blob: 3d7099ea7f67d526ce80ddf0fd3108d5523cb1c2 [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 Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_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_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
19 "@pthreadpool",
20]
21
Marat Dukhan08c4a432019-10-03 09:29:21 -070022MICROKERNEL_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 = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070039 ":ukernels_test_mode",
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
Marat Dukhan1b354632020-03-23 12:50:22 -070046OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070047 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070048 "@pthreadpool",
49 "@FP16",
50]
51
Marat Dukhan08c4a432019-10-03 09:29:21 -070052OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070053 "src/operators/argmax-pooling-nhwc.c",
54 "src/operators/average-pooling-nhwc.c",
55 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070056 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070057 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070058 "src/operators/convolution-nchw.c",
59 "src/operators/convolution-nhwc.c",
60 "src/operators/deconvolution-nhwc.c",
61 "src/operators/fully-connected-nc.c",
62 "src/operators/global-average-pooling-ncw.c",
63 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070064 "src/operators/leaky-relu-nc.c",
65 "src/operators/max-pooling-nhwc.c",
66 "src/operators/prelu-nc.c",
67 "src/operators/resize-bilinear-nhwc.c",
68 "src/operators/sigmoid-nc.c",
69 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070070 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070071 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070072]
73
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070074SUBGRAPH_SRCS = [
75 "src/subgraph/add2.c",
76 "src/subgraph/argmax-pooling-2d.c",
77 "src/subgraph/average-pooling-2d.c",
78 "src/subgraph/clamp.c",
79 "src/subgraph/convolution-2d.c",
80 "src/subgraph/deconvolution-2d.c",
81 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070082 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070083 "src/subgraph/fully-connected.c",
84 "src/subgraph/hardswish.c",
85 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/maximum2.c",
87 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070088 "src/subgraph/multiply2.c",
89 "src/subgraph/prelu.c",
90 "src/subgraph/sigmoid.c",
91 "src/subgraph/softmax.c",
92 "src/subgraph/static-constant-pad.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/squared-difference.c",
94 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/unpooling-2d.c",
96]
97
Marat Dukhan3a77ea72019-12-23 12:10:24 -080098TABLE_SRCS = [
99 "src/tables/exp2-k-over-64.c",
100 "src/tables/exp2-k-over-2048.c",
101]
102
Marat Dukhan08c4a432019-10-03 09:29:21 -0700103SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800104 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800105 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800106 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700107 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
108 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700109 "src/f32-clamp/gen/scalar-x1.c",
110 "src/f32-clamp/gen/scalar-x2.c",
111 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -0800112 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800113 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700114 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
115 "src/f32-dwconv-chw/3x3p1-scalar.c",
116 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
117 "src/f32-dwconv-chw/5x5p2-scalar.c",
118 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700119 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
120 "src/f32-dwconv/gen/up1x4-scalar.c",
121 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
122 "src/f32-dwconv/gen/up1x9-scalar.c",
123 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
124 "src/f32-dwconv/gen/up1x25-scalar.c",
125 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
126 "src/f32-dwconv/gen/up2x4-scalar.c",
127 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
128 "src/f32-dwconv/gen/up2x9-scalar.c",
129 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
130 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700131 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
132 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
133 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
134 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700135 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
136 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700137 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
138 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
139 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
140 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700141 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
142 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700143 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700144 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
145 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700146 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
147 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
148 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700149 "src/f32-gemm/gen/1x4-scalar.c",
150 "src/f32-gemm/gen/2x4-scalar.c",
151 "src/f32-gemm/gen/4x2-scalar.c",
152 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700153 "src/f32-gemm/gen/1x4-relu-scalar.c",
154 "src/f32-gemm/gen/2x4-relu-scalar.c",
155 "src/f32-gemm/gen/4x2-relu-scalar.c",
156 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700157 "src/f32-gemm/gen/1x4-minmax-scalar.c",
158 "src/f32-gemm/gen/2x4-minmax-scalar.c",
159 "src/f32-gemm/gen/4x2-minmax-scalar.c",
160 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800161 "src/f32-hswish/gen/scalar-x1.c",
162 "src/f32-hswish/gen/scalar-x2.c",
163 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700164 "src/f32-ibilinear/gen/scalar-c1.c",
165 "src/f32-ibilinear/gen/scalar-c2.c",
166 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700167 "src/f32-igemm/gen/1x4-scalar.c",
168 "src/f32-igemm/gen/2x4-scalar.c",
169 "src/f32-igemm/gen/4x2-scalar.c",
170 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700171 "src/f32-igemm/gen/1x4-relu-scalar.c",
172 "src/f32-igemm/gen/2x4-relu-scalar.c",
173 "src/f32-igemm/gen/4x2-relu-scalar.c",
174 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700175 "src/f32-igemm/gen/1x4-minmax-scalar.c",
176 "src/f32-igemm/gen/2x4-minmax-scalar.c",
177 "src/f32-igemm/gen/4x2-minmax-scalar.c",
178 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700179 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
180 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
181 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700182 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
183 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
184 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
185 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800186 "src/f32-prelu/gen/scalar-2x1.c",
187 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800188 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
189 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
190 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
191 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
192 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
193 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
194 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
195 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
196 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
197 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
198 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
199 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700200 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800201 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
202 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
203 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
204 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
205 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
206 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
207 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
208 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
209 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700210 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
211 "src/f32-spmm/gen/1x1-minmax-scalar.c",
212 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
213 "src/f32-spmm/gen/2x1-minmax-scalar.c",
214 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
215 "src/f32-spmm/gen/4x1-minmax-scalar.c",
216 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
217 "src/f32-spmm/gen/8x1-minmax-scalar.c",
218 "src/f32-spmm/gen/8x2-minmax-scalar.c",
219 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700220 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
221 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
222 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
223 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
224 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
225 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
226 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
227 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
228 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
229 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
230 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
231 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800232 "src/f32-vbinary/gen/vmax-scalar-x1.c",
233 "src/f32-vbinary/gen/vmax-scalar-x2.c",
234 "src/f32-vbinary/gen/vmax-scalar-x4.c",
235 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
236 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
237 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
238 "src/f32-vbinary/gen/vmin-scalar-x1.c",
239 "src/f32-vbinary/gen/vmin-scalar-x2.c",
240 "src/f32-vbinary/gen/vmin-scalar-x4.c",
241 "src/f32-vbinary/gen/vminc-scalar-x1.c",
242 "src/f32-vbinary/gen/vminc-scalar-x2.c",
243 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700244 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
245 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
246 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
247 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
248 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
249 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
250 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
251 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
252 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700253 "src/f32-vbinary/gen/vrsqrdiffc-scalar-x1.c",
254 "src/f32-vbinary/gen/vrsqrdiffc-scalar-x2.c",
255 "src/f32-vbinary/gen/vrsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700256 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
257 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
258 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700259 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
260 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
261 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
262 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
263 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
264 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700265 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
266 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
267 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
268 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
269 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
270 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700271 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
272 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
273 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700274 "src/f32-vunary/gen/vabs-scalar-x1.c",
275 "src/f32-vunary/gen/vabs-scalar-x2.c",
276 "src/f32-vunary/gen/vabs-scalar-x4.c",
277 "src/f32-vunary/gen/vneg-scalar-x1.c",
278 "src/f32-vunary/gen/vneg-scalar-x2.c",
279 "src/f32-vunary/gen/vneg-scalar-x4.c",
280 "src/f32-vunary/gen/vsqr-scalar-x1.c",
281 "src/f32-vunary/gen/vsqr-scalar-x2.c",
282 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800283 "src/math/expminus-scalar-lut2048-p1.c",
284 "src/math/expminus-scalar-lut64-p2.c",
285 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700286 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700287 "src/math/roundne-scalar-nearbyint.c",
288 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700289 "src/math/roundd-scalar-addsub.c",
290 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700291 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700292 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700293 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700294 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700295 "src/math/roundz-scalar-addsub.c",
296 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700297 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800298 "src/math/sigmoid-scalar-lut2048-p1-div.c",
299 "src/math/sigmoid-scalar-lut64-p2-div.c",
300 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700301 "src/q8-avgpool/9p8x-minmax-scalar-c1.c",
302 "src/q8-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700303 "src/q8-dwconv/up1x9-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700304 "src/q8-gavgpool/7p7x-minmax-scalar-c1.c",
305 "src/q8-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700306 "src/q8-gemm/2x2-minmax-scalar.c",
307 "src/q8-igemm/2x2-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700308 "src/q8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700309 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700310 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700311 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700312 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700313 "src/x32-fill/scalar-float.c",
314 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700315 "src/x32-packx/x2-scalar.c",
316 "src/x32-packx/x3-scalar.c",
317 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700318 "src/x32-pad/scalar-int.c",
319 "src/x32-pad/scalar-float.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700320 "src/x32-unpool/scalar.c",
321 "src/x32-zip/x2-scalar.c",
322 "src/x32-zip/x3-scalar.c",
323 "src/x32-zip/x4-scalar.c",
324 "src/x32-zip/xm-scalar.c",
325 "src/x8-lut/scalar.c",
326 "src/x8-zip/x2-scalar.c",
327 "src/x8-zip/x3-scalar.c",
328 "src/x8-zip/x4-scalar.c",
329 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700330 "src/requantization/precise-scalar.c",
331 "src/requantization/fp32-scalar.c",
332 "src/requantization/q31-scalar.c",
333 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700334]
335
Marat Dukhan436ebe62019-12-04 15:10:12 -0800336WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700337 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
338 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700339 "src/f32-clamp/gen/wasm-x1.c",
340 "src/f32-clamp/gen/wasm-x2.c",
341 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700342 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
343 "src/f32-dwconv/gen/up1x4-wasm.c",
344 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
345 "src/f32-dwconv/gen/up1x9-wasm.c",
346 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
347 "src/f32-dwconv/gen/up1x25-wasm.c",
348 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
349 "src/f32-dwconv/gen/up2x4-wasm.c",
350 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
351 "src/f32-dwconv/gen/up2x9-wasm.c",
352 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
353 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700354 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
355 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
356 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
357 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700358 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
359 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700360 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
361 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
362 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
363 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700364 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
365 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700366 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
367 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700368 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
369 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
370 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700371 "src/f32-gemm/gen/1x4-wasm.c",
372 "src/f32-gemm/gen/2x4-wasm.c",
373 "src/f32-gemm/gen/4x2-wasm.c",
374 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700375 "src/f32-gemm/gen/1x4-relu-wasm.c",
376 "src/f32-gemm/gen/2x4-relu-wasm.c",
377 "src/f32-gemm/gen/4x2-relu-wasm.c",
378 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700379 "src/f32-gemm/gen/1x4-minmax-wasm.c",
380 "src/f32-gemm/gen/2x4-minmax-wasm.c",
381 "src/f32-gemm/gen/4x2-minmax-wasm.c",
382 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800383 "src/f32-hswish/gen/wasm-x1.c",
384 "src/f32-hswish/gen/wasm-x2.c",
385 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700386 "src/f32-igemm/gen/1x4-wasm.c",
387 "src/f32-igemm/gen/2x4-wasm.c",
388 "src/f32-igemm/gen/4x2-wasm.c",
389 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700390 "src/f32-igemm/gen/1x4-relu-wasm.c",
391 "src/f32-igemm/gen/2x4-relu-wasm.c",
392 "src/f32-igemm/gen/4x2-relu-wasm.c",
393 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700394 "src/f32-igemm/gen/1x4-minmax-wasm.c",
395 "src/f32-igemm/gen/2x4-minmax-wasm.c",
396 "src/f32-igemm/gen/4x2-minmax-wasm.c",
397 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700398 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
399 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
400 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700401 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
402 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
403 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
404 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
405 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
406 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
407 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
408 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
409 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
410 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
411 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
412 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800413 "src/f32-vbinary/gen/vmax-wasm-x1.c",
414 "src/f32-vbinary/gen/vmax-wasm-x2.c",
415 "src/f32-vbinary/gen/vmax-wasm-x4.c",
416 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
417 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
418 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
419 "src/f32-vbinary/gen/vmin-wasm-x1.c",
420 "src/f32-vbinary/gen/vmin-wasm-x2.c",
421 "src/f32-vbinary/gen/vmin-wasm-x4.c",
422 "src/f32-vbinary/gen/vminc-wasm-x1.c",
423 "src/f32-vbinary/gen/vminc-wasm-x2.c",
424 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700425 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
426 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
427 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
428 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
429 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
430 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
431 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
432 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
433 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
434 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
435 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
436 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
437 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
438 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
439 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
440 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
441 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
442 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700443 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
444 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
445 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800446]
447
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800448PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800449 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800450 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800451 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700452 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
453 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700454 "src/f32-clamp/gen/psimd-x4.c",
455 "src/f32-clamp/gen/psimd-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700456 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
457 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
458 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
459 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
460 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
461 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
462 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
463 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
464 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
465 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
466 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
467 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700468 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
469 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700470 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
471 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
472 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
473 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
474 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
475 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
476 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
477 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
478 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
479 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
480 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
481 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
482 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
483 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
484 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
485 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
486 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
487 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
488 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800489 "src/f32-hswish/gen/psimd-x4.c",
490 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700491 "src/f32-ibilinear/gen/psimd-c4.c",
492 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700493 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
494 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
495 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
496 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
497 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
498 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
499 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
500 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
501 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
502 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700503 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
504 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
505 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700506 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800507 "src/f32-prelu/gen/psimd-2x4.c",
508 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800509 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700510 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
511 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
