma.c",
544 "src/f32-spmm/gen/8x2-neonfma.c",
545 "src/f32-spmm/gen/8x4-neonfma.c",
Marat Dukhan22aae132019-11-22 17:10:29 -0800546 "src/math/sigmoid-neonfma-p5-div.c",
Marat Dukhan91f8d862019-11-27 12:25:42 -0800547 "src/math/sigmoid-neonfma-lut2048-p1-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700548]
549
550AARCH64_NEONFP16ARITH_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800551 "src/f16-gemm/gen/4x8-neonfp16arith-ld64.c",
552 "src/f16-gemm/gen/6x8-neonfp16arith-ld64.c",
553 "src/f16-gemm/gen/8x8-neonfp16arith-ld64.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700554]
555
556SSE_UKERNELS = [
557 "src/f32-avgpool/mp9p8q-sse.c",
558 "src/f32-avgpool/up9-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800559 "src/f32-bilinear/gen/sse-c4.c",
560 "src/f32-bilinear/gen/sse-c8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700561 "src/f32-clamp/sse.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800562 "src/f32-dwconv-spchw/3x3p1-sse.c",
563 "src/f32-dwconv-spchw/3x3s2p1-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800564 "src/f32-dwconv/gen/up4x25-sse-acc2.c",
565 "src/f32-dwconv/gen/up4x25-sse.c",
566 "src/f32-dwconv/gen/up4x4-sse-acc2.c",
567 "src/f32-dwconv/gen/up4x4-sse.c",
568 "src/f32-dwconv/gen/up4x9-sse-acc2.c",
569 "src/f32-dwconv/gen/up4x9-sse.c",
570 "src/f32-dwconv/gen/up8x25-sse-acc2.c",
571 "src/f32-dwconv/gen/up8x25-sse.c",
572 "src/f32-dwconv/gen/up8x4-sse-acc2.c",
573 "src/f32-dwconv/gen/up8x4-sse.c",
574 "src/f32-dwconv/gen/up8x9-sse-acc2.c",
575 "src/f32-dwconv/gen/up8x9-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700576 "src/f32-gavgpool-spchw/sse-x4.c",
577 "src/f32-gavgpool/mp7p7q-sse.c",
578 "src/f32-gavgpool/up7-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800579 "src/f32-gemm/gen/1x8-sse-dup.c",
580 "src/f32-gemm/gen/1x8-sse-load1.c",
581 "src/f32-gemm/gen/1x8s4-sse.c",
582 "src/f32-gemm/gen/4x8-sse-dup.c",
583 "src/f32-gemm/gen/4x8-sse-load1.c",
584 "src/f32-gemm/gen/4x8s4-sse.c",
585 "src/f32-gemm/gen-inc/1x8-sse-dup.c",
586 "src/f32-gemm/gen-inc/1x8-sse-load1.c",
587 "src/f32-gemm/gen-inc/1x8s4-sse.c",
588 "src/f32-gemm/gen-inc/4x8-sse-dup.c",
589 "src/f32-gemm/gen-inc/4x8-sse-load1.c",
590 "src/f32-gemm/gen-inc/4x8s4-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700591 "src/f32-hswish/sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800592 "src/f32-igemm/gen/1x8-sse-dup.c",
593 "src/f32-igemm/gen/1x8-sse-load1.c",
594 "src/f32-igemm/gen/1x8s4-sse.c",
595 "src/f32-igemm/gen/4x2c4-sse.c",
596 "src/f32-igemm/gen/4x8-sse-dup.c",
597 "src/f32-igemm/gen/4x8-sse-load1.c",
598 "src/f32-igemm/gen/4x8s4-sse.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800599 "src/f32-maxpool/9p8x-sse-c4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700600 "src/f32-pavgpool/mp9p8q-sse.c",
601 "src/f32-pavgpool/up9-sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800602 "src/f32-ppmm/gen/4x8-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700603 "src/f32-rmax/sse.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800604 "src/f32-spmm/gen/4x1-sse.c",
605 "src/f32-spmm/gen/8x1-sse.c",
606 "src/f32-vbinary/gen/vadd-sse-x4.c",
607 "src/f32-vbinary/gen/vadd-sse-x8.c",
608 "src/f32-vbinary/gen/vaddc-sse-x4.c",
609 "src/f32-vbinary/gen/vaddc-sse-x8.c",
610 "src/f32-vbinary/gen/vmul-sse-x4.c",
611 "src/f32-vbinary/gen/vmul-sse-x8.c",
612 "src/f32-vbinary/gen/vmulc-sse-x4.c",
613 "src/f32-vbinary/gen/vmulc-sse-x8.c",
614 "src/f32-vbinary/gen/vrsubc-sse-x4.c",
615 "src/f32-vbinary/gen/vrsubc-sse-x8.c",
616 "src/f32-vbinary/gen/vsub-sse-x4.c",
617 "src/f32-vbinary/gen/vsub-sse-x8.c",
618 "src/f32-vbinary/gen/vsubc-sse-x4.c",
619 "src/f32-vbinary/gen/vsubc-sse-x8.c",
620 "src/f32-vmulcaddc/gen/c4-sse-2x.c",
621 "src/f32-vmulcaddc/gen/c8-sse-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700622 "src/x32-packx/x4-sse.c",
623]
624
625SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800626 "src/f32-argmaxpool/9p8x-sse2-c4.c",
627 "src/f32-argmaxpool/4x-sse2-c4.c",
628 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800629 "src/f32-prelu/gen/sse2-2x4.c",
630 "src/f32-prelu/gen/sse2-2x8.c",
631 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
632 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700633 "src/q8-avgpool/mp9p8q-sse2.c",
634 "src/q8-avgpool/up9-sse2.c",
635 "src/q8-igemm/4x4c2-sse2.c",
636 "src/q8-dwconv/up8x9-sse2.c",
637 "src/q8-gavgpool/mp7p7q-sse2.c",
638 "src/q8-gavgpool/up7-sse2.c",
639 "src/q8-gemm/2x4c8-sse2.c",
640 "src/q8-gemm/4x4c2-sse2.c",
641 "src/q8-vadd/sse2.c",
642 "src/u8-clamp/sse2.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800643 "src/u8-maxpool/9p8x-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700644 "src/u8-rmax/sse2.c",
645 "src/x32-pad/x2-sse2.c",
646 "src/x32-zip/x2-sse2.c",
647 "src/x32-zip/x3-sse2.c",
648 "src/x32-zip/x4-sse2.c",
649 "src/x32-zip/xm-sse2.c",
650 "src/x8-zip/x2-sse2.c",
651 "src/x8-zip/x3-sse2.c",
652 "src/x8-zip/x4-sse2.c",
653 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -0800654 "src/math/exp-sse2-p5.c",
655 "src/math/expminus-sse2-p5.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -0800656 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700657]
658
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800659SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800660 "src/f32-prelu/gen/sse41-2x4.c",
661 "src/f32-prelu/gen/sse41-2x8.c",
662 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
663 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -0800664]
665
Marat Dukhan08c4a432019-10-03 09:29:21 -0700666AVX_UKERNELS = [
Marat Dukhane2c3f292019-11-27 15:40:54 -0800667 "src/f32-clamp/avx.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800668 "src/f32-dwconv/gen/up16x4-avx-acc2.c",
669 "src/f32-dwconv/gen/up16x4-avx.c",
670 "src/f32-dwconv/gen/up8x4-avx-acc2.c",
671 "src/f32-dwconv/gen/up8x4-avx.c",
672 "src/f32-dwconv/gen/up16x9-avx-acc2.c",
673 "src/f32-dwconv/gen/up16x9-avx.c",
674 "src/f32-dwconv/gen/up8x9-avx-acc2.c",
675 "src/f32-dwconv/gen/up8x9-avx.c",
676 "src/f32-dwconv/gen/up16x25-avx-acc2.c",
677 "src/f32-dwconv/gen/up16x25-avx.c",
678 "src/f32-dwconv/gen/up8x25-avx-acc2.c",
679 "src/f32-dwconv/gen/up8x25-avx.c",
680 "src/f32-gemm/gen/1x8-avx-broadcast.c",
681 "src/f32-gemm/gen/4x8-avx-broadcast.c",
682 "src/f32-gemm/gen/5x8-avx-broadcast.c",
683 "src/f32-gemm/gen/6x8-avx-broadcast.c",
684 "src/f32-gemm/gen/7x8-avx-broadcast.c",
685 "src/f32-gemm/gen-inc/1x8-avx-broadcast.c",
686 "src/f32-gemm/gen-inc/4x8-avx-broadcast.c",
687 "src/f32-gemm/gen-inc/5x8-avx-broadcast.c",
688 "src/f32-gemm/gen-inc/6x8-avx-broadcast.c",
689 "src/f32-gemm/gen-inc/7x8-avx-broadcast.c",
690 "src/f32-igemm/gen/1x8-avx-broadcast.c",
691 "src/f32-igemm/gen/4x8-avx-broadcast.c",
692 "src/f32-igemm/gen/5x8-avx-broadcast.c",
693 "src/f32-igemm/gen/6x8-avx-broadcast.c",
694 "src/f32-igemm/gen/7x8-avx-broadcast.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700695 "src/f32-rmax/avx.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700696 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700697]
698
Marat Dukhanfda12b82019-11-21 12:27:59 -0800699FMA3_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -0800700 "src/f32-dwconv/gen/up16x4-fma3-acc2.c",
701 "src/f32-dwconv/gen/up16x4-fma3.c",
702 "src/f32-dwconv/gen/up8x4-fma3-acc2.c",
703 "src/f32-dwconv/gen/up8x4-fma3.c",
704 "src/f32-dwconv/gen/up16x9-fma3-acc2.c",
705 "src/f32-dwconv/gen/up16x9-fma3.c",
706 "src/f32-dwconv/gen/up8x9-fma3-acc2.c",
707 "src/f32-dwconv/gen/up8x9-fma3.c",
708 "src/f32-dwconv/gen/up16x25-fma3-acc2.c",
709 "src/f32-dwconv/gen/up16x25-fma3.c",
710 "src/f32-dwconv/gen/up8x25-fma3-acc2.c",
711 "src/f32-dwconv/gen/up8x25-fma3.c",
712 "src/f32-gemm/gen/1x8-fma3-broadcast.c",
713 "src/f32-gemm/gen/4x8-fma3-broadcast.c",