512 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
513 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
514 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
515 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
516 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
517 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800518 "src/f32-vbinary/gen/vmax-psimd-x4.c",
519 "src/f32-vbinary/gen/vmax-psimd-x8.c",
520 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
521 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
522 "src/f32-vbinary/gen/vmin-psimd-x4.c",
523 "src/f32-vbinary/gen/vmin-psimd-x8.c",
524 "src/f32-vbinary/gen/vminc-psimd-x4.c",
525 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700526 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
527 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
528 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
529 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
530 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
531 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700532 "src/f32-vbinary/gen/vrsqrdiffc-psimd-x4.c",
533 "src/f32-vbinary/gen/vrsqrdiffc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700534 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
535 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700536 "src/f32-vbinary/gen/vsqrdiff-psimd-x4.c",
537 "src/f32-vbinary/gen/vsqrdiff-psimd-x8.c",
538 "src/f32-vbinary/gen/vsqrdiffc-psimd-x4.c",
539 "src/f32-vbinary/gen/vsqrdiffc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700540 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
541 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
542 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
543 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700544 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
545 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700546 "src/f32-vunary/gen/vabs-psimd-x4.c",
547 "src/f32-vunary/gen/vabs-psimd-x8.c",
548 "src/f32-vunary/gen/vneg-psimd-x4.c",
549 "src/f32-vunary/gen/vneg-psimd-x8.c",
550 "src/f32-vunary/gen/vsqr-psimd-x4.c",
551 "src/f32-vunary/gen/vsqr-psimd-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700552 "src/x32-fill/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700553 "src/x32-packx/x4-psimd.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700554 "src/x32-pad/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700555 "src/x32-unpool/psimd.c",
556 "src/x32-zip/x2-psimd.c",
557 "src/x32-zip/x3-psimd.c",
558 "src/x32-zip/x4-psimd.c",
559 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700560 "src/requantization/precise-psimd.c",
561 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700562]
563
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800564PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800565 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
566 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
567 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
568 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
569 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
570 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
571 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
572 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
573 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
574 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
575 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
576 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800577 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
578 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
579 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
580 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
581 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
582 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700583 "src/math/roundne-psimd-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700584 "src/math/roundd-psimd-addsub.c",
585 "src/math/roundu-psimd-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700586 "src/math/roundz-psimd-addsub.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800587 "src/math/sigmoid-psimd-p5-div.c",
588]
589
Marat Dukhan08c4a432019-10-03 09:29:21 -0700590# ISA-specific micro-kernels
591NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700592 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
593 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700594 "src/f32-clamp/gen/neon-x4.c",
595 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700596 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
597 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700598 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
599 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700600 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
601 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700602 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
603 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700604 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
605 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
606 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
607 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700608 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
609 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
610 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
611 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700612 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
613 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
614 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
615 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700616 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700617 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
618 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700619 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
620 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
621 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
622 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
623 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
624 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
625 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
626 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
627 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
628 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
629 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
630 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
631 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
632 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
633 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
634 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
635 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
636 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
637 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
638 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
639 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
640 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
641 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
642 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
643 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
644 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
645 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
646 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
647 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
648 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
649 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800650 "src/f32-hswish/gen/neon-x4.c",
651 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700652 "src/f32-ibilinear/gen/neon-c4.c",
653 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700654 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
655 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
656 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
657 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
658 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
659 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
660 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
661 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
662 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
663 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
664 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
665 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
666 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
667 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
668 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
669 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700670 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
671 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
672 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700673 "src/f32-ppmm/gen/4x8-minmax-neon.c",
674 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800675 "src/f32-prelu/gen/neon-2x4.c",
676 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800677 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
678 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
679 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
680 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
681 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
682 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
683 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
684 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
685 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
686 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
687 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
688 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
689 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
690 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
691 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
692 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
693 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
694 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
695 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
696 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
697 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
698 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
699 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
700 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700701 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800702 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800703 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
704 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
705 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
706 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
707 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
708 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
709 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
710 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
711 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
712 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
713 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
714 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
715 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
716 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
717 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
718 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
719 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
720 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700721 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
722 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
723 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
724 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800725 "src/f32-vbinary/gen/vmax-neon-x4.c",
726 "src/f32-vbinary/gen/vmax-neon-x8.c",
727 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
728 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
729 "src/f32-vbinary/gen/vmin-neon-x4.c",
730 "src/f32-vbinary/gen/vmin-neon-x8.c",
731 "src/f32-vbinary/gen/vminc-neon-x4.c",
732 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700733 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
734 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
735 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
736 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700737 "src/f32-vbinary/gen/vrsqrdiffc-neon-x4.c",
738 "src/f32-vbinary/gen/vrsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700739 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
740 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700741 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
742 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
743 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
744 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700745 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
746 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
747 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
748 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700749 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
750 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700751 "src/f32-vunary/gen/vabs-neon-x4.c",
752 "src/f32-vunary/gen/vabs-neon-x8.c",
753 "src/f32-vunary/gen/vneg-neon-x4.c",
754 "src/f32-vunary/gen/vneg-neon-x8.c",
755 "src/f32-vunary/gen/vsqr-neon-x4.c",
756 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700757 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
758 "src/q8-avgpool/9x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700759 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700760 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
761 "src/q8-gavgpool/7x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700762 "src/q8-gemm/4x8-minmax-neon.c",
763 "src/q8-gemm/8x8-minmax-neon.c",
764 "src/q8-igemm/4x8-minmax-neon.c",
765 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700766 "src/q8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700767 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700768 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700769 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700770 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700771 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700772 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -0700773 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700774 "src/x32-zip/x2-neon.c",
775 "src/x32-zip/x3-neon.c",
776 "src/x32-zip/x4-neon.c",
777 "src/x32-zip/xm-neon.c",
778 "src/x8-zip/x2-neon.c",
779 "src/x8-zip/x3-neon.c",
780 "src/x8-zip/x4-neon.c",
781 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700782 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700783 "src/math/roundd-neon-addsub.c",
784 "src/math/roundd-neon-cvt.c",
785 "src/math/roundu-neon-addsub.c",
786 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700787 "src/math/roundz-neon-addsub.c",
788 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800789 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800790 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
791 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
792 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
793 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
794 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
795 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700796 "src/requantization/precise-neon.c",
797 "src/requantization/fp32-neon.c",
798 "src/requantization/q31-neon.c",
799 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700800]
801
802NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -0700803 "src/f32-ibilinear/gen/neonfma-c4.c",
804 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700805 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
806 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
807 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
808 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
809 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
810 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
811 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
812 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
813 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700814 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
815 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
816 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
817 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700818 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
819 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
820 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
821 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700822 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
823 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
824 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
825 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700826 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
827 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
828 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
829 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
830 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
831 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
832 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
833 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
834 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
835 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
836 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
837 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
838 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
839 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
840 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
841 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
842 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
843 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800844 "src/f32-hswish/gen/neonfma-x4.c",
845 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700846 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
847 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800848 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
849 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
850 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
851 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
852 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
853 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
854 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
855 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
856 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
857 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
858 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
859 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
860 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
861 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
862 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
863 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
864 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
865 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
866 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
867 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
868 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
869 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
870 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
871 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800872 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
873 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
874 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
875 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
876 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
877 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
878 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
879 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
880 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
881 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
882 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
883 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
884 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
885 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
886 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
887 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
888 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
889 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
890 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
891 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
892 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
893 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
894 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
895 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
896 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
897 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
898 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
899 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
900 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
901 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
902 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
903 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
904 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
905 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
906 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
907 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
908 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
909 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
910 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
911 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
912 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
913 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
914 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
915 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
916 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
917 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
918 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
919 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
920 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
921 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
922 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
923 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
924 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
925 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700926 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
927 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -0800928 "src/math/exp-neonfma-lut64-p2.c",
929 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -0800930 "src/math/expminus-neonfma-lut2048-p1.c",
931 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -0800932 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800933 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
934 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
935 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
936 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
937 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
938 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
939 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
940 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
941 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
942 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
943 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
944 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
945 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
946 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
947 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
948 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
949 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
950 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700951]
952
953AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700954 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
955 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
956 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
957 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
958 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
959 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700960 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
961 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
962 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
963 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
964 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
965 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
966 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
967 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
968 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
969 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
970 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
971 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
972 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
973 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
974 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
975 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
976 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
977 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
978 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
979 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700980 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
981 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700982 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
983 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700984 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
985 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700986 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
987 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700988 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
989 "src/f32-dwconv-chw/3x3p1-neonfma.c",
990 "src/f32-dwconv-chw/5x5p2-neonfma.c",
991 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
992 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800993 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
994 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
995 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
996 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
997 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
998 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
999 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1000 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1001 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1002 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1003 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1004 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1005 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1006 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1007 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1008 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1009 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1010 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001011 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1012 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1013 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1014 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1015 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1016 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1017 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1018 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1019 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1020 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1021 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1022 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1023 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1024 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1025 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1026 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1027 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1028 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001029 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1030 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1031 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1032 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1033 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1034 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001035]
1036
Marat Dukhan8853b822020-05-07 12:19:01 -07001037NEONV8_UKERNELS = [
1038 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001039 "src/math/roundd-neonv8.c",
1040 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001041 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001042]
1043
Marat Dukhan08c4a432019-10-03 09:29:21 -07001044AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001045 "src/f16-clamp/gen/neonfp16arith-x16.c",
1046 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001047 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1048 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1049 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1050 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1051 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1052 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1053 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1054 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1055 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1056 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1057 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1058 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001059 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1060 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001061 "src/f16-hswish/gen/neonfp16arith-x16.c",
1062 "src/f16-hswish/gen/neonfp16arith-x8.c",
1063 "src/f16-prelu/gen/neonfp16arith-2x16.c",
1064 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001065 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1066 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1067 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1068 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1069 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1070 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1071 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1072 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1073 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1074 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1075 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1076 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1077 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1078 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1079 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1080 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1081 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1082 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1083 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1084 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1085 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1086 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1087 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1088 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1089 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1090 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1091 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1092 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001093 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1094 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001095 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1096 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1097 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1098 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1099 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1100 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1101 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1102 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001103 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1104 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1105 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1106 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001107 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1108 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1109 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1110 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1111 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1112 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1113 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1114 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1115 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1116 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1117 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1118 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001119 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1120 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1121 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1122 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1123 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1124 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1125 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1126 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001127]
1128
1129SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001130 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1131 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001132 "src/f32-clamp/gen/sse-x4.c",
1133 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001134 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1135 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001136 "src/f32-dwconv-chw/3x3p1-sse.c",
1137 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001138 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1139 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1140 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1141 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1142 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1143 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1144 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1145 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1146 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1147 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1148 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1149 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001150 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001151 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1152 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001153 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1154 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1155 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1156 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1157 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1158 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1159 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1160 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1161 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1162 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1163 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1164 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1165 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001166 "src/f32-hswish/gen/sse-x4.c",
1167 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001168 "src/f32-ibilinear/gen/sse-c4.c",
1169 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001170 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1171 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1172 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1173 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1174 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1175 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1176 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001177 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1178 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1179 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001180 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001181 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001182 "src/f32-spmm/gen/4x1-minmax-sse.c",
1183 "src/f32-spmm/gen/8x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001184 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1185 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1186 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1187 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1188 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1189 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1190 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1191 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001192 "src/f32-vbinary/gen/vmax-sse-x4.c",
1193 "src/f32-vbinary/gen/vmax-sse-x8.c",
1194 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1195 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1196 "src/f32-vbinary/gen/vmin-sse-x4.c",
1197 "src/f32-vbinary/gen/vmin-sse-x8.c",
1198 "src/f32-vbinary/gen/vminc-sse-x4.c",
1199 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001200 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1201 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1202 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1203 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1204 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1205 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001206 "src/f32-vbinary/gen/vrsqrdiffc-sse-x4.c",
1207 "src/f32-vbinary/gen/vrsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001208 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1209 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001210 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1211 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1212 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1213 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001214 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1215 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1216 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1217 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001218 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1219 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001220 "src/f32-vunary/gen/vabs-sse-x4.c",
1221 "src/f32-vunary/gen/vabs-sse-x8.c",
1222 "src/f32-vunary/gen/vneg-sse-x4.c",
1223 "src/f32-vunary/gen/vneg-sse-x8.c",
1224 "src/f32-vunary/gen/vsqr-sse-x4.c",
1225 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001226 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001227 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001228 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001229 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001230 "src/math/roundd-sse-addsub.c",
1231 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001232 "src/math/roundz-sse-addsub.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001233]
1234
1235SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001236 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1237 "src/f32-argmaxpool/4x-sse2-c4.c",
1238 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001239 "src/f32-prelu/gen/sse2-2x4.c",
1240 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001241 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1242 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1243 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1244 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1245 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1246 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1247 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1248 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1249 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1250 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1251 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1252 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001253 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001254 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001255 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001256 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001257 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1258 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001259 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1260 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001261 "src/q8-igemm/4x4c2-minmax-sse2.c",
1262 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001263 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1264 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001265 "src/q8-gemm/2x4c8-minmax-sse2.c",
1266 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001267 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001268 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001269 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001270 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001271 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001272 "src/x32-zip/x2-sse2.c",
1273 "src/x32-zip/x3-sse2.c",
1274 "src/x32-zip/x4-sse2.c",
1275 "src/x32-zip/xm-sse2.c",
1276 "src/x8-zip/x2-sse2.c",
1277 "src/x8-zip/x3-sse2.c",
1278 "src/x8-zip/x4-sse2.c",
1279 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001280 "src/math/exp-sse2-p5.c",
1281 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001282 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001283 "src/math/roundd-sse2-cvt.c",
1284 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001285 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001286 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001287 "src/requantization/precise-sse2.c",
1288 "src/requantization/fp32-sse2.c",
1289 "src/requantization/q31-sse2.c",
1290 "src/requantization/gemmlowp-sse2.c",
1291]
1292
1293SSSE3_UKERNELS = [
1294 "src/requantization/precise-ssse3.c",
1295 "src/requantization/q31-ssse3.c",
1296 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001297]
1298
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001299SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001300 "src/f32-prelu/gen/sse41-2x4.c",
1301 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001302 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001303 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001304 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001305 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001306 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1307 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001308 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001309 "src/math/roundd-sse41.c",
1310 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001311 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001312 "src/requantization/precise-sse4.c",
1313 "src/requantization/q31-sse4.c",
1314 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001315]
1316
Marat Dukhan08c4a432019-10-03 09:29:21 -07001317AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001318 "src/f32-clamp/gen/avx-x8.c",
1319 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001320 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1321 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1322 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1323 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1324 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1325 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1326 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1327 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1328 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1329 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1330 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1331 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1332 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1333 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1334 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1335 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1336 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1337 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1338 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1339 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1340 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1341 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1342 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1343 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1344 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1345 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1346 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1347 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1348 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1349 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001350 "src/f32-hswish/gen/avx-x8.c",
1351 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001352 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1353 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1354 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1355 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1356 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1357 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1358 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1359 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1360 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001361 "src/f32-prelu/gen/avx-2x8.c",
1362 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001363 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001364 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1365 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1366 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1367 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1368 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1369 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1370 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1371 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001372 "src/f32-vbinary/gen/vmax-avx-x8.c",
1373 "src/f32-vbinary/gen/vmax-avx-x16.c",
1374 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1375 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1376 "src/f32-vbinary/gen/vmin-avx-x8.c",
1377 "src/f32-vbinary/gen/vmin-avx-x16.c",
1378 "src/f32-vbinary/gen/vminc-avx-x8.c",
1379 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001380 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1381 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1382 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1383 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1384 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1385 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001386 "src/f32-vbinary/gen/vrsqrdiffc-avx-x8.c",
1387 "src/f32-vbinary/gen/vrsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001388 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1389 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001390 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1391 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1392 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1393 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001394 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1395 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1396 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1397 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001398 "src/f32-vscale/avx-unroll32.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001399 "src/f32-vunary/gen/vabs-avx-x8.c",
1400 "src/f32-vunary/gen/vabs-avx-x16.c",
1401 "src/f32-vunary/gen/vneg-avx-x8.c",
1402 "src/f32-vunary/gen/vneg-avx-x16.c",