714 "src/f32-gemm/gen/5x8-fma3-broadcast.c",
715 "src/f32-gemm/gen/6x8-fma3-broadcast.c",
716 "src/f32-gemm/gen/7x8-fma3-broadcast.c",
717 "src/f32-gemm/gen/8x8-fma3-broadcast.c",
718 "src/f32-gemm/gen-inc/1x8-fma3-broadcast.c",
719 "src/f32-gemm/gen-inc/4x8-fma3-broadcast.c",
720 "src/f32-gemm/gen-inc/5x8-fma3-broadcast.c",
721 "src/f32-gemm/gen-inc/6x8-fma3-broadcast.c",
722 "src/f32-gemm/gen-inc/7x8-fma3-broadcast.c",
723 "src/f32-gemm/gen-inc/8x8-fma3-broadcast.c",
724 "src/f32-igemm/gen/1x8-fma3-broadcast.c",
725 "src/f32-igemm/gen/4x8-fma3-broadcast.c",
726 "src/f32-igemm/gen/5x8-fma3-broadcast.c",
727 "src/f32-igemm/gen/6x8-fma3-broadcast.c",
728 "src/f32-igemm/gen/7x8-fma3-broadcast.c",
729 "src/f32-igemm/gen/8x8-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -0800730]
731
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700732AVX2_UKERNELS = [
Marat Dukhan97579532019-10-18 16:40:39 -0700733 "src/f32-raddexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700734 "src/f32-raddextexp/avx2-p5-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700735 "src/f32-raddstoreexpminusmax/avx2-p5-unroll64.c",
736 "src/f32-vscaleexpminusmax/avx2-p5-unroll64.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700737 "src/f32-vscaleextexp/avx2-p5-unroll64.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700738 "src/math/exp-avx2-p5.c",
739 "src/math/exp-avx2-perm-p3.c",
740 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -0700741 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700742 "src/math/extexp-avx2-p5.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700743]
744
Marat Dukhan08c4a432019-10-03 09:29:21 -0700745AVX512F_UKERNELS = [
Marat Dukhane2c3f292019-11-27 15:40:54 -0800746 "src/f32-clamp/avx512f.c",
Marat Dukhan479f87e2019-11-27 15:17:06 -0800747 "src/f32-dwconv/gen/up32x4-avx512f-acc2.c",
748 "src/f32-dwconv/gen/up32x4-avx512f.c",
749 "src/f32-dwconv/gen/up16x4-avx512f-acc2.c",
750 "src/f32-dwconv/gen/up16x4-avx512f.c",
751 "src/f32-dwconv/gen/up32x9-avx512f-acc2.c",
752 "src/f32-dwconv/gen/up32x9-avx512f.c",
753 "src/f32-dwconv/gen/up16x9-avx512f-acc2.c",
754 "src/f32-dwconv/gen/up16x9-avx512f.c",
755 "src/f32-dwconv/gen/up32x25-avx512f-acc2.c",
756 "src/f32-dwconv/gen/up32x25-avx512f.c",
757 "src/f32-dwconv/gen/up16x25-avx512f-acc2.c",
758 "src/f32-dwconv/gen/up16x25-avx512f.c",
Marat Dukhan0f349c42019-11-27 11:58:54 -0800759 "src/f32-gemm/gen/1x16-avx512f-broadcast.c",
760 "src/f32-gemm/gen/4x16-avx512f-broadcast.c",
761 "src/f32-gemm/gen/5x16-avx512f-broadcast.c",
762 "src/f32-gemm/gen/6x16-avx512f-broadcast.c",
763 "src/f32-gemm/gen/7x16-avx512f-broadcast.c",
764 "src/f32-gemm/gen/8x16-avx512f-broadcast.c",
765 "src/f32-gemm/gen-inc/1x16-avx512f-broadcast.c",
766 "src/f32-gemm/gen-inc/4x16-avx512f-broadcast.c",
767 "src/f32-gemm/gen-inc/5x16-avx512f-broadcast.c",
768 "src/f32-gemm/gen-inc/6x16-avx512f-broadcast.c",
769 "src/f32-gemm/gen-inc/7x16-avx512f-broadcast.c",
770 "src/f32-gemm/gen-inc/8x16-avx512f-broadcast.c",
771 "src/f32-igemm/gen/1x16-avx512f-broadcast.c",
772 "src/f32-igemm/gen/4x16-avx512f-broadcast.c",
773 "src/f32-igemm/gen/5x16-avx512f-broadcast.c",
774 "src/f32-igemm/gen/6x16-avx512f-broadcast.c",
775 "src/f32-igemm/gen/7x16-avx512f-broadcast.c",
776 "src/f32-igemm/gen/8x16-avx512f-broadcast.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700777 "src/f32-raddexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700778 "src/f32-raddextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700779 "src/f32-raddstoreexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700780 "src/f32-rmax/avx512f.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700781 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan97579532019-10-18 16:40:39 -0700782 "src/f32-vscaleexpminusmax/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700783 "src/f32-vscaleextexp/avx512f-p5-scalef-unroll128.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700784 "src/math/exp-avx512f-p5-scalef.c",
785 "src/math/exp-avx512f-p5.c",
786 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -0700787 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -0700788 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700789]
790
791AARCH32_ASM_UKERNELS = [
792 "src/q8-dwconv/up8x9-aarch32-neon.S",
Frank Barchard8b0f0262019-11-27 23:18:40 -0800793 "src/f32-gemm/4x8-aarch32-neon-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700794]
795
796AARCH64_ASM_UKERNELS = [
797 "src/f32-dwconv/up4x9-aarch64-neonfma-cortex-a55.S",
798 "src/f32-dwconv/up4x9-aarch64-neonfma.S",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800799 "src/f32-gemm/gen/1x12-aarch64-neonfma-cortex-a53.S",
800 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a53.S",
801 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a57.S",
802 "src/f32-gemm/gen/1x8-aarch64-neonfma-cortex-a75.S",
803 "src/f32-gemm/gen/4x12-aarch64-neonfma-cortex-a53.S",
804 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a53.S",
805 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a57.S",
806 "src/f32-gemm/gen/4x8-aarch64-neonfma-cortex-a75.S",
807 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld128.S",
808 "src/f32-gemm/gen/4x8-aarch64-neonfma-ld64.S",
809 "src/f32-gemm/gen/5x8-aarch64-neonfma-cortex-a75.S",
810 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a53.S",
811 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a57.S",
812 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a73.S",
813 "src/f32-gemm/gen/6x8-aarch64-neonfma-cortex-a75.S",
814 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld128.S",
815 "src/f32-gemm/gen/6x8-aarch64-neonfma-ld64.S",
816 "src/f32-gemm/gen-inc/1x12-aarch64-neonfma-cortex-a53.S",
817 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a53.S",
818 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a57.S",
819 "src/f32-gemm/gen-inc/1x8-aarch64-neonfma-cortex-a75.S",
820 "src/f32-gemm/gen-inc/4x12-aarch64-neonfma-cortex-a53.S",
821 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a53.S",
822 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a57.S",
823 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-cortex-a75.S",
824 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld128.S",
825 "src/f32-gemm/gen-inc/4x8-aarch64-neonfma-ld64.S",
826 "src/f32-gemm/gen-inc/5x8-aarch64-neonfma-cortex-a75.S",
827 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a53.S",
828 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a57.S",
829 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a73.S",
830 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-cortex-a75.S",
831 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld128.S",
832 "src/f32-gemm/gen-inc/6x8-aarch64-neonfma-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700833 "src/f32-igemm/1x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard21be34f2019-10-09 19:32:19 -0700834 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700835 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a57.S",
836 "src/f32-igemm/1x8-aarch64-neonfma-cortex-a75.S",
837 "src/f32-igemm/4x12-aarch64-neonfma-cortex-a53.S",
Frank Barchard46fb8072019-10-25 12:54:22 -0700838 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700839 "src/f32-igemm/4x8-aarch64-neonfma-cortex-a75.S",
840 "src/f32-igemm/5x8-aarch64-neonfma-cortex-a75.S",
Frank Barcharda7fb8552019-10-23 17:14:17 -0700841 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a53.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700842 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a57.S",
843 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a73.S",
844 "src/f32-igemm/6x8-aarch64-neonfma-cortex-a75.S",
845]
846
847INTERNAL_MICROKERNEL_HDRS = [
848 "src/xnnpack/argmaxpool.h",