1403 "src/f32-vunary/gen/vsqr-avx-x8.c",
1404 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001405]
1406
Marat Dukhanfda12b82019-11-21 12:27:59 -08001407FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001408 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1409 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1410 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1411 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1412 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1413 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1414 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1415 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1416 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1417 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1418 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1419 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1420 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1421 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1422 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1423 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1424 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1425 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1426 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1427 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1428 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1429 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1430 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1431 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1432 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1433 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1434 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1435 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1436 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1437 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1438 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1439 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1440 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1441 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1442 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1443 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1444 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1445 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1446 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1447 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001448 "src/f32-hswish/gen/fma3-x8.c",
1449 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001450 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1451 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1452 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1453 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1454 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1455 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1456 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1457 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1458 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1459 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1460 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1461 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1462 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1463 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001464]
1465
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001466AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001467 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1468 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1469 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1470 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1471 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1472 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1473 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1474 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1475 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1476 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1477 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1478 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1479 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1480 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1481 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1482 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1483 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1484 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1485 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1486 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1487 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1488 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1489 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1490 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1491 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1492 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1493 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1494 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1495 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1496 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1497 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1498 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1499 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1500 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1501 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1502 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001503 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1504 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1505 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1506 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1507 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1508 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1509 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1510 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1511 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1512 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1513 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1514 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1515 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1516 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1517 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1518 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1519 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1520 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1521 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1522 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1523 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1524 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1525 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1526 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1527 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1528 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1529 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1530 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1531 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1532 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001533 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1534 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1535 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1536 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1537 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1538 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1539 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1540 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1541 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1542 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1543 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1544 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1545 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1546 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1547 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1548 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1549 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1550 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1551 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1552 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1553 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1554 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1555 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1556 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001557 "src/math/exp-avx2-p5.c",
1558 "src/math/exp-avx2-perm-p3.c",
1559 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001560 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001561 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001562 "src/math/sigmoid-avx2-rr2-p5-div.c",
1563 "src/math/sigmoid-avx2-rr1-p5-div.c",
1564 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1565 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1566 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1567 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001568]
1569
Marat Dukhan08c4a432019-10-03 09:29:21 -07001570AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001571 "src/f32-clamp/gen/avx512f-x16.c",
1572 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001573 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1574 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1575 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1576 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1577 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1578 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1579 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1580 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1581 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1582 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1583 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1584 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1585 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1586 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1587 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1588 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1589 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1590 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1591 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1592 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1593 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1594 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1595 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1596 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001597 "src/f32-hswish/gen/avx512f-x16.c",
1598 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001599 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1600 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1601 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1602 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1603 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1604 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001605 "src/f32-prelu/gen/avx512f-2x16.c",
1606 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001607 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1608 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1609 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1610 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1611 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1612 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1613 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1614 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1615 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1616 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1617 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1618 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1619 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1620 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1621 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1622 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1623 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1624 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1625 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1626 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1627 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1628 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1629 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1630 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1631 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1632 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1633 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1634 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1635 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1636 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1637 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1638 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1639 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1640 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1641 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1642 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001643 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001644 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
1645 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
1646 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
1647 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
1648 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
1649 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
1650 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
1651 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001652 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1653 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1654 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1655 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1656 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1657 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1658 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1659 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001660 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
1661 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
1662 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
1663 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
1664 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
1665 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001666 "src/f32-vbinary/gen/vrsqrdiffc-avx512f-x16.c",
1667 "src/f32-vbinary/gen/vrsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001668 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
1669 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001670 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
1671 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
1672 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
1673 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001674 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
1675 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
1676 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
1677 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001678 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001679 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1680 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1681 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1682 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1683 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1684 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1685 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1686 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1687 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1688 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1689 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1690 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1691 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1692 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1693 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1694 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1695 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1696 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1697 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1698 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1699 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1700 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1701 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1702 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001703 "src/f32-vunary/gen/vabs-avx512f-x16.c",
1704 "src/f32-vunary/gen/vabs-avx512f-x32.c",
1705 "src/f32-vunary/gen/vneg-avx512f-x16.c",
1706 "src/f32-vunary/gen/vneg-avx512f-x32.c",
1707 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
1708 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001709 "src/math/exp-avx512f-p5-scalef.c",
1710 "src/math/exp-avx512f-p5.c",
1711 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001712 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001713 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001714]
1715
1716AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07001717 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07001718 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001719 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1720 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1721 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1722 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1723 "src/f32-gemm/4x8-minmax-aarch32-neon-ld64.S",
1724 "src/f32-igemm/4x8-minmax-aarch32-neon-ld64.S",
1725 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1726 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1727 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1728 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001729]
1730
1731AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07001732 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001733 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
1734 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07001735 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001736 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1737 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001738 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
1739 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
1740 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001741 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001742 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1743 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1744 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001745 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001746 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1747 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001748 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001749 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1750 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1751 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1752 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1753 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1754 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1755 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1756 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1757 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1758 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
1759 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
1760 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1761 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1762 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1763 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1764 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1765 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1766 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1767 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
1768 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
1769 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001770 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001771 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1772 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1773 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1774 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1775 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
1776 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1777 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1778 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1779 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1780 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
1781 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
1782 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1783 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1784 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
1785 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
1786 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
1787 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
1788 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
1789 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ios.S",
1790 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
1791 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
1792 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1793 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1794 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1795 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1796 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1797 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1798 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1799 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
1800 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
1801 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
1802 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
1803 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
1804 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
1805 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
1806 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
1807 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
1808 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-ios.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001809]
1810
Marat Dukhan1b354632020-03-23 12:50:22 -07001811INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001812 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001813 "src/xnnpack/argmaxpool.h",
1814 "src/xnnpack/avgpool.h",
1815 "src/xnnpack/clamp.h",
1816 "src/xnnpack/common.h",
1817 "src/xnnpack/conv.h",
1818 "src/xnnpack/dwconv.h",
1819 "src/xnnpack/gavgpool.h",
1820 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001821 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001822 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07001823 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001824 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08001825 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001826 "src/xnnpack/lut.h",
1827 "src/xnnpack/math.h",
1828 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001829 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001830 "src/xnnpack/packx.h",
1831 "src/xnnpack/pad.h",
1832 "src/xnnpack/params.h",
1833 "src/xnnpack/pavgpool.h",
1834 "src/xnnpack/ppmm.h",
1835 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001836 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001837 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001838 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001839 "src/xnnpack/rmax.h",
1840 "src/xnnpack/scalar-utils.h",
1841 "src/xnnpack/spmm.h",
1842 "src/xnnpack/unpool.h",
1843 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001844 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001845 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001846 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07001847 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07001848 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08001849 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001850 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07001851]
1852
1853INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001854 "include/xnnpack.h",
1855 "src/xnnpack/allocator.h",
1856 "src/xnnpack/compute.h",
1857 "src/xnnpack/im2col.h",
1858 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001859 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07001860 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001861 "src/xnnpack/operator.h",
1862 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001863 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001864 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001865 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08001866 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001867]
1868
Marat Dukhan1b354632020-03-23 12:50:22 -07001869ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001870 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001871]
1872
Marat Dukhan1b354632020-03-23 12:50:22 -07001873MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001874 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001875 "include/xnnpack.h",
1876]
1877
Marat Dukhan1b354632020-03-23 12:50:22 -07001878MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07001879 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07001880 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001881 "src/xnnpack/requantization.h",
1882 "include/xnnpack.h",
1883]
1884
1885OPERATOR_TEST_PARAMS_HDRS = [
1886 "src/xnnpack/params.h",
1887 "src/xnnpack/common.h",
1888]
1889
1890WEIGHTS_PACK_HDRS = [
1891 "src/xnnpack/pack.h",
1892 "src/xnnpack/operator.h",
1893 "src/xnnpack/compute.h",
1894]
1895
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001896LOGGING_COPTS = select({
1897 # No logging in optimized mode
1898 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
1899 # Full logging in debug mode
1900 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
1901 # Error-only logging in default (fastbuild) mode
1902 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
1903})
1904
Marat Dukhan3b59de22020-06-03 20:15:19 -07001905LOGGING_SRCS = select({
1906 # No logging in optimized mode
1907 ":optimized_build": [],
1908 "//conditions:default": [
1909 "src/operator-strings.c",
1910 "src/subgraph-strings.c",
1911 ],
1912})
1913
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07001914LOGGING_HDRS = [
1915 "src/xnnpack/log.h",
1916]
1917
Marat Dukhan08c4a432019-10-03 09:29:21 -07001918xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001919 name = "tables",
1920 srcs = TABLE_SRCS,
1921 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001922 gcc_copts = xnnpack_gcc_std_copts(),
1923 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001924)
1925
1926xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07001927 name = "scalar_ukernels",
1928 srcs = SCALAR_UKERNELS,
1929 hdrs = INTERNAL_HDRS,
1930 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07001931 gcc_copts = xnnpack_gcc_std_copts(),
1932 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07001933 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001934 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001935 "@FP16",
1936 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08001937 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001938 ],
1939)
1940
1941xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07001942 name = "scalar_ukernels_test_mode",
1943 srcs = SCALAR_UKERNELS,
1944 hdrs = INTERNAL_HDRS,
1945 aarch32_copts = ["-marm"],
1946 copts = [
1947 "-UNDEBUG",
1948 "-DXNN_TEST_MODE=1",
1949 ],
1950 gcc_copts = xnnpack_gcc_std_copts(),
1951 msvc_copts = xnnpack_msvc_std_copts(),
1952 deps = [
1953 ":tables",
1954 "@FP16",
1955 "@FXdiv",
1956 "@pthreadpool",
1957 ],
1958)
1959
1960xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08001961 name = "wasm_ukernels",
1962 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07001963 gcc_copts = xnnpack_gcc_std_copts(),
1964 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08001965 wasm_srcs = WASM_UKERNELS,
1966 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08001967 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08001968 "@FP16",
1969 "@FXdiv",
1970 "@pthreadpool",
1971 ],
1972)
1973
1974xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07001975 name = "wasm_ukernels_test_mode",
1976 hdrs = INTERNAL_HDRS,
1977 copts = [
1978 "-UNDEBUG",
1979 "-DXNN_TEST_MODE=1",
1980 ],
1981 gcc_copts = xnnpack_gcc_std_copts(),
1982 msvc_copts = xnnpack_msvc_std_copts(),
1983 wasm_srcs = WASM_UKERNELS,
1984 deps = [
1985 ":tables",
1986 "@FP16",
1987 "@FXdiv",
1988 "@pthreadpool",
1989 ],
1990)
1991
1992xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001993 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001994 hdrs = INTERNAL_HDRS,
1995 aarch32_copts = [
1996 "-marm",
1997 "-mfpu=neon",
1998 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07001999 gcc_copts = xnnpack_gcc_std_copts(),
2000 gcc_x86_copts = ["-msse2"],
2001 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002002 optimized_copts = [
2003 "-O3",
2004 "-ffast-math",
2005 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002006 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002007 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002008 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002009 "@FP16",
2010 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002011 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002012 ],
2013)
2014
2015xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002016 name = "psimd_fastmath_ukernels_test_mode",
2017 hdrs = INTERNAL_HDRS,
2018 aarch32_copts = [
2019 "-marm",
2020 "-mfpu=neon",
2021 ],
2022 copts = [
2023 "-UNDEBUG",
2024 "-DXNN_TEST_MODE=1",
2025 ],
2026 gcc_copts = xnnpack_gcc_std_copts(),
2027 gcc_x86_copts = ["-msse2"],
2028 msvc_copts = xnnpack_msvc_std_copts(),
2029 optimized_copts = [
2030 "-O3",
2031 "-ffast-math",
2032 ],
2033 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2034 deps = [
2035 ":tables",
2036 "@FP16",
2037 "@psimd",
2038 "@pthreadpool",
2039 ],
2040)
2041
2042xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002043 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002044 hdrs = INTERNAL_HDRS,
2045 aarch32_copts = [
2046 "-marm",
2047 "-mfpu=neon",
2048 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002049 gcc_copts = xnnpack_gcc_std_copts(),
2050 gcc_x86_copts = ["-msse2"],
2051 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002052 optimized_copts = [
2053 "-O3",
2054 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002055 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002056 deps = [
2057 ":tables",
2058 "@FP16",
2059 "@psimd",
2060 "@pthreadpool",
2061 ],
2062)
2063
2064xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002065 name = "psimd_accmath_ukernels_test_mode",
2066 hdrs = INTERNAL_HDRS,
2067 aarch32_copts = [
2068 "-marm",
2069 "-mfpu=neon",
2070 ],
2071 copts = [
2072 "-UNDEBUG",
2073 "-DXNN_TEST_MODE=1",
2074 ],
2075 gcc_copts = xnnpack_gcc_std_copts(),
2076 gcc_x86_copts = ["-msse2"],
2077 msvc_copts = xnnpack_msvc_std_copts(),
2078 optimized_copts = [
2079 "-O3",
2080 ],
2081 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2082 deps = [
2083 ":tables",
2084 "@FP16",
2085 "@psimd",
2086 "@pthreadpool",
2087 ],
2088)
2089
2090xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002091 name = "neon_ukernels",
2092 hdrs = INTERNAL_HDRS,
2093 aarch32_copts = [
2094 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002095 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002096 "-mfpu=neon",
2097 ],
2098 aarch32_srcs = NEON_UKERNELS,
2099 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002100 gcc_copts = xnnpack_gcc_std_copts(),
2101 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002102 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002103 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002104 "@FP16",
2105 "@pthreadpool",
2106 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002107)
2108
2109xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002110 name = "neon_ukernels_test_mode",
2111 hdrs = INTERNAL_HDRS,
2112 aarch32_copts = [
2113 "-marm",
2114 "-march=armv7-a",
2115 "-mfpu=neon",
2116 ],
2117 aarch32_srcs = NEON_UKERNELS,
2118 aarch64_srcs = NEON_UKERNELS,
2119 copts = [
2120 "-UNDEBUG",
2121 "-DXNN_TEST_MODE=1",
2122 ],
2123 gcc_copts = xnnpack_gcc_std_copts(),
2124 msvc_copts = xnnpack_msvc_std_copts(),
2125 deps = [
2126 ":tables",
2127 "@FP16",
2128 "@pthreadpool",
2129 ],
2130)
2131
2132xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002133 name = "neonfma_ukernels",
2134 hdrs = INTERNAL_HDRS,
2135 aarch32_copts = [
2136 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002137 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002138 "-mfpu=neon-vfpv4",
2139 ],
2140 aarch32_srcs = NEONFMA_UKERNELS,
2141 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002142 gcc_copts = xnnpack_gcc_std_copts(),
2143 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002144 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002145 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002146 "@FP16",
2147 "@pthreadpool",
2148 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002149)
2150
2151xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002152 name = "neonfma_ukernels_test_mode",
2153 hdrs = INTERNAL_HDRS,
2154 aarch32_copts = [
2155 "-marm",
2156 "-march=armv7-a",
2157 "-mfpu=neon-vfpv4",
2158 ],
2159 aarch32_srcs = NEONFMA_UKERNELS,
2160 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
2161 copts = [
2162 "-UNDEBUG",
2163 "-DXNN_TEST_MODE=1",
2164 ],
2165 gcc_copts = xnnpack_gcc_std_copts(),
2166 msvc_copts = xnnpack_msvc_std_copts(),
2167 deps = [
2168 ":tables",
2169 "@FP16",
2170 "@pthreadpool",
2171 ],
2172)
2173
2174xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002175 name = "neonv8_ukernels",
2176 hdrs = INTERNAL_HDRS,
2177 aarch32_copts = [
2178 "-marm",
2179 "-march=armv8-a",
2180 "-mfpu=neon-fp-armv8",
2181 ],
2182 aarch32_srcs = NEONV8_UKERNELS,
2183 aarch64_srcs = NEONV8_UKERNELS,
2184 gcc_copts = xnnpack_gcc_std_copts(),
2185 msvc_copts = xnnpack_msvc_std_copts(),
2186 deps = [
2187 ":tables",
2188 "@FP16",
2189 "@pthreadpool",
2190 ],
2191)
2192
2193xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002194 name = "neonv8_ukernels_test_mode",
2195 hdrs = INTERNAL_HDRS,
2196 aarch32_copts = [
2197 "-marm",
2198 "-march=armv8-a",
2199 "-mfpu=neon-fp-armv8",
2200 ],
2201 aarch32_srcs = NEONV8_UKERNELS,
2202 aarch64_srcs = NEONV8_UKERNELS,
2203 copts = [
2204 "-UNDEBUG",
2205 "-DXNN_TEST_MODE=1",
2206 ],
2207 gcc_copts = xnnpack_gcc_std_copts(),
2208 msvc_copts = xnnpack_msvc_std_copts(),
2209 deps = [
2210 ":tables",
2211 "@FP16",
2212 "@pthreadpool",
2213 ],
2214)
2215
2216xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002217 name = "neonfp16arith_ukernels",
2218 hdrs = INTERNAL_HDRS,
2219 aarch64_copts = ["-march=armv8.2-a+fp16"],
2220 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002221 gcc_copts = xnnpack_gcc_std_copts(),
2222 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002223 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002224 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002225 "@FP16",
2226 "@pthreadpool",
2227 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002228)
2229
2230xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002231 name = "neonfp16arith_ukernels_test_mode",
2232 hdrs = INTERNAL_HDRS,
2233 aarch64_copts = ["-march=armv8.2-a+fp16"],
2234 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2235 copts = [
2236 "-UNDEBUG",
2237 "-DXNN_TEST_MODE=1",
2238 ],
2239 gcc_copts = xnnpack_gcc_std_copts(),
2240 msvc_copts = xnnpack_msvc_std_copts(),
2241 deps = [
2242 ":tables",
2243 "@FP16",
2244 "@pthreadpool",
2245 ],
2246)
2247
2248xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002249 name = "sse2_ukernels",
2250 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002251 gcc_copts = xnnpack_gcc_std_copts(),
2252 gcc_x86_copts = ["-msse2"],
2253 msvc_copts = xnnpack_msvc_std_copts(),
2254 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002255 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002256 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002257 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002258 "@FP16",
2259 "@pthreadpool",
2260 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002261)
2262
2263xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002264 name = "sse2_ukernels_test_mode",
2265 hdrs = INTERNAL_HDRS,
2266 copts = [
2267 "-UNDEBUG",
2268 "-DXNN_TEST_MODE=1",
2269 ],
2270 gcc_copts = xnnpack_gcc_std_copts(),
2271 gcc_x86_copts = ["-msse2"],
2272 msvc_copts = xnnpack_msvc_std_copts(),
2273 msvc_x86_32_copts = ["/arch:SSE2"],
2274 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2275 deps = [
2276 ":tables",
2277 "@FP16",
2278 "@pthreadpool",
2279 ],
2280)
2281
2282xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002283 name = "ssse3_ukernels",
2284 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002285 gcc_copts = xnnpack_gcc_std_copts(),
2286 gcc_x86_copts = ["-mssse3"],
2287 msvc_copts = xnnpack_msvc_std_copts(),
2288 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002289 x86_srcs = SSSE3_UKERNELS,
2290 deps = [
2291 ":tables",
2292 "@FP16",
2293 "@pthreadpool",
2294 ],
2295)
2296
2297xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002298 name = "ssse3_ukernels_test_mode",
2299 hdrs = INTERNAL_HDRS,
2300 copts = [
2301 "-UNDEBUG",
2302 "-DXNN_TEST_MODE=1",
2303 ],
2304 gcc_copts = xnnpack_gcc_std_copts(),
2305 gcc_x86_copts = ["-mssse3"],
2306 msvc_copts = xnnpack_msvc_std_copts(),
2307 msvc_x86_32_copts = ["/arch:SSE2"],
2308 x86_srcs = SSSE3_UKERNELS,
2309 deps = [
2310 ":tables",
2311 "@FP16",
2312 "@pthreadpool",
2313 ],
2314)
2315
2316xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002317 name = "sse41_ukernels",
2318 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002319 gcc_copts = xnnpack_gcc_std_copts(),
2320 gcc_x86_copts = ["-msse4.1"],
2321 msvc_copts = xnnpack_msvc_std_copts(),
2322 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002323 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002324 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002325 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002326 "@FP16",
2327 "@pthreadpool",
2328 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002329)
2330
2331xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002332 name = "sse41_ukernels_test_mode",
2333 hdrs = INTERNAL_HDRS,
2334 copts = [
2335 "-UNDEBUG",
2336 "-DXNN_TEST_MODE=1",
2337 ],
2338 gcc_copts = xnnpack_gcc_std_copts(),
2339 gcc_x86_copts = ["-msse4.1"],
2340 msvc_copts = xnnpack_msvc_std_copts(),
2341 msvc_x86_32_copts = ["/arch:SSE2"],
2342 x86_srcs = SSE41_UKERNELS,
2343 deps = [
2344 ":tables",
2345 "@FP16",
2346 "@pthreadpool",
2347 ],
2348)
2349
2350xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002351 name = "avx_ukernels",
2352 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002353 gcc_copts = xnnpack_gcc_std_copts(),
2354 gcc_x86_copts = ["-mavx"],
2355 msvc_copts = xnnpack_msvc_std_copts(),
2356 msvc_x86_32_copts = ["/arch:AVX"],
2357 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002358 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002359 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002360 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002361 "@FP16",
2362 "@pthreadpool",
2363 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002364)
2365
2366xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002367 name = "avx_ukernels_test_mode",
2368 hdrs = INTERNAL_HDRS,
2369 copts = [
2370 "-UNDEBUG",
2371 "-DXNN_TEST_MODE=1",
2372 ],
2373 gcc_copts = xnnpack_gcc_std_copts(),
2374 gcc_x86_copts = ["-mavx"],
2375 msvc_copts = xnnpack_msvc_std_copts(),
2376 msvc_x86_32_copts = ["/arch:AVX"],
2377 msvc_x86_64_copts = ["/arch:AVX"],
2378 x86_srcs = AVX_UKERNELS,
2379 deps = [
2380 ":tables",
2381 "@FP16",
2382 "@pthreadpool",
2383 ],
2384)
2385
2386xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002387 name = "fma3_ukernels",
2388 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002389 gcc_copts = xnnpack_gcc_std_copts(),
2390 gcc_x86_copts = ["-mfma"],
2391 msvc_copts = xnnpack_msvc_std_copts(),
2392 msvc_x86_32_copts = ["/arch:AVX"],
2393 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002394 x86_srcs = FMA3_UKERNELS,
2395 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002396 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002397 "@FP16",
2398 "@pthreadpool",
2399 ],
2400)
2401
2402xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002403 name = "fma3_ukernels_test_mode",
2404 hdrs = INTERNAL_HDRS,
2405 copts = [
2406 "-UNDEBUG",
2407 "-DXNN_TEST_MODE=1",
2408 ],
2409 gcc_copts = xnnpack_gcc_std_copts(),
2410 gcc_x86_copts = ["-mfma"],
2411 msvc_copts = xnnpack_msvc_std_copts(),
2412 msvc_x86_32_copts = ["/arch:AVX"],
2413 msvc_x86_64_copts = ["/arch:AVX"],
2414 x86_srcs = FMA3_UKERNELS,
2415 deps = [
2416 ":tables",
2417 "@FP16",
2418 "@pthreadpool",
2419 ],
2420)
2421
2422xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002423 name = "avx2_ukernels",
2424 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002425 gcc_copts = xnnpack_gcc_std_copts(),
2426 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002427 "-mfma",
2428 "-mavx2",
2429 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002430 msvc_copts = xnnpack_msvc_std_copts(),
2431 msvc_x86_32_copts = ["/arch:AVX2"],
2432 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002433 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002434 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002435 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002436 "@FP16",
2437 "@pthreadpool",
2438 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002439)
2440
2441xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002442 name = "avx2_ukernels_test_mode",
2443 hdrs = INTERNAL_HDRS,
2444 copts = [
2445 "-UNDEBUG",
2446 "-DXNN_TEST_MODE=1",
2447 ],
2448 gcc_copts = xnnpack_gcc_std_copts(),
2449 gcc_x86_copts = [
2450 "-mfma",
2451 "-mavx2",
2452 ],
2453 msvc_copts = xnnpack_msvc_std_copts(),
2454 msvc_x86_32_copts = ["/arch:AVX2"],
2455 msvc_x86_64_copts = ["/arch:AVX2"],
2456 x86_srcs = AVX2_UKERNELS,
2457 deps = [
2458 ":tables",
2459 "@FP16",
2460 "@pthreadpool",
2461 ],
2462)
2463
2464xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002465 name = "avx512f_ukernels",
2466 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002467 gcc_copts = xnnpack_gcc_std_copts(),
2468 gcc_x86_copts = ["-mavx512f"],
2469 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2470 msvc_copts = xnnpack_msvc_std_copts(),
2471 msvc_x86_32_copts = ["/arch:AVX512"],
2472 msvc_x86_64_copts = ["/arch:AVX512"],
2473 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002474 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002475 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002476 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002477 "@FP16",
2478 "@pthreadpool",
2479 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002480)
2481
2482xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002483 name = "avx512f_ukernels_test_mode",
2484 hdrs = INTERNAL_HDRS,
2485 copts = [
2486 "-UNDEBUG",
2487 "-DXNN_TEST_MODE=1",
2488 ],
2489 gcc_copts = xnnpack_gcc_std_copts(),
2490 gcc_x86_copts = ["-mavx512f"],
2491 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2492 msvc_copts = xnnpack_msvc_std_copts(),
2493 msvc_x86_32_copts = ["/arch:AVX512"],
2494 msvc_x86_64_copts = ["/arch:AVX512"],
2495 msys_copts = ["-fno-asynchronous-unwind-tables"],
2496 x86_srcs = AVX512F_UKERNELS,
2497 deps = [
2498 ":tables",
2499 "@FP16",
2500 "@pthreadpool",
2501 ],
2502)
2503
2504xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002505 name = "asm_ukernels",
2506 hdrs = ["src/xnnpack/assembly.h"],
2507 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002508 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002509 aarch64_srcs = AARCH64_ASM_UKERNELS,
2510)
2511
Marat Dukhan3b59de22020-06-03 20:15:19 -07002512xnnpack_cc_library(
2513 name = "logging_utils",
2514 srcs = LOGGING_SRCS,
2515 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2516 copts = LOGGING_COPTS + [
2517 "-Isrc",
2518 "-Iinclude",
2519 ] + select({
2520 ":debug_build": [],
2521 "//conditions:default": xnnpack_min_size_copts(),
2522 }),
2523 gcc_copts = xnnpack_gcc_std_copts(),
2524 msvc_copts = xnnpack_msvc_std_copts(),
2525 visibility = xnnpack_visibility(),
2526 deps = [
2527 "@FP16",
2528 "@clog",
2529 "@pthreadpool",
2530 ],
2531)
2532
Marat Dukhan08c4a432019-10-03 09:29:21 -07002533xnnpack_aggregate_library(
2534 name = "ukernels",
2535 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002536 ":neon_ukernels",
2537 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002538 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002539 ":asm_ukernels",
2540 ],
2541 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002542 ":neon_ukernels",
2543 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002544 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002545 ":neonfp16arith_ukernels",
2546 ":asm_ukernels",
2547 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002548 generic_deps = [
2549 ":scalar_ukernels",
2550 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002551 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002552 ":psimd_fastmath_ukernels",
2553 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002554 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002555 wasm_deps = [
2556 ":wasm_ukernels",
2557 ],
2558 wasmsimd_deps = [
2559 ":wasm_ukernels",
2560 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002561 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002562 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002563 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002564 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002565 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002566 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002567 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002568 ":avx512f_ukernels",
2569 ],
2570)
2571
Marat Dukhan33fcf782020-05-24 14:27:15 -07002572xnnpack_aggregate_library(
2573 name = "ukernels_test_mode",
2574 aarch32_deps = [
2575 ":neon_ukernels_test_mode",
2576 ":neonfma_ukernels_test_mode",
2577 ":neonv8_ukernels_test_mode",
2578 ":asm_ukernels",
2579 ],
2580 aarch64_deps = [
2581 ":neon_ukernels_test_mode",
2582 ":neonfma_ukernels_test_mode",
2583 ":neonv8_ukernels_test_mode",
2584 ":neonfp16arith_ukernels_test_mode",
2585 ":asm_ukernels",
2586 ],
2587 generic_deps = [
2588 ":scalar_ukernels_test_mode",
2589 ],
2590 psimd_deps = [
2591 ":psimd_fastmath_ukernels_test_mode",
2592 ":psimd_accmath_ukernels_test_mode",
2593 ],
2594 wasm_deps = [
2595 ":wasm_ukernels_test_mode",
2596 ],
2597 wasmsimd_deps = [
2598 ":wasm_ukernels_test_mode",
2599 ],
2600 x86_deps = [
2601 ":sse2_ukernels_test_mode",
2602 ":ssse3_ukernels_test_mode",
2603 ":sse41_ukernels_test_mode",
2604 ":avx_ukernels_test_mode",
2605 ":fma3_ukernels_test_mode",
2606 ":avx2_ukernels_test_mode",
2607 ":avx512f_ukernels_test_mode",
2608 ],
2609)
2610
Marat Dukhan08c4a432019-10-03 09:29:21 -07002611xnnpack_cc_library(
2612 name = "im2col",
2613 srcs = ["src/im2col.c"],
2614 hdrs = [
2615 "src/xnnpack/common.h",
2616 "src/xnnpack/im2col.h",
2617 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002618 gcc_copts = xnnpack_gcc_std_copts(),
2619 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002620)
2621
2622xnnpack_cc_library(
2623 name = "indirection",
2624 srcs = ["src/indirection.c"],
2625 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002626 gcc_copts = xnnpack_gcc_std_copts(),
2627 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002628 deps = [
2629 "@FP16",
2630 "@FXdiv",
2631 "@pthreadpool",
2632 ],
2633)
2634
2635xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002636 name = "indirection_test_mode",
2637 srcs = ["src/indirection.c"],