849 "src/xnnpack/avgpool.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -0800850 "src/xnnpack/bilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700851 "src/xnnpack/clamp.h",
852 "src/xnnpack/common.h",
853 "src/xnnpack/conv.h",
854 "src/xnnpack/dwconv.h",
855 "src/xnnpack/gavgpool.h",
856 "src/xnnpack/gemm.h",
857 "src/xnnpack/hswish.h",
858 "src/xnnpack/igemm.h",
859 "src/xnnpack/lut.h",
860 "src/xnnpack/math.h",
861 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -0800862 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700863 "src/xnnpack/packx.h",
864 "src/xnnpack/pad.h",
865 "src/xnnpack/params.h",
866 "src/xnnpack/pavgpool.h",
867 "src/xnnpack/ppmm.h",
868 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700869 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700870 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700871 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700872 "src/xnnpack/rmax.h",
873 "src/xnnpack/scalar-utils.h",
874 "src/xnnpack/spmm.h",
875 "src/xnnpack/unpool.h",
876 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800877 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700878 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -0700879 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -0700880 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -0700881 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800882 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700883 "src/xnnpack/zip.h",
884]
885
886INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
887 "include/xnnpack.h",
888 "src/xnnpack/allocator.h",
889 "src/xnnpack/compute.h",
890 "src/xnnpack/im2col.h",
891 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700892 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700893 "src/xnnpack/operator.h",
894 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700895 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700896 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -0700897 "src/xnnpack/requantization.h",
898]
899
900ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
901 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700902]
903
904MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700905 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700906 "include/xnnpack.h",
907]
908
909MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
910 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -0700911 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700912 "src/xnnpack/requantization.h",
913 "include/xnnpack.h",
914]
915
916OPERATOR_TEST_PARAMS_HDRS = [
917 "src/xnnpack/params.h",
918 "src/xnnpack/common.h",
919]
920
921WEIGHTS_PACK_HDRS = [
922 "src/xnnpack/pack.h",
923 "src/xnnpack/operator.h",
924 "src/xnnpack/compute.h",
925]
926
Marat Dukhanc8e00eb2019-10-04 14:55:26 -0700927LOGGING_COPTS = select({
928 # No logging in optimized mode
929 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
930 # Full logging in debug mode
931 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
932 # Error-only logging in default (fastbuild) mode
933 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
934})
935
936LOGGING_HDRS = [
937 "src/xnnpack/log.h",
938]
939
Marat Dukhan08c4a432019-10-03 09:29:21 -0700940xnnpack_cc_library(
941 name = "scalar_ukernels",
942 srcs = SCALAR_UKERNELS,
943 hdrs = INTERNAL_HDRS,
944 aarch32_copts = ["-marm"],
945 copts = xnnpack_std_copts(),
946 deps = [
947 "@FP16",
948 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -0800949 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700950 ],
951)
952
953xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -0800954 name = "wasm_ukernels",
955 hdrs = INTERNAL_HDRS,
956 copts = xnnpack_std_copts(),
957 wasm_srcs = WASM_UKERNELS,
958 deps = [
959 "@FP16",
960 "@FXdiv",
961 "@pthreadpool",
962 ],
963)
964
965xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -0700966 name = "psimd_ukernels",
967 srcs = PSIMD_UKERNELS,
968 hdrs = INTERNAL_HDRS,
969 aarch32_copts = [
970 "-marm",
971 "-mfpu=neon",
972 ],
973 copts = xnnpack_std_copts(),
974 optimized_copts = [
975 "-O3",
976 "-ffast-math",
977 ],
978 deps = [
979 "@FP16",
980 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -0800981 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700982 ],
983)
984
985xnnpack_cc_library(
986 name = "neon_ukernels",
987 hdrs = INTERNAL_HDRS,
988 aarch32_copts = [
989 "-marm",
990 "-mfpu=neon",
991 ],
992 aarch32_srcs = NEON_UKERNELS,
993 aarch64_srcs = NEON_UKERNELS,
994 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -0800995 deps = [
996 "@FP16",
997 "@pthreadpool",
998 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -0700999)
1000
1001xnnpack_cc_library(
1002 name = "neonfma_ukernels",
1003 hdrs = INTERNAL_HDRS,
1004 aarch32_copts = [
1005 "-marm",
1006 "-mfpu=neon-vfpv4",
1007 ],
1008 aarch32_srcs = NEONFMA_UKERNELS,
1009 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
1010 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001011 deps = [
1012 "@FP16",
1013 "@pthreadpool",
1014 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001015)
1016
1017xnnpack_cc_library(
1018 name = "neonfp16arith_ukernels",
1019 hdrs = INTERNAL_HDRS,
1020 aarch64_copts = ["-march=armv8.2-a+fp16"],
1021 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
1022 copts = xnnpack_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08001023 deps = [
1024 "@FP16",
1025 "@pthreadpool",
1026 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001027)
1028
1029xnnpack_cc_library(
1030 name = "sse2_ukernels",
1031 hdrs = INTERNAL_HDRS,
1032 copts = xnnpack_std_copts(),
1033 x86_copts = ["-msse2"],
1034 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001035 deps = [
1036 "@FP16",
1037 "@pthreadpool",
1038 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001039)
1040
1041xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001042 name = "sse41_ukernels",
1043 hdrs = INTERNAL_HDRS,
1044 copts = xnnpack_std_copts(),
1045 x86_copts = ["-msse4.1"],
1046 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001047 deps = [
1048 "@FP16",
1049 "@pthreadpool",
1050 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001051)
1052
1053xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001054 name = "avx_ukernels",
1055 hdrs = INTERNAL_HDRS,
1056 copts = xnnpack_std_copts(),
1057 x86_copts = ["-mavx"],
1058 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001059 deps = [
1060 "@FP16",
1061 "@pthreadpool",
1062 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001063)
1064
1065xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08001066 name = "fma3_ukernels",
1067 hdrs = INTERNAL_HDRS,
1068 copts = xnnpack_std_copts(),
1069 x86_copts = [
1070 "-mfma",
1071 ],
1072 x86_srcs = FMA3_UKERNELS,
1073 deps = [
1074 "@FP16",
1075 "@pthreadpool",
1076 ],
1077)
1078
1079xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001080 name = "avx2_ukernels",
1081 hdrs = INTERNAL_HDRS,
1082 copts = xnnpack_std_copts(),
1083 x86_copts = [
1084 "-mfma",
1085 "-mavx2",
1086 ],
1087 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001088 deps = [
1089 "@FP16",
1090 "@pthreadpool",
1091 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001092)
1093
1094xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001095 name = "avx512f_ukernels",
1096 hdrs = INTERNAL_HDRS,
1097 copts = xnnpack_std_copts(),
1098 x86_copts = ["-mavx512f"],
1099 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08001100 deps = [
1101 "@FP16",
1102 "@pthreadpool",
1103 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001104)
1105
1106xnnpack_cc_library(
1107 name = "asm_ukernels",
1108 hdrs = ["src/xnnpack/assembly.h"],
1109 aarch32_srcs = AARCH32_ASM_UKERNELS,
1110 aarch64_srcs = AARCH64_ASM_UKERNELS,
1111)
1112
1113xnnpack_aggregate_library(
1114 name = "ukernels",
1115 aarch32_deps = [
1116 ":psimd_ukernels",
1117 ":neon_ukernels",
1118 ":neonfma_ukernels",
1119 ":asm_ukernels",
1120 ],
1121 aarch64_deps = [
1122 ":psimd_ukernels",
1123 ":neon_ukernels",
1124 ":neonfma_ukernels",
1125 ":neonfp16arith_ukernels",
1126 ":asm_ukernels",
1127 ],
1128 generic_deps = [":scalar_ukernels"],
Marat Dukhan436ebe62019-12-04 15:10:12 -08001129 wasm_deps = [
1130 ":wasm_ukernels",
1131 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001132 wasmsimd_deps = [
Marat Dukhan436ebe62019-12-04 15:10:12 -08001133 ":wasm_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001134 ":psimd_ukernels",
1135 ],
1136 x86_deps = [
1137 ":psimd_ukernels",
1138 ":sse2_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001139 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001140 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001141 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001142 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001143 ":avx512f_ukernels",
1144 ],
1145)
1146
1147xnnpack_cc_library(
1148 name = "im2col",
1149 srcs = ["src/im2col.c"],
1150 hdrs = [
1151 "src/xnnpack/common.h",
1152 "src/xnnpack/im2col.h",
1153 ],
1154 copts = xnnpack_std_copts(),
1155)
1156
1157xnnpack_cc_library(
1158 name = "indirection",
1159 srcs = ["src/indirection.c"],
1160 hdrs = INTERNAL_HDRS,
1161 copts = xnnpack_std_copts(),
1162 deps = [
1163 "@FP16",
1164 "@FXdiv",
1165 "@pthreadpool",
1166 ],
1167)
1168
1169xnnpack_cc_library(
1170 name = "operator_run",
1171 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001172 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
1173 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001174 # Wrappers for multi-pass microkernels use VLAs for temporary buffers.
1175 "-Wno-vla",
1176 ],
1177 deps = [
1178 "@FP16",
1179 "@FXdiv",
1180 "@clog",
1181 "@pthreadpool",
1182 ],
1183)
1184
1185cc_library(
1186 name = "enable_assembly",
1187 defines = select({
1188 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
1189 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07001190 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001191 }),
1192)
1193
Marat Dukhancf056b22019-10-07 10:26:29 -07001194xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001195 name = "operators",
1196 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08001197 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001198 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07001199 ],
1200 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001201 copts = xnnpack_std_copts() + LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001202 "-Isrc",
1203 "-Iinclude",
1204 ] + select({
1205 ":debug_build": [],
1206 "//conditions:default": xnnpack_min_size_copts(),
1207 }),
Marat Dukhancf056b22019-10-07 10:26:29 -07001208 wasm_srcs = ["src/wasm-stubs.c"],
1209 wasmsimd_srcs = ["src/wasm-stubs.c"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001210 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001211 ":indirection",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001212 "@FP16",
1213 "@FXdiv",
1214 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001215 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001216 ],
1217)
1218
1219cc_library(
1220 name = "XNNPACK",
1221 srcs = [
1222 "src/init.c",
1223 ],
1224 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1225 "-Isrc",
1226 "-Iinclude",
1227 ] + select({
1228 ":debug_build": [],
1229 "//conditions:default": xnnpack_min_size_copts(),
1230 }),
1231 includes = ["include"],
1232 linkstatic = True,
1233 textual_hdrs = ["include/xnnpack.h"],
1234 visibility = xnnpack_visibility(),
1235 deps = [
1236 ":enable_assembly",
1237 ":ukernels",
1238 ":operator_run",
1239 ":operators",
1240 "@clog",
1241 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07001242 ] + select({
1243 ":emscripten": [],
1244 "//conditions:default": ["@cpuinfo"],
1245 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001246)
1247
1248cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001249 name = "xnnpack_operators_nhwc_f32",
1250 srcs = [
1251 "src/init.c",
1252 ],
1253 copts = xnnpack_std_copts() + LOGGING_COPTS + [
1254 "-Isrc",
1255 "-Iinclude",
1256 ] + select({
1257 ":debug_build": [],
1258 "//conditions:default": xnnpack_min_size_copts(),
1259 }),
1260 defines = [
1261 "XNN_NO_Q8_OPERATORS",
1262 "XNN_NO_U8_OPERATORS",
1263 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08001264 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001265 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001266 includes = ["include"],
1267 linkstatic = True,
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001268 textual_hdrs = ["include/xnnpack.h"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001269 visibility = xnnpack_visibility(),
1270 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001271 ":enable_assembly",
1272 ":ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001273 ":operator_run",
1274 ":operators",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001275 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001276 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07001277 ] + select({
1278 ":emscripten": [],
1279 "//conditions:default": ["@cpuinfo"],
1280 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001281)
1282
Marat Dukhancf056b22019-10-07 10:26:29 -07001283xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001284 name = "bench_utils",
1285 srcs = ["bench/utils.cc"],
1286 hdrs = ["bench/utils.h"],
1287 copts = ["-Wno-unused-result"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08001288 deps = [
1289 "@com_google_benchmark//:benchmark",
1290 "@cpuinfo",
1291 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001292)
1293
Frank Barchard7e955972019-10-11 10:34:25 -07001294######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07001295
1296xnnpack_benchmark(
1297 name = "q8_gemm_bench",
1298 srcs = [
1299 "bench/gemm.h",
1300 "bench/q8-gemm.cc",
1301 "src/xnnpack/AlignedAllocator.h",
1302 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1303 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
1304 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
1305)
1306
1307xnnpack_benchmark(
1308 name = "f16_gemm_bench",
1309 srcs = [
1310 "bench/f16-gemm.cc",
1311 "bench/gemm.h",
1312 "src/xnnpack/AlignedAllocator.h",
1313 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1314 copts = ["-Wno-unused-function"],
1315 deps = MICROKERNEL_BENCHMARK_DEPS,
1316)
1317
1318xnnpack_benchmark(
1319 name = "f32_igemm_bench",
1320 srcs = [
1321 "bench/f32-igemm.cc",
1322 "bench/conv.h",
1323 "src/xnnpack/AlignedAllocator.h",
1324 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07001325 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07001326)
1327
1328xnnpack_benchmark(
1329 name = "f32_conv_hwc_bench",
1330 srcs = [
1331 "bench/f32-conv-hwc.cc",
1332 "bench/dconv.h",
1333 "src/xnnpack/AlignedAllocator.h",
1334 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1335 copts = ["-Wno-unused-function"],
1336 deps = MICROKERNEL_BENCHMARK_DEPS,
1337)
1338
1339xnnpack_benchmark(
Erich Elsen563df5f2019-10-23 08:02:21 -07001340 name = "f32_conv_hwc2spchw_bench",
1341 srcs = [
1342 "bench/f32-conv-hwc2spchw.cc",
1343 "bench/dconv.h",
1344 "src/xnnpack/AlignedAllocator.h",
1345 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1346 copts = ["-Wno-unused-function"],
1347 deps = MICROKERNEL_BENCHMARK_DEPS,
1348)
1349
1350xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001351 name = "f32_dwconv_bench",
1352 srcs = [
1353 "bench/f32-dwconv.cc",
1354 "bench/dwconv.h",
1355 "src/xnnpack/AlignedAllocator.h",
1356 ] + MICROKERNEL_BENCHMARK_HDRS,
1357 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1358)
1359
1360xnnpack_benchmark(
1361 name = "f32_dwconv_spchw_bench",
1362 srcs = [
1363 "bench/f32-dwconv-spchw.cc",
1364 "bench/dwconv.h",