2638 hdrs = INTERNAL_HDRS,
2639 copts = [
2640 "-UNDEBUG",
2641 "-DXNN_TEST_MODE=1",
2642 ],
2643 gcc_copts = xnnpack_gcc_std_copts(),
2644 msvc_copts = xnnpack_msvc_std_copts(),
2645 deps = [
2646 "@FP16",
2647 "@FXdiv",
2648 "@pthreadpool",
2649 ],
2650)
2651
2652xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002653 name = "operator_run",
2654 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002655 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002656 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07002657 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2658 "//conditions:default": [],
2659 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002660 gcc_copts = xnnpack_gcc_std_copts(),
2661 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002662 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07002663 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002664 "@FP16",
2665 "@FXdiv",
2666 "@clog",
2667 "@pthreadpool",
2668 ],
2669)
2670
Chao Mei6ddfc602020-05-13 22:29:36 -07002671xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002672 name = "operator_run_test_mode",
2673 srcs = ["src/operator-run.c"],
2674 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2675 copts = LOGGING_COPTS + [
2676 "-UNDEBUG",
2677 "-DXNN_TEST_MODE=1",
2678 ] + select({
2679 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2680 "//conditions:default": [],
2681 }),
2682 gcc_copts = xnnpack_gcc_std_copts(),
2683 msvc_copts = xnnpack_msvc_std_copts(),
2684 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07002685 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07002686 "@FP16",
2687 "@FXdiv",
2688 "@clog",
2689 "@pthreadpool",
2690 ],
2691)
2692
2693xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07002694 name = "memory_planner",
2695 srcs = ["src/memory-planner.c"],
2696 hdrs = INTERNAL_HDRS,
2697 defines = select({
2698 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
2699 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
2700 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
2701 }),
2702 gcc_copts = xnnpack_gcc_std_copts(),
2703 msvc_copts = xnnpack_msvc_std_copts(),
2704 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07002705 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07002706 "@pthreadpool",
2707 ],
2708)
2709
Marat Dukhan33fcf782020-05-24 14:27:15 -07002710xnnpack_cc_library(
2711 name = "memory_planner_test_mode",
2712 srcs = ["src/memory-planner.c"],
2713 hdrs = INTERNAL_HDRS,
2714 copts = [
2715 "-UNDEBUG",
2716 "-DXNN_TEST_MODE=1",
2717 ],
2718 defines = select({
2719 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
2720 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
2721 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
2722 }),
2723 gcc_copts = xnnpack_gcc_std_copts(),
2724 msvc_copts = xnnpack_msvc_std_copts(),
2725 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07002726 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07002727 "@pthreadpool",
2728 ],
2729)
2730
Marat Dukhan08c4a432019-10-03 09:29:21 -07002731cc_library(
2732 name = "enable_assembly",
2733 defines = select({
2734 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
2735 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07002736 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002737 }),
2738)
2739
Marat Dukhancf056b22019-10-07 10:26:29 -07002740xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002741 name = "operators",
2742 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08002743 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002744 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07002745 ],
2746 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002747 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002748 "-Isrc",
2749 "-Iinclude",
2750 ] + select({
2751 ":debug_build": [],
2752 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002753 }) + select({
2754 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2755 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002756 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002757 gcc_copts = xnnpack_gcc_std_copts(),
2758 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002759 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002760 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002761 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002762 "@FP16",
2763 "@FXdiv",
2764 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002765 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002766 ],
2767)
2768
Marat Dukhan10a38082020-04-17 03:58:35 -07002769xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002770 name = "operators_test_mode",
2771 srcs = OPERATOR_SRCS + [
2772 "src/memory.c",
2773 "src/operator-delete.c",
2774 ],
2775 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2776 copts = LOGGING_COPTS + [
2777 "-Isrc",
2778 "-Iinclude",
2779 "-UNDEBUG",
2780 "-DXNN_TEST_MODE=1",
2781 ] + select({
2782 ":debug_build": [],
2783 "//conditions:default": xnnpack_min_size_copts(),
2784 }) + select({
2785 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2786 "//conditions:default": [],
2787 }),
2788 gcc_copts = xnnpack_gcc_std_copts(),
2789 msvc_copts = xnnpack_msvc_std_copts(),
2790 deps = [
2791 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002792 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07002793 "@FP16",
2794 "@FXdiv",
2795 "@clog",
2796 "@pthreadpool",
2797 ],
2798)
2799
2800xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002801 name = "XNNPACK",
2802 srcs = [
2803 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08002804 "src/runtime.c",
2805 "src/subgraph.c",
2806 "src/tensor.c",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002807 ] + SUBGRAPH_SRCS + LOGGING_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002808 hdrs = ["include/xnnpack.h"],
2809 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002810 "-Isrc",
2811 "-Iinclude",
2812 ] + select({
2813 ":debug_build": [],
2814 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002815 }) + select({
2816 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2817 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002818 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002819 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002820 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002821 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002822 visibility = xnnpack_visibility(),
2823 deps = [
2824 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002825 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07002826 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002827 ":operator_run",
2828 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002829 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002830 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07002831 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002832 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07002833 ] + select({
2834 ":emscripten": [],
2835 "//conditions:default": ["@cpuinfo"],
2836 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002837)
2838
Marat Dukhan10a38082020-04-17 03:58:35 -07002839xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002840 name = "XNNPACK_test_mode",
2841 srcs = [
2842 "src/init.c",
2843 "src/runtime.c",
2844 "src/subgraph.c",
2845 "src/tensor.c",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002846 ] + SUBGRAPH_SRCS + LOGGING_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002847 hdrs = ["include/xnnpack.h"],
2848 copts = LOGGING_COPTS + [
2849 "-Isrc",
2850 "-Iinclude",
2851 "-UNDEBUG",
2852 "-DXNN_TEST_MODE=1",
2853 ] + select({
2854 ":debug_build": [],
2855 "//conditions:default": xnnpack_min_size_copts(),
2856 }) + select({
2857 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2858 "//conditions:default": [],
2859 }),
2860 gcc_copts = xnnpack_gcc_std_copts(),
2861 includes = ["include"],
2862 msvc_copts = xnnpack_msvc_std_copts(),
2863 visibility = xnnpack_visibility(),
2864 deps = [
2865 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002866 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07002867 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07002868 ":operator_run_test_mode",
2869 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002870 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07002871 "@clog",
2872 "@FP16",
2873 "@pthreadpool",
2874 ] + select({
2875 ":emscripten": [],
2876 "//conditions:default": ["@cpuinfo"],
2877 }),
2878)
2879
2880xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002881 name = "xnnpack_operators_nhwc_f32",
2882 srcs = [
2883 "src/init.c",
2884 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002885 hdrs = ["include/xnnpack.h"],
2886 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002887 "-Isrc",
2888 "-Iinclude",
2889 ] + select({
2890 ":debug_build": [],
2891 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07002892 }) + select({
2893 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2894 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002895 }),
2896 defines = [
2897 "XNN_NO_Q8_OPERATORS",
2898 "XNN_NO_U8_OPERATORS",
2899 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08002900 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002901 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002902 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002903 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002904 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002905 visibility = xnnpack_visibility(),
2906 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002907 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002908 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002909 ":operator_run",
2910 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07002911 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002912 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002913 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07002914 ] + select({
2915 ":emscripten": [],
2916 "//conditions:default": ["@cpuinfo"],
2917 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002918)
2919
Marat Dukhancf056b22019-10-07 10:26:29 -07002920xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002921 name = "bench_utils",
2922 srcs = ["bench/utils.cc"],
2923 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08002924 deps = [
2925 "@com_google_benchmark//:benchmark",
2926 "@cpuinfo",
2927 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002928)
2929
Frank Barchard7e955972019-10-11 10:34:25 -07002930######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07002931
2932xnnpack_benchmark(
2933 name = "q8_gemm_bench",
2934 srcs = [
2935 "bench/gemm.h",
2936 "bench/q8-gemm.cc",
2937 "src/xnnpack/AlignedAllocator.h",
2938 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002939 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002940 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
2941)
2942
2943xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07002944 name = "f16_igemm_bench",
2945 srcs = [
2946 "bench/f16-igemm.cc",
2947 "bench/conv.h",
2948 "bench/google/conv.h",
2949 "src/xnnpack/AlignedAllocator.h",
2950 ] + MICROKERNEL_BENCHMARK_HDRS,
2951 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
2952)
2953
2954xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002955 name = "f16_gemm_bench",
2956 srcs = [
2957 "bench/f16-gemm.cc",
2958 "bench/gemm.h",
2959 "src/xnnpack/AlignedAllocator.h",
2960 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002961 deps = MICROKERNEL_BENCHMARK_DEPS,
2962)
2963
2964xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002965 name = "f16_spmm_bench",
2966 srcs = [
2967 "bench/f16-spmm.cc",
2968 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002969 "src/xnnpack/AlignedAllocator.h",
2970 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08002971 deps = MICROKERNEL_BENCHMARK_DEPS,
2972)
2973
2974xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002975 name = "f32_igemm_bench",
2976 srcs = [
2977 "bench/f32-igemm.cc",
2978 "bench/conv.h",
2979 "src/xnnpack/AlignedAllocator.h",
2980 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07002981 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002982)
2983
2984xnnpack_benchmark(
2985 name = "f32_conv_hwc_bench",
2986 srcs = [
2987 "bench/f32-conv-hwc.cc",
2988 "bench/dconv.h",
2989 "src/xnnpack/AlignedAllocator.h",
2990 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002991 deps = MICROKERNEL_BENCHMARK_DEPS,
2992)
2993
2994xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07002995 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07002996 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07002997 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07002998 "bench/dconv.h",
2999 "src/xnnpack/AlignedAllocator.h",
3000 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Erich Elsen563df5f2019-10-23 08:02:21 -07003001 deps = MICROKERNEL_BENCHMARK_DEPS,
3002)
3003
3004xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003005 name = "f16_dwconv_bench",
3006 srcs = [
3007 "bench/f16-dwconv.cc",
3008 "bench/dwconv.h",
3009 "bench/google/dwconv.h",
3010 "src/xnnpack/AlignedAllocator.h",
3011 ] + MICROKERNEL_BENCHMARK_HDRS,
3012 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3013)
3014
3015xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003016 name = "f32_dwconv_bench",
3017 srcs = [
3018 "bench/f32-dwconv.cc",
3019 "bench/dwconv.h",
3020 "src/xnnpack/AlignedAllocator.h",
3021 ] + MICROKERNEL_BENCHMARK_HDRS,
3022 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3023)
3024
3025xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003026 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003027 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003028 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003029 "bench/dwconv.h",
3030 "src/xnnpack/AlignedAllocator.h",
3031 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3032 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3033)
3034
3035xnnpack_benchmark(
3036 name = "f32_gemm_bench",
3037 srcs = [
3038 "bench/f32-gemm.cc",
3039 "bench/gemm.h",
3040 "src/xnnpack/AlignedAllocator.h",
3041 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003042 copts = xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07003043 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003044)
3045
3046xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003047 name = "f32_raddexpminusmax_bench",
3048 srcs = [
3049 "bench/f32-raddexpminusmax.cc",
3050 "src/xnnpack/AlignedAllocator.h",
3051 ] + MICROKERNEL_BENCHMARK_HDRS,
3052 deps = MICROKERNEL_BENCHMARK_DEPS,
3053)
3054
3055xnnpack_benchmark(
3056 name = "f32_raddextexp_bench",
3057 srcs = [
3058 "bench/f32-raddextexp.cc",
3059 "src/xnnpack/AlignedAllocator.h",
3060 ] + MICROKERNEL_BENCHMARK_HDRS,
3061 deps = MICROKERNEL_BENCHMARK_DEPS,
3062)
3063
3064xnnpack_benchmark(
3065 name = "f32_raddstoreexpminusmax_bench",
3066 srcs = [
3067 "bench/f32-raddstoreexpminusmax.cc",
3068 "src/xnnpack/AlignedAllocator.h",
3069 ] + MICROKERNEL_BENCHMARK_HDRS,
3070 deps = MICROKERNEL_BENCHMARK_DEPS,
3071)
3072
3073xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003074 name = "f32_rmax_bench",
3075 srcs = [
3076 "bench/f32-rmax.cc",
3077 "src/xnnpack/AlignedAllocator.h",
3078 ] + MICROKERNEL_BENCHMARK_HDRS,
3079 deps = MICROKERNEL_BENCHMARK_DEPS,
3080)
3081
3082xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003083 name = "f32_sigmoid_bench",
3084 srcs = [
3085 "bench/f32-sigmoid.cc",
3086 "src/xnnpack/AlignedAllocator.h",
3087 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003088 deps = MICROKERNEL_BENCHMARK_DEPS,
3089)
3090
3091xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003092 name = "f32_spmm_bench",
3093 srcs = [
3094 "bench/f32-spmm.cc",
3095 "bench/gemm.h",
3096 "src/xnnpack/AlignedAllocator.h",
3097 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003098 deps = MICROKERNEL_BENCHMARK_DEPS,
3099)
3100
3101xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003102 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003103 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003104 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003105 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003106 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003107 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003108)
3109
3110xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003111 name = "f32_vscaleexpminusmax_bench",
3112 srcs = [
3113 "bench/f32-vscaleexpminusmax.cc",
3114 "src/xnnpack/AlignedAllocator.h",
3115 ] + MICROKERNEL_BENCHMARK_HDRS,
3116 deps = MICROKERNEL_BENCHMARK_DEPS,
3117)
3118
3119xnnpack_benchmark(
3120 name = "f32_vscaleextexp_bench",
3121 srcs = [
3122 "bench/f32-vscaleextexp.cc",
3123 "src/xnnpack/AlignedAllocator.h",
3124 ] + MICROKERNEL_BENCHMARK_HDRS,
3125 deps = MICROKERNEL_BENCHMARK_DEPS,
3126)
3127
3128xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003129 name = "f32_im2col_gemm_bench",
3130 srcs = [
3131 "bench/f32-im2col-gemm.cc",
3132 "bench/conv.h",
3133 "src/xnnpack/AlignedAllocator.h",
3134 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3135 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
3136)
3137
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003138xnnpack_benchmark(
3139 name = "requantization_bench",
3140 srcs = [
3141 "bench/requantization.cc",
3142 "src/xnnpack/requantization-stubs.h",
3143 "src/xnnpack/AlignedAllocator.h",
3144 ] + MICROKERNEL_BENCHMARK_HDRS,
3145 deps = MICROKERNEL_BENCHMARK_DEPS,
3146)
3147
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003148xnnpack_benchmark(
3149 name = "rounding_bench",
3150 srcs = [
3151 "bench/rounding.cc",
3152 "src/xnnpack/math-stubs.h",
3153 "src/xnnpack/AlignedAllocator.h",
3154 ] + MICROKERNEL_BENCHMARK_HDRS,
3155 deps = MICROKERNEL_BENCHMARK_DEPS,
3156)
3157
Marat Dukhan08c4a432019-10-03 09:29:21 -07003158########################### Benchmarks for operators ###########################
3159
3160xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003161 name = "average_pooling_bench",
3162 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003163 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003164 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003165 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003166)
3167
3168xnnpack_benchmark(
3169 name = "channel_shuffle_bench",
3170 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003171 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003172)
3173
3174xnnpack_benchmark(
3175 name = "convolution_bench",
3176 srcs = ["bench/convolution.cc"],
3177 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003178 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003179 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003180)
3181
3182xnnpack_benchmark(
3183 name = "deconvolution_bench",
3184 srcs = ["bench/deconvolution.cc"],
3185 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003186 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003187 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003188)
3189
3190xnnpack_benchmark(
3191 name = "global_average_pooling_bench",
3192 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003193 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003194)
3195
3196xnnpack_benchmark(
3197 name = "max_pooling_bench",
3198 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003199 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003200)
3201
3202xnnpack_benchmark(
3203 name = "sigmoid_bench",
3204 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003205 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003206 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003207 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003208)
3209
3210xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003211 name = "prelu_bench",
3212 srcs = ["bench/prelu.cc"],
3213 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003214 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003215 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003216)
3217
3218xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003219 name = "softmax_bench",
3220 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003221 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003222 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003223 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003224)
3225
Marat Dukhanc068bb62019-10-04 13:24:39 -07003226############################# End-to-end benchmarks ############################
3227
3228cc_library(
3229 name = "mobilenet_v1",
3230 srcs = ["models/mobilenet-v1.cc"],
3231 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003232 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003233 linkstatic = True,
3234 deps = [
3235 ":XNNPACK",
3236 "@pthreadpool",
3237 ],
3238)
3239
3240cc_library(
3241 name = "mobilenet_v2",
3242 srcs = ["models/mobilenet-v2.cc"],
3243 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003244 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003245 linkstatic = True,
3246 deps = [
3247 ":XNNPACK",
3248 "@pthreadpool",
3249 ],
3250)
3251
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003252cc_library(
3253 name = "mobilenet_v3_large",
3254 srcs = ["models/mobilenet-v3-large.cc"],
3255 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003256 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003257 linkstatic = True,
3258 deps = [
3259 ":XNNPACK",
3260 "@pthreadpool",
3261 ],
3262)
3263
3264cc_library(
3265 name = "mobilenet_v3_small",
3266 srcs = ["models/mobilenet-v3-small.cc"],
3267 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003268 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003269 linkstatic = True,
3270 deps = [
3271 ":XNNPACK",
3272 "@pthreadpool",
3273 ],
3274)
3275
Marat Dukhanc068bb62019-10-04 13:24:39 -07003276xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07003277 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003278 srcs = [
3279 "bench/f32-dwconv-e2e.cc",
3280 "bench/end2end.h",
3281 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07003282 deps = MICROKERNEL_BENCHMARK_DEPS + [
3283 ":XNNPACK",
3284 ":mobilenet_v1",
3285 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003286 ":mobilenet_v3_large",
3287 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07003288 ],
3289)
3290
3291xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07003292 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003293 srcs = [
3294 "bench/f32-gemm-e2e.cc",
3295 "bench/end2end.h",
3296 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07003297 deps = MICROKERNEL_BENCHMARK_DEPS + [
3298 ":XNNPACK",
3299 ":mobilenet_v1",
3300 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003301 ":mobilenet_v3_large",
3302 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07003303 ],
3304)
3305
3306xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07003307 name = "end2end_bench",
3308 srcs = ["bench/end2end.cc"],
3309 deps = [
3310 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07003311 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07003312 ":mobilenet_v1",
3313 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003314 ":mobilenet_v3_large",
3315 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07003316 "@pthreadpool",
3317 ],
3318)
3319
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003320#################### Accuracy evaluation for math functions ####################
3321
3322xnnpack_benchmark(
3323 name = "f32_exp_eval",
3324 srcs = [
3325 "eval/f32-exp.cc",
3326 "src/xnnpack/AlignedAllocator.h",
3327 ] + ACCURACY_EVAL_HDRS,
3328 deps = ACCURACY_EVAL_DEPS,
3329)
3330
Marat Dukhan515c9772019-10-17 18:07:57 -07003331xnnpack_benchmark(
3332 name = "f32_expminus_eval",
3333 srcs = [
3334 "eval/f32-expminus.cc",
3335 "src/xnnpack/AlignedAllocator.h",
3336 ] + ACCURACY_EVAL_HDRS,
3337 deps = ACCURACY_EVAL_DEPS,
3338)
3339
Marat Dukhan98ba4412019-10-23 02:14:28 -07003340xnnpack_benchmark(
3341 name = "f32_extexp_eval",
3342 srcs = [
3343 "eval/f32-extexp.cc",
3344 "src/xnnpack/AlignedAllocator.h",
3345 ] + ACCURACY_EVAL_HDRS,
3346 deps = ACCURACY_EVAL_DEPS,
3347)
3348
Marat Dukhan8853b822020-05-07 12:19:01 -07003349xnnpack_unit_test(
3350 name = "f32_roundne_eval",
3351 srcs = [
3352 "eval/f32-roundne.cc",
3353 "src/xnnpack/AlignedAllocator.h",
3354 "src/xnnpack/math-stubs.h",