1365 "src/xnnpack/AlignedAllocator.h",
1366 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1367 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
1368)
1369
1370xnnpack_benchmark(
1371 name = "f32_gemm_bench",
1372 srcs = [
1373 "bench/f32-gemm.cc",
1374 "bench/gemm.h",
1375 "src/xnnpack/AlignedAllocator.h",
1376 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1377 copts = ["-Wno-unused-function"] + xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07001378 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001379)
1380
1381xnnpack_benchmark(
1382 name = "f32_rmax_bench",
1383 srcs = [
1384 "bench/f32-rmax.cc",
1385 "src/xnnpack/AlignedAllocator.h",
1386 ] + MICROKERNEL_BENCHMARK_HDRS,
1387 deps = MICROKERNEL_BENCHMARK_DEPS,
1388)
1389
1390xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08001391 name = "f32_sigmoid_bench",
1392 srcs = [
1393 "bench/f32-sigmoid.cc",
1394 "src/xnnpack/AlignedAllocator.h",
1395 ] + MICROKERNEL_BENCHMARK_HDRS,
1396 copts = ["-Wno-unused-function"],
1397 deps = MICROKERNEL_BENCHMARK_DEPS,
1398)
1399
1400xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001401 name = "f32_spmm_bench",
1402 srcs = [
1403 "bench/f32-spmm.cc",
1404 "bench/gemm.h",
1405 "src/xnnpack/AlignedAllocator.h",
1406 ] + MICROKERNEL_BENCHMARK_HDRS,
1407 copts = ["-Wno-unused-function"],
1408 deps = MICROKERNEL_BENCHMARK_DEPS,
1409)
1410
1411xnnpack_benchmark(
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07001412 name = "f32_softargmax_bench",
1413 srcs = [
1414 "bench/f32-softargmax.cc",
1415 ] + MICROKERNEL_BENCHMARK_HDRS,
1416 copts = ["-Wno-unused-function"],
1417 deps = MICROKERNEL_BENCHMARK_DEPS,
1418)
1419
1420xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001421 name = "f32_im2col_gemm_bench",
1422 srcs = [
1423 "bench/f32-im2col-gemm.cc",
1424 "bench/conv.h",
1425 "src/xnnpack/AlignedAllocator.h",
1426 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
1427 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
1428)
1429
1430########################### Benchmarks for operators ###########################
1431
1432xnnpack_benchmark(
1433 name = "add_bench",
1434 srcs = ["bench/add.cc"],
1435 deps = OPERATOR_BENCHMARK_DEPS,
1436)
1437
1438xnnpack_benchmark(
1439 name = "average_pooling_bench",
1440 srcs = ["bench/average-pooling.cc"],
1441 deps = OPERATOR_BENCHMARK_DEPS,
1442)
1443
1444xnnpack_benchmark(
1445 name = "channel_shuffle_bench",
1446 srcs = ["bench/channel-shuffle.cc"],
1447 deps = OPERATOR_BENCHMARK_DEPS,
1448)
1449
1450xnnpack_benchmark(
1451 name = "convolution_bench",
1452 srcs = ["bench/convolution.cc"],
1453 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
1454 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
1455)
1456
1457xnnpack_benchmark(
1458 name = "deconvolution_bench",
1459 srcs = ["bench/deconvolution.cc"],
1460 copts = xnnpack_optional_tflite_copts(),
1461 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1462)
1463
1464xnnpack_benchmark(
1465 name = "global_average_pooling_bench",
1466 srcs = ["bench/global-average-pooling.cc"],
1467 deps = OPERATOR_BENCHMARK_DEPS,
1468)
1469
1470xnnpack_benchmark(
1471 name = "max_pooling_bench",
1472 srcs = ["bench/max-pooling.cc"],
1473 deps = OPERATOR_BENCHMARK_DEPS,
1474)
1475
1476xnnpack_benchmark(
1477 name = "sigmoid_bench",
1478 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08001479 copts = xnnpack_optional_tflite_copts(),
1480 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001481)
1482
1483xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07001484 name = "prelu_bench",
1485 srcs = ["bench/prelu.cc"],
1486 copts = xnnpack_optional_tflite_copts(),
1487 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
1488)
1489
1490xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001491 name = "softargmax_bench",
1492 srcs = ["bench/softargmax.cc"],
1493 deps = OPERATOR_BENCHMARK_DEPS,
1494)
1495
Marat Dukhanc068bb62019-10-04 13:24:39 -07001496############################# End-to-end benchmarks ############################
1497
1498cc_library(
1499 name = "mobilenet_v1",
1500 srcs = ["models/mobilenet-v1.cc"],
1501 hdrs = ["models/models.h"],
1502 linkstatic = True,
1503 deps = [
1504 ":XNNPACK",
1505 "@pthreadpool",
1506 ],
1507)
1508
1509cc_library(
1510 name = "mobilenet_v2",
1511 srcs = ["models/mobilenet-v2.cc"],
1512 hdrs = ["models/models.h"],
1513 linkstatic = True,
1514 deps = [
1515 ":XNNPACK",
1516 "@pthreadpool",
1517 ],
1518)
1519
1520xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07001521 name = "f32_dwconv_e2e_bench",
1522 srcs = ["bench/f32-dwconv-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1523 copts = ["-Wno-unused-function"],
1524 deps = MICROKERNEL_BENCHMARK_DEPS + [
1525 ":XNNPACK",
1526 ":mobilenet_v1",
1527 ":mobilenet_v2",
1528 ],
1529)
1530
1531xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07001532 name = "f32_gemm_e2e_bench",
1533 srcs = ["bench/f32-gemm-e2e.cc"] + MICROKERNEL_BENCHMARK_HDRS,
1534 copts = ["-Wno-unused-function"],
1535 deps = MICROKERNEL_BENCHMARK_DEPS + [
1536 ":XNNPACK",
1537 ":mobilenet_v1",
1538 ":mobilenet_v2",
1539 ],
1540)
1541
1542xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07001543 name = "end2end_bench",
1544 srcs = ["bench/end2end.cc"],
1545 deps = [
1546 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07001547 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07001548 ":mobilenet_v1",
1549 ":mobilenet_v2",
1550 "@pthreadpool",
1551 ],
1552)
1553
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001554#################### Accuracy evaluation for math functions ####################
1555
1556xnnpack_benchmark(
1557 name = "f32_exp_eval",
1558 srcs = [
1559 "eval/f32-exp.cc",
1560 "src/xnnpack/AlignedAllocator.h",
1561 ] + ACCURACY_EVAL_HDRS,
1562 deps = ACCURACY_EVAL_DEPS,
1563)
1564
Marat Dukhan515c9772019-10-17 18:07:57 -07001565xnnpack_benchmark(
1566 name = "f32_expminus_eval",
1567 srcs = [
1568 "eval/f32-expminus.cc",
1569 "src/xnnpack/AlignedAllocator.h",
1570 ] + ACCURACY_EVAL_HDRS,
1571 deps = ACCURACY_EVAL_DEPS,
1572)
1573
Marat Dukhan98ba4412019-10-23 02:14:28 -07001574xnnpack_benchmark(
1575 name = "f32_extexp_eval",
1576 srcs = [
1577 "eval/f32-extexp.cc",
1578 "src/xnnpack/AlignedAllocator.h",
1579 ] + ACCURACY_EVAL_HDRS,
1580 deps = ACCURACY_EVAL_DEPS,
1581)
1582
Marat Dukhan346a9e52019-11-15 09:06:30 -08001583xnnpack_benchmark(
1584 name = "f32_sigmoid_eval",
1585 srcs = [
1586 "eval/f32-sigmoid.cc",
1587 "src/xnnpack/AlignedAllocator.h",
1588 ] + ACCURACY_EVAL_HDRS,
1589 deps = ACCURACY_EVAL_DEPS,
1590)
1591
Marat Dukhan08c4a432019-10-03 09:29:21 -07001592######################### Unit tests for micro-kernels #########################
1593
1594xnnpack_unit_test(
1595 name = "f16_gemm_test",
1596 srcs = [
1597 "test/f16-gemm.cc",
1598 "test/gemm-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_argmaxpool_test",
1606 srcs = [
1607 "test/f32-argmaxpool.cc",
1608 "test/argmaxpool-microkernel-tester.h",
1609 "src/xnnpack/AlignedAllocator.h",
1610 ] + MICROKERNEL_TEST_HDRS,
1611 deps = MICROKERNEL_TEST_DEPS,
1612)
1613
1614xnnpack_unit_test(
1615 name = "f32_avgpool_test",
1616 srcs = [
1617 "test/f32-avgpool.cc",
1618 "test/avgpool-microkernel-tester.h",
1619 "src/xnnpack/AlignedAllocator.h",
1620 ] + MICROKERNEL_TEST_HDRS,
1621 deps = MICROKERNEL_TEST_DEPS,
1622)
1623
1624xnnpack_unit_test(
Marat Dukhan35dacfb2019-11-07 19:18:16 -08001625 name = "f32_bilinear_test",
1626 srcs = [
1627 "test/f32-bilinear.cc",
1628 "test/bilinear-microkernel-tester.h",
1629 "src/xnnpack/AlignedAllocator.h",
1630 ] + MICROKERNEL_TEST_HDRS,
1631 deps = MICROKERNEL_TEST_DEPS,
1632)
1633
1634xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001635 name = "f32_clamp_test",
1636 srcs = [
1637 "test/f32-clamp.cc",
1638 "test/clamp-microkernel-tester.h",
1639 ] + MICROKERNEL_TEST_HDRS,
1640 deps = MICROKERNEL_TEST_DEPS,
1641)
1642
1643xnnpack_unit_test(
1644 name = "f32_igemm_test",
1645 srcs = [
1646 "test/f32-igemm.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_conv_hwc_test",
1655 srcs = [
1656 "test/f32-conv-hwc.cc",
1657 "test/conv-hwc-microkernel-tester.h",
1658 "src/xnnpack/AlignedAllocator.h",
1659 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1660 deps = MICROKERNEL_TEST_DEPS,
1661)
1662
1663xnnpack_unit_test(
1664 name = "f32_conv_hwc2spchw_test",
1665 srcs = [
1666 "test/f32-conv-hwc2spchw.cc",
1667 "test/conv-hwc2spchw-microkernel-tester.h",
1668 "src/xnnpack/AlignedAllocator.h",
1669 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1670 deps = MICROKERNEL_TEST_DEPS,
1671)
1672
1673xnnpack_unit_test(
1674 name = "f32_dwconv_test",
1675 srcs = [
1676 "test/f32-dwconv.cc",
1677 "test/dwconv-microkernel-tester.h",
1678 "src/xnnpack/AlignedAllocator.h",
1679 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1680 deps = MICROKERNEL_TEST_DEPS,
1681)
1682
1683xnnpack_unit_test(
1684 name = "f32_dwconv_spchw_test",
1685 srcs = [
1686 "test/f32-dwconv-spchw.cc",
1687 "test/dwconv-spchw-microkernel-tester.h",
1688 "src/xnnpack/AlignedAllocator.h",
1689 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1690 deps = MICROKERNEL_TEST_DEPS,
1691)
1692
1693xnnpack_unit_test(
1694 name = "f32_gavgpool_test",
1695 srcs = [
1696 "test/f32-gavgpool.cc",
1697 "test/gavgpool-microkernel-tester.h",
1698 "src/xnnpack/AlignedAllocator.h",
1699 ] + MICROKERNEL_TEST_HDRS,
1700 deps = MICROKERNEL_TEST_DEPS,
1701)
1702
1703xnnpack_unit_test(
1704 name = "f32_gavgpool_spchw_test",
1705 srcs = [
1706 "test/f32-gavgpool-spchw.cc",
1707 "test/gavgpool-spchw-microkernel-tester.h",
1708 "src/xnnpack/AlignedAllocator.h",
1709 ] + MICROKERNEL_TEST_HDRS,
1710 deps = MICROKERNEL_TEST_DEPS,
1711)
1712
1713xnnpack_unit_test(
1714 name = "f32_gemm_test",
1715 srcs = [
1716 "test/f32-gemm.cc",
1717 "test/gemm-microkernel-tester.h",
1718 "src/xnnpack/AlignedAllocator.h",
1719 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1720 deps = MICROKERNEL_TEST_DEPS,
1721)
1722
1723xnnpack_unit_test(
1724 name = "f32_gemminc_test",
1725 srcs = [
1726 "test/f32-gemminc.cc",
1727 "test/gemm-microkernel-tester.h",
1728 "src/xnnpack/AlignedAllocator.h",
1729 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1730 deps = MICROKERNEL_TEST_DEPS,
1731)
1732
1733xnnpack_unit_test(
1734 name = "f32_hswish_test",
1735 srcs = [
1736 "test/f32-hswish.cc",
1737 "test/hswish-microkernel-tester.h",
1738 ] + MICROKERNEL_TEST_HDRS,
1739 deps = MICROKERNEL_TEST_DEPS,
1740)
1741
1742xnnpack_unit_test(
1743 name = "f32_maxpool_test",
1744 srcs = [
1745 "test/f32-maxpool.cc",
1746 "test/maxpool-microkernel-tester.h",
1747 ] + MICROKERNEL_TEST_HDRS,
1748 deps = MICROKERNEL_TEST_DEPS,
1749)
1750
1751xnnpack_unit_test(
1752 name = "f32_pavgpool_test",
1753 srcs = [
1754 "test/f32-pavgpool.cc",
1755 "test/avgpool-microkernel-tester.h",
1756 "src/xnnpack/AlignedAllocator.h",
1757 ] + MICROKERNEL_TEST_HDRS,
1758 deps = MICROKERNEL_TEST_DEPS,
1759)
1760
1761xnnpack_unit_test(
1762 name = "f32_ppmm_test",
1763 srcs = [
1764 "test/f32-ppmm.cc",
1765 "test/gemm-microkernel-tester.h",
1766 "src/xnnpack/AlignedAllocator.h",
1767 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1768 deps = MICROKERNEL_TEST_DEPS,
1769)
1770
1771xnnpack_unit_test(
1772 name = "f32_prelu_test",
1773 srcs = [
1774 "test/f32-prelu.cc",
1775 "test/prelu-microkernel-tester.h",
1776 "src/xnnpack/AlignedAllocator.h",
1777 ] + MICROKERNEL_TEST_HDRS,
1778 deps = MICROKERNEL_TEST_DEPS,
1779)
1780
1781xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001782 name = "f32_raddexpminusmax_test",
1783 srcs = [
1784 "test/f32-raddexpminusmax.cc",
1785 "test/raddexpminusmax-microkernel-tester.h",
1786 ] + MICROKERNEL_TEST_HDRS,
1787 deps = MICROKERNEL_TEST_DEPS,
1788)
1789
1790xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001791 name = "f32_raddextexp_test",
1792 srcs = [
1793 "test/f32-raddextexp.cc",
1794 "test/raddextexp-microkernel-tester.h",
1795 ] + MICROKERNEL_TEST_HDRS,
1796 deps = MICROKERNEL_TEST_DEPS,
1797)
1798
1799xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001800 name = "f32_raddstoreexpminusmax_test",
1801 srcs = [
1802 "test/f32-raddstoreexpminusmax.cc",
1803 "test/raddstoreexpminusmax-microkernel-tester.h",
1804 ] + MICROKERNEL_TEST_HDRS,
1805 deps = MICROKERNEL_TEST_DEPS,
1806)
1807
1808xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001809 name = "f32_rmax_test",
1810 srcs = [
1811 "test/f32-rmax.cc",
1812 "test/rmax-microkernel-tester.h",
1813 ] + MICROKERNEL_TEST_HDRS,
1814 deps = MICROKERNEL_TEST_DEPS,
1815)
1816
1817xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08001818 name = "f32_sigmoid_test",
1819 srcs = [
1820 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001821 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001822 ] + MICROKERNEL_TEST_HDRS,
1823 deps = MICROKERNEL_TEST_DEPS,
1824)
1825
1826xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001827 name = "f32_spmm_test",
1828 srcs = [
1829 "test/f32-spmm.cc",
1830 "test/spmm-microkernel-tester.h",
1831 "src/xnnpack/AlignedAllocator.h",
1832 ] + MICROKERNEL_TEST_HDRS,
1833 deps = MICROKERNEL_TEST_DEPS,
1834)
1835
1836xnnpack_unit_test(
1837 name = "f32_vadd_test",
1838 srcs = [
1839 "test/f32-vadd.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001840 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001841 ] + MICROKERNEL_TEST_HDRS,
1842 deps = MICROKERNEL_TEST_DEPS,
1843)
1844
1845xnnpack_unit_test(
1846 name = "f32_vaddc_test",
1847 srcs = [
1848 "test/f32-vaddc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001849 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001850 ] + MICROKERNEL_TEST_HDRS,
1851 deps = MICROKERNEL_TEST_DEPS,
1852)
1853
1854xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001855 name = "f32_vmul_test",
1856 srcs = [
1857 "test/f32-vmul.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001858 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001859 ] + MICROKERNEL_TEST_HDRS,
1860 deps = MICROKERNEL_TEST_DEPS,
1861)
1862
1863xnnpack_unit_test(
1864 name = "f32_vmulc_test",
1865 srcs = [
1866 "test/f32-vmulc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001867 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001868 ] + MICROKERNEL_TEST_HDRS,
1869 deps = MICROKERNEL_TEST_DEPS,
1870)
1871
1872xnnpack_unit_test(
1873 name = "f32_vmulcaddc_test",
1874 srcs = [
1875 "test/f32-vmulcaddc.cc",
1876 "test/vmulcaddc-microkernel-tester.h",
1877 "src/xnnpack/AlignedAllocator.h",
1878 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1879 deps = MICROKERNEL_TEST_DEPS,
1880)
1881
1882xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001883 name = "f32_vscale_test",
1884 srcs = [
1885 "test/f32-vscale.cc",
1886 "test/vscale-microkernel-tester.h",
1887 ] + MICROKERNEL_TEST_HDRS,
1888 deps = MICROKERNEL_TEST_DEPS,
1889)
1890
1891xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001892 name = "f32_vscaleexpminusmax_test",
1893 srcs = [
1894 "test/f32-vscaleexpminusmax.cc",
1895 "test/vscaleexpminusmax-microkernel-tester.h",
1896 ] + MICROKERNEL_TEST_HDRS,
1897 deps = MICROKERNEL_TEST_DEPS,
1898)
1899
1900xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001901 name = "f32_vscaleextexp_test",
1902 srcs = [
1903 "test/f32-vscaleextexp.cc",
1904 "test/vscaleextexp-microkernel-tester.h",
1905 ] + MICROKERNEL_TEST_HDRS,
1906 deps = MICROKERNEL_TEST_DEPS,
1907)
1908
1909xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07001910 name = "f32_vsub_test",
1911 srcs = [
1912 "test/f32-vsub.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001913 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001914 ] + MICROKERNEL_TEST_HDRS,
1915 deps = MICROKERNEL_TEST_DEPS,
1916)
1917