3355 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07003356 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07003357 deps = MICROKERNEL_TEST_DEPS,
3358)
3359
Marat Dukhan2dbb9442020-05-12 20:43:43 -07003360xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07003361 name = "f32_roundd_eval",
3362 srcs = [
3363 "eval/f32-roundd.cc",
3364 "src/xnnpack/AlignedAllocator.h",
3365 "src/xnnpack/math-stubs.h",
3366 ] + MICROKERNEL_TEST_HDRS,
3367 automatic = False,
3368 deps = MICROKERNEL_TEST_DEPS,
3369)
3370
3371xnnpack_unit_test(
3372 name = "f32_roundu_eval",
3373 srcs = [
3374 "eval/f32-roundu.cc",
3375 "src/xnnpack/AlignedAllocator.h",
3376 "src/xnnpack/math-stubs.h",
3377 ] + MICROKERNEL_TEST_HDRS,
3378 automatic = False,
3379 deps = MICROKERNEL_TEST_DEPS,
3380)
3381
3382xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07003383 name = "f32_roundz_eval",
3384 srcs = [
3385 "eval/f32-roundz.cc",
3386 "src/xnnpack/AlignedAllocator.h",
3387 "src/xnnpack/math-stubs.h",
3388 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07003389 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07003390 deps = MICROKERNEL_TEST_DEPS,
3391)
3392
Marat Dukhan346a9e52019-11-15 09:06:30 -08003393xnnpack_benchmark(
3394 name = "f32_sigmoid_eval",
3395 srcs = [
3396 "eval/f32-sigmoid.cc",
3397 "src/xnnpack/AlignedAllocator.h",
3398 ] + ACCURACY_EVAL_HDRS,
3399 deps = ACCURACY_EVAL_DEPS,
3400)
3401
Marat Dukhan08c4a432019-10-03 09:29:21 -07003402######################### Unit tests for micro-kernels #########################
3403
3404xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003405 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003406 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003407 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003408 "test/gemm-microkernel-tester.h",
3409 "src/xnnpack/AlignedAllocator.h",
3410 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3411 deps = MICROKERNEL_TEST_DEPS,
3412)
3413
3414xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07003415 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003416 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07003417 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003418 "test/spmm-microkernel-tester.h",
3419 "src/xnnpack/AlignedAllocator.h",
3420 ] + MICROKERNEL_TEST_HDRS,
3421 deps = MICROKERNEL_TEST_DEPS,
3422)
3423
3424xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003425 name = "f32_argmaxpool_test",
3426 srcs = [
3427 "test/f32-argmaxpool.cc",
3428 "test/argmaxpool-microkernel-tester.h",
3429 "src/xnnpack/AlignedAllocator.h",
3430 ] + MICROKERNEL_TEST_HDRS,
3431 deps = MICROKERNEL_TEST_DEPS,
3432)
3433
3434xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003435 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003436 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003437 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003438 "test/avgpool-microkernel-tester.h",
3439 "src/xnnpack/AlignedAllocator.h",
3440 ] + MICROKERNEL_TEST_HDRS,
3441 deps = MICROKERNEL_TEST_DEPS,
3442)
3443
3444xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07003445 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08003446 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07003447 "test/f32-ibilinear.cc",
3448 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08003449 "src/xnnpack/AlignedAllocator.h",
3450 ] + MICROKERNEL_TEST_HDRS,
3451 deps = MICROKERNEL_TEST_DEPS,
3452)
3453
3454xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07003455 name = "f16_clamp_test",
3456 srcs = [
3457 "test/f16-clamp.cc",
3458 "test/clamp-microkernel-tester.h",
3459 ] + MICROKERNEL_TEST_HDRS,
3460 deps = MICROKERNEL_TEST_DEPS,
3461)
3462
3463xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003464 name = "f32_clamp_test",
3465 srcs = [
3466 "test/f32-clamp.cc",
3467 "test/clamp-microkernel-tester.h",
3468 ] + MICROKERNEL_TEST_HDRS,
3469 deps = MICROKERNEL_TEST_DEPS,
3470)
3471
3472xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07003473 name = "f32_igemm_test",
3474 srcs = [
3475 "test/f32-igemm.cc",
3476 "test/gemm-microkernel-tester.h",
3477 "src/xnnpack/AlignedAllocator.h",
3478 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3479 deps = MICROKERNEL_TEST_DEPS,
3480)
3481
3482xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07003483 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003484 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07003485 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003486 "test/gemm-microkernel-tester.h",
3487 "src/xnnpack/AlignedAllocator.h",
3488 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3489 deps = MICROKERNEL_TEST_DEPS,
3490)
3491
3492xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07003493 name = "f32_igemm_minmax_test",
3494 srcs = [
3495 "test/f32-igemm-minmax.cc",
3496 "test/gemm-microkernel-tester.h",
3497 "src/xnnpack/AlignedAllocator.h",
3498 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3499 deps = MICROKERNEL_TEST_DEPS,
3500)
3501
3502xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07003503 name = "f16_igemm_minmax_test",
3504 srcs = [
3505 "test/f16-igemm-minmax.cc",
3506 "test/gemm-microkernel-tester.h",
3507 "src/xnnpack/AlignedAllocator.h",
3508 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3509 deps = MICROKERNEL_TEST_DEPS,
3510)
3511
3512xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003513 name = "f32_conv_hwc_test",
3514 srcs = [
3515 "test/f32-conv-hwc.cc",
3516 "test/conv-hwc-microkernel-tester.h",
3517 "src/xnnpack/AlignedAllocator.h",
3518 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3519 deps = MICROKERNEL_TEST_DEPS,
3520)
3521
3522xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003523 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003524 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003525 "test/f32-conv-hwc2chw.cc",
3526 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003527 "src/xnnpack/AlignedAllocator.h",
3528 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3529 deps = MICROKERNEL_TEST_DEPS,
3530)
3531
3532xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07003533 name = "f32_dwconv_test",
3534 srcs = [
3535 "test/f32-dwconv.cc",
3536 "test/dwconv-microkernel-tester.h",
3537 "src/xnnpack/AlignedAllocator.h",
3538 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3539 deps = MICROKERNEL_TEST_DEPS,
3540)
3541
3542xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07003543 name = "f16_dwconv_minmax_test",
3544 srcs = [
3545 "test/f16-dwconv-minmax.cc",
3546 "test/dwconv-microkernel-tester.h",
3547 "src/xnnpack/AlignedAllocator.h",
3548 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3549 deps = MICROKERNEL_TEST_DEPS,
3550)
3551
3552xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07003553 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003554 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003555 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003556 "test/dwconv-microkernel-tester.h",
3557 "src/xnnpack/AlignedAllocator.h",
3558 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3559 deps = MICROKERNEL_TEST_DEPS,
3560)
3561
3562xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003563 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003564 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003565 "test/f32-dwconv-chw.cc",
3566 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003567 "src/xnnpack/AlignedAllocator.h",
3568 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3569 deps = MICROKERNEL_TEST_DEPS,
3570)
3571
3572xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07003573 name = "f16_gavgpool_minmax_test",
3574 srcs = [
3575 "test/f16-gavgpool-minmax.cc",
3576 "test/gavgpool-microkernel-tester.h",
3577 "src/xnnpack/AlignedAllocator.h",
3578 ] + MICROKERNEL_TEST_HDRS,
3579 deps = MICROKERNEL_TEST_DEPS,
3580)
3581
3582xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003583 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003584 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003585 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003586 "test/gavgpool-microkernel-tester.h",
3587 "src/xnnpack/AlignedAllocator.h",
3588 ] + MICROKERNEL_TEST_HDRS,
3589 deps = MICROKERNEL_TEST_DEPS,
3590)
3591
3592xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003593 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003594 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003595 "test/f32-gavgpool-cw.cc",
3596 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003597 "src/xnnpack/AlignedAllocator.h",
3598 ] + MICROKERNEL_TEST_HDRS,
3599 deps = MICROKERNEL_TEST_DEPS,
3600)
3601
3602xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07003603 name = "f32_gemm_test",
3604 srcs = [
3605 "test/f32-gemm.cc",
3606 "test/gemm-microkernel-tester.h",
3607 "src/xnnpack/AlignedAllocator.h",
3608 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3609 deps = MICROKERNEL_TEST_DEPS,
3610)
3611
3612xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07003613 name = "f32_gemm_relu_test",
3614 srcs = [
3615 "test/f32-gemm-relu.cc",
3616 "test/gemm-microkernel-tester.h",
3617 "src/xnnpack/AlignedAllocator.h",
3618 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3619 deps = MICROKERNEL_TEST_DEPS,
3620)
3621
3622xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07003623 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003624 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003625 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003626 "test/gemm-microkernel-tester.h",
3627 "src/xnnpack/AlignedAllocator.h",
3628 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3629 deps = MICROKERNEL_TEST_DEPS,
3630)
3631
3632xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07003633 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003634 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003635 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003636 "test/gemm-microkernel-tester.h",
3637 "src/xnnpack/AlignedAllocator.h",
3638 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3639 deps = MICROKERNEL_TEST_DEPS,
3640)
3641
3642xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07003643 name = "f16_hswish_test",
3644 srcs = [
3645 "test/f16-hswish.cc",
3646 "test/hswish-microkernel-tester.h",
3647 ] + MICROKERNEL_TEST_HDRS,
3648 deps = MICROKERNEL_TEST_DEPS,
3649)
3650
3651xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003652 name = "f32_hswish_test",
3653 srcs = [
3654 "test/f32-hswish.cc",
3655 "test/hswish-microkernel-tester.h",
3656 ] + MICROKERNEL_TEST_HDRS,
3657 deps = MICROKERNEL_TEST_DEPS,
3658)
3659
3660xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003661 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003662 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003663 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003664 "test/maxpool-microkernel-tester.h",
3665 ] + MICROKERNEL_TEST_HDRS,
3666 deps = MICROKERNEL_TEST_DEPS,
3667)
3668
3669xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003670 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003671 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003672 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003673 "test/avgpool-microkernel-tester.h",
3674 "src/xnnpack/AlignedAllocator.h",
3675 ] + MICROKERNEL_TEST_HDRS,
3676 deps = MICROKERNEL_TEST_DEPS,
3677)
3678
3679xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07003680 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003681 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003682 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003683 "test/gemm-microkernel-tester.h",
3684 "src/xnnpack/AlignedAllocator.h",
3685 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3686 deps = MICROKERNEL_TEST_DEPS,
3687)
3688
3689xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07003690 name = "f16_prelu_test",
3691 srcs = [
3692 "test/f16-prelu.cc",
3693 "test/prelu-microkernel-tester.h",
3694 "src/xnnpack/AlignedAllocator.h",
3695 ] + MICROKERNEL_TEST_HDRS,
3696 deps = MICROKERNEL_TEST_DEPS,
3697)
3698
3699xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003700 name = "f32_prelu_test",
3701 srcs = [
3702 "test/f32-prelu.cc",
3703 "test/prelu-microkernel-tester.h",
3704 "src/xnnpack/AlignedAllocator.h",
3705 ] + MICROKERNEL_TEST_HDRS,
3706 deps = MICROKERNEL_TEST_DEPS,
3707)
3708
3709xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07003710 name = "f32_raddexpminusmax_test",
3711 srcs = [
3712 "test/f32-raddexpminusmax.cc",
3713 "test/raddexpminusmax-microkernel-tester.h",
3714 ] + MICROKERNEL_TEST_HDRS,
3715 deps = MICROKERNEL_TEST_DEPS,
3716)
3717
3718xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003719 name = "f32_raddextexp_test",
3720 srcs = [
3721 "test/f32-raddextexp.cc",
3722 "test/raddextexp-microkernel-tester.h",
3723 ] + MICROKERNEL_TEST_HDRS,
3724 deps = MICROKERNEL_TEST_DEPS,
3725)
3726
3727xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07003728 name = "f32_raddstoreexpminusmax_test",
3729 srcs = [
3730 "test/f32-raddstoreexpminusmax.cc",
3731 "test/raddstoreexpminusmax-microkernel-tester.h",
3732 ] + MICROKERNEL_TEST_HDRS,
3733 deps = MICROKERNEL_TEST_DEPS,
3734)
3735
3736xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003737 name = "f32_rmax_test",
3738 srcs = [
3739 "test/f32-rmax.cc",
3740 "test/rmax-microkernel-tester.h",
3741 ] + MICROKERNEL_TEST_HDRS,
3742 deps = MICROKERNEL_TEST_DEPS,
3743)
3744
3745xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08003746 name = "f32_sigmoid_test",
3747 srcs = [
3748 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003749 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08003750 ] + MICROKERNEL_TEST_HDRS,
3751 deps = MICROKERNEL_TEST_DEPS,
3752)
3753
3754xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07003755 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003756 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07003757 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003758 "test/spmm-microkernel-tester.h",
3759 "src/xnnpack/AlignedAllocator.h",
3760 ] + MICROKERNEL_TEST_HDRS,
3761 deps = MICROKERNEL_TEST_DEPS,
3762)
3763
3764xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07003765 name = "f16_vadd_minmax_test",
3766 srcs = [
3767 "test/f16-vadd-minmax.cc",
3768 "test/vbinary-microkernel-tester.h",
3769 ] + MICROKERNEL_TEST_HDRS,
3770 deps = MICROKERNEL_TEST_DEPS,
3771)
3772
3773xnnpack_unit_test(
3774 name = "f16_vaddc_minmax_test",
3775 srcs = [
3776 "test/f16-vaddc-minmax.cc",
3777 "test/vbinaryc-microkernel-tester.h",
3778 ] + MICROKERNEL_TEST_HDRS,
3779 deps = MICROKERNEL_TEST_DEPS,
3780)
3781
3782xnnpack_unit_test(
3783 name = "f16_vdiv_minmax_test",
3784 srcs = [
3785 "test/f16-vdiv-minmax.cc",
3786 "test/vbinary-microkernel-tester.h",
3787 ] + MICROKERNEL_TEST_HDRS,
3788 deps = MICROKERNEL_TEST_DEPS,
3789)
3790
3791xnnpack_unit_test(
3792 name = "f16_vdivc_minmax_test",
3793 srcs = [
3794 "test/f16-vdivc-minmax.cc",
3795 "test/vbinaryc-microkernel-tester.h",
3796 ] + MICROKERNEL_TEST_HDRS,
3797 deps = MICROKERNEL_TEST_DEPS,
3798)
3799
3800xnnpack_unit_test(
3801 name = "f16_vrdivc_minmax_test",
3802 srcs = [
3803 "test/f16-vrdivc-minmax.cc",
3804 "test/vbinaryc-microkernel-tester.h",
3805 ] + MICROKERNEL_TEST_HDRS,
3806 deps = MICROKERNEL_TEST_DEPS,
3807)
3808
3809xnnpack_unit_test(
3810 name = "f16_vmax_test",
3811 srcs = [
3812 "test/f16-vmax.cc",
3813 "test/vbinary-microkernel-tester.h",
3814 ] + MICROKERNEL_TEST_HDRS,
3815 deps = MICROKERNEL_TEST_DEPS,
3816)
3817
3818xnnpack_unit_test(
3819 name = "f16_vmaxc_test",
3820 srcs = [
3821 "test/f16-vmaxc.cc",
3822 "test/vbinaryc-microkernel-tester.h",
3823 ] + MICROKERNEL_TEST_HDRS,
3824 deps = MICROKERNEL_TEST_DEPS,
3825)
3826
3827xnnpack_unit_test(
3828 name = "f16_vmin_test",
3829 srcs = [
3830 "test/f16-vmin.cc",
3831 "test/vbinary-microkernel-tester.h",
3832 ] + MICROKERNEL_TEST_HDRS,
3833 deps = MICROKERNEL_TEST_DEPS,
3834)
3835
3836xnnpack_unit_test(
3837 name = "f16_vminc_test",
3838 srcs = [
3839 "test/f16-vminc.cc",
3840 "test/vbinaryc-microkernel-tester.h",
3841 ] + MICROKERNEL_TEST_HDRS,
3842 deps = MICROKERNEL_TEST_DEPS,
3843)
3844
3845xnnpack_unit_test(
3846 name = "f16_vmul_minmax_test",
3847 srcs = [
3848 "test/f16-vmul-minmax.cc",
3849 "test/vbinary-microkernel-tester.h",
3850 ] + MICROKERNEL_TEST_HDRS,
3851 deps = MICROKERNEL_TEST_DEPS,
3852)
3853
3854xnnpack_unit_test(
3855 name = "f16_vmulc_minmax_test",
3856 srcs = [
3857 "test/f16-vmulc-minmax.cc",
3858 "test/vbinaryc-microkernel-tester.h",
3859 ] + MICROKERNEL_TEST_HDRS,
3860 deps = MICROKERNEL_TEST_DEPS,
3861)
3862
3863xnnpack_unit_test(
3864 name = "f16_vsub_minmax_test",
3865 srcs = [
3866 "test/f16-vsub-minmax.cc",
3867 "test/vbinary-microkernel-tester.h",
3868 ] + MICROKERNEL_TEST_HDRS,
3869 deps = MICROKERNEL_TEST_DEPS,
3870)
3871
3872xnnpack_unit_test(
3873 name = "f16_vsubc_minmax_test",
3874 srcs = [
3875 "test/f16-vsubc-minmax.cc",
3876 "test/vbinaryc-microkernel-tester.h",
3877 ] + MICROKERNEL_TEST_HDRS,
3878 deps = MICROKERNEL_TEST_DEPS,
3879)
3880
3881xnnpack_unit_test(
3882 name = "f16_vrsubc_minmax_test",
3883 srcs = [
3884 "test/f16-vrsubc-minmax.cc",
3885 "test/vbinaryc-microkernel-tester.h",
3886 ] + MICROKERNEL_TEST_HDRS,
3887 deps = MICROKERNEL_TEST_DEPS,
3888)
3889
3890xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07003891 name = "f32_vabs_test",
3892 srcs = [
3893 "test/f32-vabs.cc",
3894 "test/vunary-microkernel-tester.h",
3895 ] + MICROKERNEL_TEST_HDRS,
3896 deps = MICROKERNEL_TEST_DEPS,
3897)
3898
3899xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003900 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003901 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003902 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003903 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003904 ] + MICROKERNEL_TEST_HDRS,
3905 deps = MICROKERNEL_TEST_DEPS,
3906)
3907
3908xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003909 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003910 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003911 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003912 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003913 ] + MICROKERNEL_TEST_HDRS,
3914 deps = MICROKERNEL_TEST_DEPS,
3915)
3916
3917xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003918 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003919 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003920 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003921 "test/vbinary-microkernel-tester.h",
3922 ] + MICROKERNEL_TEST_HDRS,
3923 deps = MICROKERNEL_TEST_DEPS,
3924)
3925
3926xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003927 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003928 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003929 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003930 "test/vbinaryc-microkernel-tester.h",
3931 ] + MICROKERNEL_TEST_HDRS,
3932 deps = MICROKERNEL_TEST_DEPS,
3933)
3934
3935xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003936 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003937 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003938 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08003939 "test/vbinaryc-microkernel-tester.h",
3940 ] + MICROKERNEL_TEST_HDRS,
3941 deps = MICROKERNEL_TEST_DEPS,
3942)
3943
3944xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08003945 name = "f32_vmax_test",
3946 srcs = [
3947 "test/f32-vmax.cc",
3948 "test/vbinary-microkernel-tester.h",
3949 ] + MICROKERNEL_TEST_HDRS,
3950 deps = MICROKERNEL_TEST_DEPS,
3951)
3952
3953xnnpack_unit_test(
3954 name = "f32_vmaxc_test",
3955 srcs = [
3956 "test/f32-vmaxc.cc",
3957 "test/vbinaryc-microkernel-tester.h",
3958 ] + MICROKERNEL_TEST_HDRS,
3959 deps = MICROKERNEL_TEST_DEPS,
3960)
3961
3962xnnpack_unit_test(
3963 name = "f32_vmin_test",
3964 srcs = [
3965 "test/f32-vmin.cc",
3966 "test/vbinary-microkernel-tester.h",
3967 ] + MICROKERNEL_TEST_HDRS,
3968 deps = MICROKERNEL_TEST_DEPS,
3969)
3970
3971xnnpack_unit_test(
3972 name = "f32_vminc_test",
3973 srcs = [
3974 "test/f32-vminc.cc",
3975 "test/vbinaryc-microkernel-tester.h",
3976 ] + MICROKERNEL_TEST_HDRS,
3977 deps = MICROKERNEL_TEST_DEPS,
3978)
3979
3980xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003981 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003982 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003983 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003984 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003985 ] + MICROKERNEL_TEST_HDRS,
3986 deps = MICROKERNEL_TEST_DEPS,
3987)
3988
3989xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003990 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08003991 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003992 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003993 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003994 ] + MICROKERNEL_TEST_HDRS,
3995 deps = MICROKERNEL_TEST_DEPS,
3996)
3997
3998xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07003999 name = "f16_vmulcaddc_minmax_test",
4000 srcs = [
4001 "test/f16-vmulcaddc-minmax.cc",
4002 "test/vmulcaddc-microkernel-tester.h",
4003 "src/xnnpack/AlignedAllocator.h",
4004 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4005 deps = MICROKERNEL_TEST_DEPS,
4006)
4007
4008xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004009 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004010 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004011 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004012 "test/vmulcaddc-microkernel-tester.h",
4013 "src/xnnpack/AlignedAllocator.h",
4014 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4015 deps = MICROKERNEL_TEST_DEPS,
4016)
4017
4018xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004019 name = "f32_vneg_test",
4020 srcs = [
4021 "test/f32-vneg.cc",
4022 "test/vunary-microkernel-tester.h",
4023 ] + MICROKERNEL_TEST_HDRS,
4024 deps = MICROKERNEL_TEST_DEPS,
4025)
4026
4027xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004028 name = "f32_vscale_test",
4029 srcs = [
4030 "test/f32-vscale.cc",
4031 "test/vscale-microkernel-tester.h",
4032 ] + MICROKERNEL_TEST_HDRS,
4033 deps = MICROKERNEL_TEST_DEPS,
4034)
4035
4036xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004037 name = "f32_vscaleexpminusmax_test",
4038 srcs = [
4039 "test/f32-vscaleexpminusmax.cc",
4040 "test/vscaleexpminusmax-microkernel-tester.h",
4041 ] + MICROKERNEL_TEST_HDRS,
4042 deps = MICROKERNEL_TEST_DEPS,
4043)
4044
4045xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004046 name = "f32_vscaleextexp_test",
4047 srcs = [
4048 "test/f32-vscaleextexp.cc",
4049 "test/vscaleextexp-microkernel-tester.h",
4050 ] + MICROKERNEL_TEST_HDRS,
4051 deps = MICROKERNEL_TEST_DEPS,
4052)
4053
4054xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004055 name = "f32_vsqr_test",
4056 srcs = [
4057 "test/f32-vsqr.cc",
4058 "test/vunary-microkernel-tester.h",
4059 ] + MICROKERNEL_TEST_HDRS,
4060 deps = MICROKERNEL_TEST_DEPS,
4061)
4062
4063xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004064 name = "f32_vsqrdiff_test",
4065 srcs = [
4066 "test/f32-vsqrdiff.cc",
4067 "test/vbinary-microkernel-tester.h",
4068 ] + MICROKERNEL_TEST_HDRS,
4069 deps = MICROKERNEL_TEST_DEPS,
4070)
4071
4072xnnpack_unit_test(
4073 name = "f32_vsqrdiffc_test",
4074 srcs = [
4075 "test/f32-vsqrdiffc.cc",
4076 "test/vbinaryc-microkernel-tester.h",
4077 ] + MICROKERNEL_TEST_HDRS,
4078 deps = MICROKERNEL_TEST_DEPS,
4079)
4080
4081xnnpack_unit_test(
4082 name = "f32_vrsqrdiffc_test",
4083 srcs = [
4084 "test/f32-vrsqrdiffc.cc",
4085 "test/vbinaryc-microkernel-tester.h",
4086 ] + MICROKERNEL_TEST_HDRS,
4087 deps = MICROKERNEL_TEST_DEPS,
4088)
4089
4090xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004091 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07004092 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004093 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004094 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004095 ] + MICROKERNEL_TEST_HDRS,
4096 deps = MICROKERNEL_TEST_DEPS,
4097)
4098
4099xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004100 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004101 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004102 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004103 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004104 ] + MICROKERNEL_TEST_HDRS,
4105 deps = MICROKERNEL_TEST_DEPS,
4106)
4107