1918xnnpack_unit_test(
1919 name = "f32_vsubc_test",
1920 srcs = [
1921 "test/f32-vsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001922 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08001923 ] + MICROKERNEL_TEST_HDRS,
1924 deps = MICROKERNEL_TEST_DEPS,
1925)
1926
1927xnnpack_unit_test(
1928 name = "f32_vrsubc_test",
1929 srcs = [
1930 "test/f32-vrsubc.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001931 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001932 ] + MICROKERNEL_TEST_HDRS,
1933 deps = MICROKERNEL_TEST_DEPS,
1934)
1935
1936xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001937 name = "q8_avgpool_test",
1938 srcs = [
1939 "test/q8-avgpool.cc",
1940 "test/avgpool-microkernel-tester.h",
1941 "src/xnnpack/AlignedAllocator.h",
1942 ] + MICROKERNEL_TEST_HDRS,
1943 deps = MICROKERNEL_TEST_DEPS,
1944)
1945
1946xnnpack_unit_test(
1947 name = "q8_igemm_test",
1948 srcs = [
1949 "test/q8-igemm.cc",
1950 "test/gemm-microkernel-tester.h",
1951 "src/xnnpack/AlignedAllocator.h",
1952 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1953 deps = MICROKERNEL_TEST_DEPS,
1954)
1955
1956xnnpack_unit_test(
1957 name = "q8_dwconv_test",
1958 srcs = [
1959 "test/q8-dwconv.cc",
1960 "test/dwconv-microkernel-tester.h",
1961 "src/xnnpack/AlignedAllocator.h",
1962 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1963 deps = MICROKERNEL_TEST_DEPS,
1964)
1965
1966xnnpack_unit_test(
1967 name = "q8_gavgpool_test",
1968 srcs = [
1969 "test/q8-gavgpool.cc",
1970 "test/gavgpool-microkernel-tester.h",
1971 "src/xnnpack/AlignedAllocator.h",
1972 ] + MICROKERNEL_TEST_HDRS,
1973 deps = MICROKERNEL_TEST_DEPS,
1974)
1975
1976xnnpack_unit_test(
1977 name = "q8_gemm_test",
1978 srcs = [
1979 "test/q8-gemm.cc",
1980 "test/gemm-microkernel-tester.h",
1981 "src/xnnpack/AlignedAllocator.h",
1982 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
1983 deps = MICROKERNEL_TEST_DEPS,
1984)
1985
1986xnnpack_unit_test(
1987 name = "q8_vadd_test",
1988 srcs = [
1989 "test/q8-vadd.cc",
1990 "test/vadd-microkernel-tester.h",
1991 ] + MICROKERNEL_TEST_HDRS,
1992 deps = MICROKERNEL_TEST_DEPS,
1993)
1994
1995xnnpack_unit_test(
1996 name = "u8_clamp_test",
1997 srcs = [
1998 "test/u8-clamp.cc",
1999 "test/clamp-microkernel-tester.h",
2000 ] + MICROKERNEL_TEST_HDRS,
2001 deps = MICROKERNEL_TEST_DEPS,
2002)
2003
2004xnnpack_unit_test(
2005 name = "u8_lut32norm_test",
2006 srcs = [
2007 "test/u8-lut32norm.cc",
2008 "test/lut-norm-microkernel-tester.h",
2009 ] + MICROKERNEL_TEST_HDRS,
2010 deps = MICROKERNEL_TEST_DEPS,
2011)
2012
2013xnnpack_unit_test(
2014 name = "u8_maxpool_test",
2015 srcs = [
2016 "test/u8-maxpool.cc",
2017 "test/maxpool-microkernel-tester.h",
2018 ] + MICROKERNEL_TEST_HDRS,
2019 deps = MICROKERNEL_TEST_DEPS,
2020)
2021
2022xnnpack_unit_test(
2023 name = "u8_rmax_test",
2024 srcs = [
2025 "test/u8-rmax.cc",
2026 "test/rmax-microkernel-tester.h",
2027 ] + MICROKERNEL_TEST_HDRS,
2028 deps = MICROKERNEL_TEST_DEPS,
2029)
2030
2031xnnpack_unit_test(
2032 name = "x32_packx_test",
2033 srcs = [
2034 "test/x32-packx.cc",
2035 "test/pack-microkernel-tester.h",
2036 "src/xnnpack/AlignedAllocator.h",
2037 ] + MICROKERNEL_TEST_HDRS,
2038 deps = MICROKERNEL_TEST_DEPS,
2039)
2040
2041xnnpack_unit_test(
2042 name = "x32_pad_test",
2043 srcs = [
2044 "test/x32-pad.cc",
2045 "test/pad-microkernel-tester.h",
2046 ] + MICROKERNEL_TEST_HDRS,
2047 deps = MICROKERNEL_TEST_DEPS,
2048)
2049
2050xnnpack_unit_test(
2051 name = "x32_unpool_test",
2052 srcs = [
2053 "test/x32-unpool.cc",
2054 "test/unpool-microkernel-tester.h",
2055 ] + MICROKERNEL_TEST_HDRS,
2056 deps = MICROKERNEL_TEST_DEPS,
2057)
2058
2059xnnpack_unit_test(
2060 name = "x32_zip_test",
2061 srcs = [
2062 "test/x32-zip.cc",
2063 "test/zip-microkernel-tester.h",
2064 ] + MICROKERNEL_TEST_HDRS,
2065 deps = MICROKERNEL_TEST_DEPS,
2066)
2067
2068xnnpack_unit_test(
2069 name = "x8_lut_test",
2070 srcs = [
2071 "test/x8-lut.cc",
2072 "test/lut-microkernel-tester.h",
2073 ] + MICROKERNEL_TEST_HDRS,
2074 deps = MICROKERNEL_TEST_DEPS,
2075)
2076
2077xnnpack_unit_test(
2078 name = "x8_zip_test",
2079 srcs = [
2080 "test/x8-zip.cc",
2081 "test/zip-microkernel-tester.h",
2082 ] + MICROKERNEL_TEST_HDRS,
2083 deps = MICROKERNEL_TEST_DEPS,
2084)
2085
2086########################### Size test for the library ##########################
2087
2088xnnpack_binary(
2089 name = "size_test",
2090 srcs = ["test/size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002091 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002092)
2093
2094########################### Unit tests for operators ###########################
2095
2096xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002097 name = "add_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002098 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002099 "test/add-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002100 "test/add-operator-tester.h",
2101 ],
2102 deps = OPERATOR_TEST_DEPS,
2103)
2104
2105xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08002106 name = "add_nd_test",
2107 srcs = [
2108 "test/add-nd.cc",
2109 "test/binary-elementwise-operator-tester.h",
2110 ],
2111 deps = OPERATOR_TEST_DEPS,
2112)
2113
2114xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002115 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002116 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002117 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002118 "test/argmax-pooling-operator-tester.h",
2119 ] + OPERATOR_TEST_PARAMS_HDRS,
2120 deps = OPERATOR_TEST_DEPS,
2121)
2122
2123xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002124 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002125 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002126 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002127 "test/average-pooling-operator-tester.h",
2128 ] + OPERATOR_TEST_PARAMS_HDRS,
2129 deps = OPERATOR_TEST_DEPS,
2130)
2131
2132xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002133 name = "channel_pad_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002134 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002135 "test/channel-pad-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002136 "test/channel-pad-operator-tester.h",
2137 ] + OPERATOR_TEST_PARAMS_HDRS,
2138 deps = OPERATOR_TEST_DEPS,
2139)
2140
2141xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002142 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002143 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002144 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002145 "test/channel-shuffle-operator-tester.h",
2146 ],
2147 deps = OPERATOR_TEST_DEPS,
2148)
2149
2150xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002151 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002152 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002153 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002154 "test/clamp-operator-tester.h",
2155 ],
2156 deps = OPERATOR_TEST_DEPS,
2157)
2158
2159xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002160 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002161 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002162 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002163 "test/convolution-operator-tester.h",
2164 ],
2165 deps = OPERATOR_TEST_DEPS,
2166)
2167
2168xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002169 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002170 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002171 "test/convolution-nchw.cc",
2172 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002173 ],
2174 deps = OPERATOR_TEST_DEPS,
2175)
2176
2177xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002178 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002179 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002180 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002181 "test/deconvolution-operator-tester.h",
2182 ] + OPERATOR_TEST_PARAMS_HDRS,
2183 deps = OPERATOR_TEST_DEPS,
2184)
2185
2186xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002187 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002188 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002189 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002190 "test/fully-connected-operator-tester.h",
2191 ],
2192 deps = OPERATOR_TEST_DEPS,
2193)
2194
2195xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002196 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002197 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002198 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002199 "test/global-average-pooling-operator-tester.h",
2200 ] + OPERATOR_TEST_PARAMS_HDRS,
2201 deps = OPERATOR_TEST_DEPS,
2202)
2203
2204xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002205 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002206 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002207 "test/global-average-pooling-ncw.cc",
2208 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002209 ],
2210 deps = OPERATOR_TEST_DEPS,
2211)
2212
2213xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002214 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002215 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002216 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002217 "test/hardswish-operator-tester.h",
2218 ],
2219 deps = OPERATOR_TEST_DEPS,
2220)
2221
2222xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002223 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002224 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002225 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002226 "test/leaky-relu-operator-tester.h",
2227 ],
2228 deps = OPERATOR_TEST_DEPS,
2229)
2230
2231xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002232 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002233 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002234 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002235 "test/max-pooling-operator-tester.h",
2236 ] + OPERATOR_TEST_PARAMS_HDRS,
2237 deps = OPERATOR_TEST_DEPS,
2238)
2239
2240xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002241 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08002242 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08002243 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002244 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08002245 ],
2246 deps = OPERATOR_TEST_DEPS,
2247)
2248
2249xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002250 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002251 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002252 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002253 "test/prelu-operator-tester.h",
2254 ] + OPERATOR_TEST_PARAMS_HDRS,
2255 deps = OPERATOR_TEST_DEPS,
2256)
2257
2258xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002259 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08002260 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002261 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08002262 "test/resize-bilinear-operator-tester.h",
2263 ] + OPERATOR_TEST_PARAMS_HDRS,
2264 deps = OPERATOR_TEST_DEPS,
2265)
2266
2267xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002268 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002269 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002270 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002271 "test/sigmoid-operator-tester.h",
2272 ],
2273 deps = OPERATOR_TEST_DEPS,
2274)
2275
2276xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002277 name = "softargmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002278 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002279 "test/softargmax-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002280 "test/softargmax-operator-tester.h",
2281 ],
2282 deps = OPERATOR_TEST_DEPS,
2283)
2284
2285xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08002286 name = "subtract_nd_test",
2287 srcs = [
2288 "test/binary-elementwise-operator-tester.h",
2289 "test/subtract-nd.cc",
2290 ],
2291 deps = OPERATOR_TEST_DEPS,
2292)
2293
2294xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08002295 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002296 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08002297 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002298 "test/unpooling-operator-tester.h",
2299 ],
2300 deps = OPERATOR_TEST_DEPS,
2301)
2302
2303############################# Build configurations #############################
2304
Marat Dukhanb8642352019-10-30 15:43:02 -07002305# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07002306config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07002307 name = "xnn_enable_assembly_explicit_true",
2308 define_values = {"xnn_enable_assembly": "true"},
2309)
2310
2311# Disables usage of assembly kernels.
2312config_setting(
2313 name = "xnn_enable_assembly_explicit_false",
2314 define_values = {"xnn_enable_assembly": "false"},
2315)
2316
2317# Builds with -c dbg
2318config_setting(
2319 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002320 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07002321 "compilation_mode": "dbg",
2322 },
2323)
2324
2325# Builds with -c opt
2326config_setting(
2327 name = "optimized_build",
2328 values = {
2329 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002330 },
2331)
2332
2333config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07002334 name = "linux_k8",
2335 values = {"cpu": "k8"},
2336)
2337
2338config_setting(
Marat Dukhan4e45e662019-10-03 15:40:24 -07002339 name = "linux_aarch64",
Marat Dukhanb8642352019-10-30 15:43:02 -07002340 values = {"cpu": "aarch64"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07002341)
2342
2343config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002344 name = "android",
2345 values = {"crosstool_top": "//external:android/crosstool"},
2346)
2347
2348config_setting(
2349 name = "android_armv7",
2350 values = {
2351 "crosstool_top": "//external:android/crosstool",
2352 "cpu": "armeabi-v7a",
2353 },
2354)
2355
2356config_setting(
2357 name = "android_arm64",
2358 values = {
2359 "crosstool_top": "//external:android/crosstool",
2360 "cpu": "arm64-v8a",
2361 },
2362)
2363
2364config_setting(
2365 name = "android_x86",
2366 values = {
2367 "crosstool_top": "//external:android/crosstool",
2368 "cpu": "x86",
2369 },
2370)
2371
2372config_setting(
2373 name = "android_x86_64",
2374 values = {
2375 "crosstool_top": "//external:android/crosstool",
2376 "cpu": "x86_64",
2377 },
2378)
2379
2380config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07002381 name = "macos_x86_64",
2382 values = {
2383 "apple_platform_type": "macos",
2384 "cpu": "darwin",
2385 },
2386)
2387
2388config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002389 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002390 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07002391)
2392
2393config_setting(
2394 name = "emscripten_wasm",
2395 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002396 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002397 "cpu": "wasm",
2398 },
2399)
2400
2401config_setting(
2402 name = "emscripten_wasmsimd",
2403 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002404 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002405 "cpu": "wasm",
Marat Dukhan8c19e3c2019-10-30 12:14:58 -07002406 "features": "wasm_simd",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002407 },
2408)
2409
2410config_setting(
2411 name = "emscripten_asmjs",
2412 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07002413 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002414 "cpu": "asmjs",
2415 },
2416)