4108xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004109 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004110 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004111 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004112 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07004113 ] + MICROKERNEL_TEST_HDRS,
4114 deps = MICROKERNEL_TEST_DEPS,
4115)
4116
4117xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004118 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004119 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004120 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004121 "test/avgpool-microkernel-tester.h",
4122 "src/xnnpack/AlignedAllocator.h",
4123 ] + MICROKERNEL_TEST_HDRS,
4124 deps = MICROKERNEL_TEST_DEPS,
4125)
4126
4127xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004128 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004129 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004130 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004131 "test/gemm-microkernel-tester.h",
4132 "src/xnnpack/AlignedAllocator.h",
4133 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4134 deps = MICROKERNEL_TEST_DEPS,
4135)
4136
4137xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004138 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004139 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004140 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004141 "test/dwconv-microkernel-tester.h",
4142 "src/xnnpack/AlignedAllocator.h",
4143 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4144 deps = MICROKERNEL_TEST_DEPS,
4145)
4146
4147xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004148 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004149 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004150 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004151 "test/gavgpool-microkernel-tester.h",
4152 "src/xnnpack/AlignedAllocator.h",
4153 ] + MICROKERNEL_TEST_HDRS,
4154 deps = MICROKERNEL_TEST_DEPS,
4155)
4156
4157xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004158 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004159 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004160 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004161 "test/gemm-microkernel-tester.h",
4162 "src/xnnpack/AlignedAllocator.h",
4163 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4164 deps = MICROKERNEL_TEST_DEPS,
4165)
4166
4167xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004168 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004169 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004170 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004171 "test/vadd-microkernel-tester.h",
4172 ] + MICROKERNEL_TEST_HDRS,
4173 deps = MICROKERNEL_TEST_DEPS,
4174)
4175
4176xnnpack_unit_test(
4177 name = "u8_clamp_test",
4178 srcs = [
4179 "test/u8-clamp.cc",
4180 "test/clamp-microkernel-tester.h",
4181 ] + MICROKERNEL_TEST_HDRS,
4182 deps = MICROKERNEL_TEST_DEPS,
4183)
4184
4185xnnpack_unit_test(
4186 name = "u8_lut32norm_test",
4187 srcs = [
4188 "test/u8-lut32norm.cc",
4189 "test/lut-norm-microkernel-tester.h",
4190 ] + MICROKERNEL_TEST_HDRS,
4191 deps = MICROKERNEL_TEST_DEPS,
4192)
4193
4194xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004195 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004196 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004197 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004198 "test/maxpool-microkernel-tester.h",
4199 ] + MICROKERNEL_TEST_HDRS,
4200 deps = MICROKERNEL_TEST_DEPS,
4201)
4202
4203xnnpack_unit_test(
4204 name = "u8_rmax_test",
4205 srcs = [
4206 "test/u8-rmax.cc",
4207 "test/rmax-microkernel-tester.h",
4208 ] + MICROKERNEL_TEST_HDRS,
4209 deps = MICROKERNEL_TEST_DEPS,
4210)
4211
4212xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07004213 name = "x32_fill_test",
4214 srcs = [
4215 "test/x32-fill.cc",
4216 "test/fill-microkernel-tester.h",
4217 ] + MICROKERNEL_TEST_HDRS,
4218 deps = MICROKERNEL_TEST_DEPS,
4219)
4220
4221xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004222 name = "x32_packx_test",
4223 srcs = [
4224 "test/x32-packx.cc",
4225 "test/pack-microkernel-tester.h",
4226 "src/xnnpack/AlignedAllocator.h",
4227 ] + MICROKERNEL_TEST_HDRS,
4228 deps = MICROKERNEL_TEST_DEPS,
4229)
4230
4231xnnpack_unit_test(
4232 name = "x32_pad_test",
4233 srcs = [
4234 "test/x32-pad.cc",
4235 "test/pad-microkernel-tester.h",
4236 ] + MICROKERNEL_TEST_HDRS,
4237 deps = MICROKERNEL_TEST_DEPS,
4238)
4239
4240xnnpack_unit_test(
4241 name = "x32_unpool_test",
4242 srcs = [
4243 "test/x32-unpool.cc",
4244 "test/unpool-microkernel-tester.h",
4245 ] + MICROKERNEL_TEST_HDRS,
4246 deps = MICROKERNEL_TEST_DEPS,
4247)
4248
4249xnnpack_unit_test(
4250 name = "x32_zip_test",
4251 srcs = [
4252 "test/x32-zip.cc",
4253 "test/zip-microkernel-tester.h",
4254 ] + MICROKERNEL_TEST_HDRS,
4255 deps = MICROKERNEL_TEST_DEPS,
4256)
4257
4258xnnpack_unit_test(
4259 name = "x8_lut_test",
4260 srcs = [
4261 "test/x8-lut.cc",
4262 "test/lut-microkernel-tester.h",
4263 ] + MICROKERNEL_TEST_HDRS,
4264 deps = MICROKERNEL_TEST_DEPS,
4265)
4266
4267xnnpack_unit_test(
4268 name = "x8_zip_test",
4269 srcs = [
4270 "test/x8-zip.cc",
4271 "test/zip-microkernel-tester.h",
4272 ] + MICROKERNEL_TEST_HDRS,
4273 deps = MICROKERNEL_TEST_DEPS,
4274)
4275
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004276xnnpack_unit_test(
4277 name = "requantization_test",
4278 srcs = [
4279 "src/xnnpack/requantization-stubs.h",
4280 "test/requantization.cc",
4281 "test/requantization-tester.h",
4282 ] + MICROKERNEL_TEST_HDRS,
4283 deps = MICROKERNEL_TEST_DEPS,
4284)
4285
Marat Dukhan20c3b922020-03-10 03:45:06 -07004286########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004287
4288xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07004289 name = "operator_size_test",
4290 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004291 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004292)
4293
Marat Dukhan20c3b922020-03-10 03:45:06 -07004294xnnpack_binary(
4295 name = "subgraph_size_test",
4296 srcs = ["test/subgraph-size.c"],
4297 deps = [":XNNPACK"],
4298)
4299
4300########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004301
4302xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004303 name = "abs_nc_test",
4304 srcs = [
4305 "test/abs-nc.cc",
4306 "test/abs-operator-tester.h",
4307 ],
4308 deps = OPERATOR_TEST_DEPS,
4309)
4310
4311xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08004312 name = "add_nd_test",
4313 srcs = [
4314 "test/add-nd.cc",
4315 "test/binary-elementwise-operator-tester.h",
4316 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004317 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08004318)
4319
4320xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004321 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004322 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004323 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004324 "test/argmax-pooling-operator-tester.h",
4325 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004326 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004327)
4328
4329xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004330 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004331 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004332 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004333 "test/average-pooling-operator-tester.h",
4334 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004335 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004336)
4337
4338xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004339 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004340 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004341 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004342 "test/channel-shuffle-operator-tester.h",
4343 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004344 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004345)
4346
4347xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004348 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004349 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004350 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004351 "test/clamp-operator-tester.h",
4352 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004353 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004354)
4355
4356xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07004357 name = "constant_pad_nd_test",
4358 srcs = [
4359 "test/constant-pad-nd.cc",
4360 "test/constant-pad-operator-tester.h",
4361 ],
4362 deps = OPERATOR_TEST_DEPS,
4363)
4364
4365xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004366 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004367 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004368 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004369 "test/convolution-operator-tester.h",
4370 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004371 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004372)
4373
4374xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004375 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004376 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004377 "test/convolution-nchw.cc",
4378 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004379 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004380 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004381)
4382
4383xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07004384 name = "copy_nc_test",
4385 srcs = [
4386 "test/copy-nc.cc",
4387 "test/copy-operator-tester.h",
4388 ],
4389 deps = OPERATOR_TEST_DEPS,
4390)
4391
4392xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004393 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004394 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004395 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004396 "test/deconvolution-operator-tester.h",
4397 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004398 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004399)
4400
4401xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08004402 name = "divide_nd_test",
4403 srcs = [
4404 "test/binary-elementwise-operator-tester.h",
4405 "test/divide-nd.cc",
4406 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004407 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08004408)
4409
4410xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004411 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004412 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004413 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004414 "test/fully-connected-operator-tester.h",
4415 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004416 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004417)
4418
4419xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004420 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004421 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004422 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004423 "test/global-average-pooling-operator-tester.h",
4424 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004425 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004426)
4427
4428xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004429 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004430 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004431 "test/global-average-pooling-ncw.cc",
4432 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004433 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004434 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004435)
4436
4437xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004438 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004439 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004440 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004441 "test/hardswish-operator-tester.h",
4442 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004443 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004444)
4445
4446xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004447 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004448 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004449 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004450 "test/leaky-relu-operator-tester.h",
4451 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004452 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004453)
4454
4455xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004456 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004457 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004458 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004459 "test/max-pooling-operator-tester.h",
4460 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004461 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004462)
4463
4464xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08004465 name = "maximum_nd_test",
4466 srcs = [
4467 "test/binary-elementwise-operator-tester.h",
4468 "test/maximum-nd.cc",
4469 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004470 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08004471)
4472
4473xnnpack_unit_test(
4474 name = "minimum_nd_test",
4475 srcs = [
4476 "test/binary-elementwise-operator-tester.h",
4477 "test/minimum-nd.cc",
4478 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004479 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08004480)
4481
4482xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004483 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08004484 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08004485 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004486 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08004487 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004488 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08004489)
4490
4491xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004492 name = "negate_nc_test",
4493 srcs = [
4494 "test/negate-nc.cc",
4495 "test/negate-operator-tester.h",
4496 ],
4497 deps = OPERATOR_TEST_DEPS,
4498)
4499
4500xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004501 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004502 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004503 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004504 "test/prelu-operator-tester.h",
4505 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004506 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004507)
4508
4509xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004510 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08004511 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004512 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08004513 "test/resize-bilinear-operator-tester.h",
4514 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004515 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08004516)
4517
4518xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004519 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004520 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004521 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004522 "test/sigmoid-operator-tester.h",
4523 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004524 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004525)
4526
4527xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004528 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004529 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004530 "test/softmax-nc.cc",
4531 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004532 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004533 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004534)
4535
4536xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004537 name = "square_nc_test",
4538 srcs = [
4539 "test/square-nc.cc",
4540 "test/square-operator-tester.h",
4541 ],
4542 deps = OPERATOR_TEST_DEPS,
4543)
4544
4545xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07004546 name = "squared_difference_nd_test",
4547 srcs = [
4548 "test/binary-elementwise-operator-tester.h",
4549 "test/squared-difference-nd.cc",
4550 ],
4551 deps = OPERATOR_TEST_DEPS,
4552)
4553
4554xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08004555 name = "subtract_nd_test",
4556 srcs = [
4557 "test/binary-elementwise-operator-tester.h",
4558 "test/subtract-nd.cc",
4559 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004560 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08004561)
4562
4563xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004564 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004565 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004566 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004567 "test/unpooling-operator-tester.h",
4568 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004569 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004570)
4571
Chao Mei6ddfc602020-05-13 22:29:36 -07004572############################### Misc unit tests ###############################
4573
4574xnnpack_unit_test(
4575 name = "memory_planner_test",
4576 srcs = [
4577 "test/memory-planner-test.cc",
4578 ],
4579 deps = [
4580 ":XNNPACK",
4581 ":memory_planner",
4582 ],
4583)
4584
Marat Dukhan08c4a432019-10-03 09:29:21 -07004585############################# Build configurations #############################
4586
Marat Dukhanb8642352019-10-30 15:43:02 -07004587# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07004588config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07004589 name = "xnn_enable_assembly_explicit_true",
4590 define_values = {"xnn_enable_assembly": "true"},
4591)
4592
4593# Disables usage of assembly kernels.
4594config_setting(
4595 name = "xnn_enable_assembly_explicit_false",
4596 define_values = {"xnn_enable_assembly": "false"},
4597)
4598
Marat Dukhan05702cf2020-03-26 15:41:33 -07004599# Disables usage of HMP-aware optimizations.
4600config_setting(
4601 name = "xnn_enable_hmp_explicit_false",
4602 define_values = {"xnn_enable_hmp": "false"},
4603)
4604
Chao Mei6ddfc602020-05-13 22:29:36 -07004605# Enable usage of optimized memory allocation
4606config_setting(
4607 name = "xnn_enable_memopt_explicit_true",
4608 define_values = {"xnnpack_enable_memopt": "true"},
4609)
4610
4611# Disable usage of optimized memory allocation
4612config_setting(
4613 name = "xnn_enable_memopt_explicit_false",
4614 define_values = {"xnnpack_enable_memopt": "false"},
4615)
4616
Marat Dukhanb8642352019-10-30 15:43:02 -07004617# Builds with -c dbg
4618config_setting(
4619 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004620 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07004621 "compilation_mode": "dbg",
4622 },
4623)
4624
4625# Builds with -c opt
4626config_setting(
4627 name = "optimized_build",
4628 values = {
4629 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004630 },
4631)
4632
4633config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07004634 name = "linux_k8",
4635 values = {"cpu": "k8"},
4636)
4637
4638config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07004639 name = "linux_arm",
4640 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07004641)
4642
4643config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07004644 name = "linux_armhf",
4645 values = {"cpu": "armhf"},
4646)
4647
4648config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07004649 name = "linux_aarch64",
4650 values = {"cpu": "aarch64"},
4651)
4652
4653config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004654 name = "android",
4655 values = {"crosstool_top": "//external:android/crosstool"},
4656)
4657
4658config_setting(
4659 name = "android_armv7",
4660 values = {
4661 "crosstool_top": "//external:android/crosstool",
4662 "cpu": "armeabi-v7a",
4663 },
4664)
4665
4666config_setting(
4667 name = "android_arm64",
4668 values = {
4669 "crosstool_top": "//external:android/crosstool",
4670 "cpu": "arm64-v8a",
4671 },
4672)
4673
4674config_setting(
4675 name = "android_x86",
4676 values = {
4677 "crosstool_top": "//external:android/crosstool",
4678 "cpu": "x86",
4679 },
4680)
4681
4682config_setting(
4683 name = "android_x86_64",
4684 values = {
4685 "crosstool_top": "//external:android/crosstool",
4686 "cpu": "x86_64",
4687 },
4688)
4689
4690config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07004691 name = "windows_x86_64",
4692 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07004693)
4694
4695config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07004696 name = "windows_x86_64_clang",
4697 values = {
4698 "compiler": "clang-cl",
4699 "cpu": "x64_windows",
4700 },
4701)
4702
4703config_setting(
4704 name = "windows_x86_64_mingw",
4705 values = {
4706 "compiler": "mingw-gcc",
4707 "cpu": "x64_windows",
4708 },
4709)
4710
4711config_setting(
4712 name = "windows_x86_64_msys",
4713 values = {
4714 "compiler": "msys-gcc",
4715 "cpu": "x64_windows",
4716 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07004717)
4718
4719config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07004720 name = "macos_x86_64",
4721 values = {
4722 "apple_platform_type": "macos",
4723 "cpu": "darwin",
4724 },
4725)
4726
4727config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004728 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07004729 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07004730)
4731
4732config_setting(
4733 name = "emscripten_wasm",
4734 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07004735 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004736 "cpu": "wasm",
4737 },
4738)
4739
4740config_setting(
4741 name = "emscripten_wasmsimd",
4742 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07004743 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004744 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07004745 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004746 },
4747)
4748
4749config_setting(
4750 name = "emscripten_asmjs",
4751 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07004752 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004753 "cpu": "asmjs",
4754 },
4755)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004756
4757config_setting(
4758 name = "ios_armv7",
4759 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004760 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004761 "cpu": "ios_armv7",
4762 },
4763)
4764
4765config_setting(
4766 name = "ios_arm64",
4767 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004768 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004769 "cpu": "ios_arm64",
4770 },
4771)
4772
4773config_setting(
4774 name = "ios_arm64e",
4775 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004776 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004777 "cpu": "ios_arm64e",
4778 },
4779)
4780
4781config_setting(
4782 name = "ios_x86",
4783 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004784 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004785 "cpu": "ios_i386",
4786 },
4787)
4788
4789config_setting(
4790 name = "ios_x86_64",
4791 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004792 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004793 "cpu": "ios_x86_64",
4794 },
4795)
4796
4797config_setting(
4798 name = "watchos_armv7k",
4799 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004800 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004801 "cpu": "watchos_armv7k",
4802 },
4803)
4804
4805config_setting(
4806 name = "watchos_arm64_32",
4807 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004808 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004809 "cpu": "watchos_arm64_32",
4810 },
4811)
4812
4813config_setting(
4814 name = "watchos_x86",
4815 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004816 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004817 "cpu": "watchos_i386",
4818 },
4819)
4820
4821config_setting(
4822 name = "watchos_x86_64",
4823 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004824 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004825 "cpu": "watchos_x86_64",
4826 },
4827)
4828
4829config_setting(
4830 name = "tvos_arm64",
4831 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004832 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004833 "cpu": "tvos_arm64",
4834 },
4835)
4836
4837config_setting(
4838 name = "tvos_x86_64",
4839 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08004840 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08004841 "cpu": "tvos_x86_64",
4842 },
4843)