blob: a06eb213b2921a71a4fdc7ecb60adf292635ba9f [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",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
62 "src/operators/fully-connected-nc.c",
63 "src/operators/global-average-pooling-ncw.c",
64 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070065 "src/operators/leaky-relu-nc.c",
66 "src/operators/max-pooling-nhwc.c",
67 "src/operators/prelu-nc.c",
68 "src/operators/resize-bilinear-nhwc.c",
69 "src/operators/sigmoid-nc.c",
70 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070071 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070072 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070073]
74
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070075SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070076 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077 "src/subgraph/add2.c",
78 "src/subgraph/argmax-pooling-2d.c",
79 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070080 "src/subgraph/bankers-rounding.c",
81 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070082 "src/subgraph/clamp.c",
83 "src/subgraph/convolution-2d.c",
84 "src/subgraph/deconvolution-2d.c",
85 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070087 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070089 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070090 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070091 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070092 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/maximum2.c",
94 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070096 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/prelu.c",
98 "src/subgraph/sigmoid.c",
99 "src/subgraph/softmax.c",
100 "src/subgraph/static-constant-pad.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700101 "src/subgraph/square.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700102 "src/subgraph/squared-difference.c",
103 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700104 "src/subgraph/unpooling-2d.c",
105]
106
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800107TABLE_SRCS = [
108 "src/tables/exp2-k-over-64.c",
109 "src/tables/exp2-k-over-2048.c",
110]
111
Marat Dukhan08c4a432019-10-03 09:29:21 -0700112SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800113 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800114 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800115 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700116 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
117 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700118 "src/f32-clamp/gen/scalar-x1.c",
119 "src/f32-clamp/gen/scalar-x2.c",
120 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -0800121 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800122 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700123 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
124 "src/f32-dwconv-chw/3x3p1-scalar.c",
125 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
126 "src/f32-dwconv-chw/5x5p2-scalar.c",
127 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700128 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
129 "src/f32-dwconv/gen/up1x4-scalar.c",
130 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
131 "src/f32-dwconv/gen/up1x9-scalar.c",
132 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
133 "src/f32-dwconv/gen/up1x25-scalar.c",
134 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
135 "src/f32-dwconv/gen/up2x4-scalar.c",
136 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
137 "src/f32-dwconv/gen/up2x9-scalar.c",
138 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
139 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700140 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
141 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
142 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
143 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700144 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
145 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700146 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
147 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
148 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
149 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700150 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
151 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700152 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700153 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
154 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700155 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
156 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
157 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700158 "src/f32-gemm/gen/1x4-scalar.c",
159 "src/f32-gemm/gen/2x4-scalar.c",
160 "src/f32-gemm/gen/4x2-scalar.c",
161 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700162 "src/f32-gemm/gen/1x4-relu-scalar.c",
163 "src/f32-gemm/gen/2x4-relu-scalar.c",
164 "src/f32-gemm/gen/4x2-relu-scalar.c",
165 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700166 "src/f32-gemm/gen/1x4-minmax-scalar.c",
167 "src/f32-gemm/gen/2x4-minmax-scalar.c",
168 "src/f32-gemm/gen/4x2-minmax-scalar.c",
169 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800170 "src/f32-hswish/gen/scalar-x1.c",
171 "src/f32-hswish/gen/scalar-x2.c",
172 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700173 "src/f32-ibilinear/gen/scalar-c1.c",
174 "src/f32-ibilinear/gen/scalar-c2.c",
175 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700176 "src/f32-igemm/gen/1x4-scalar.c",
177 "src/f32-igemm/gen/2x4-scalar.c",
178 "src/f32-igemm/gen/4x2-scalar.c",
179 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700180 "src/f32-igemm/gen/1x4-relu-scalar.c",
181 "src/f32-igemm/gen/2x4-relu-scalar.c",
182 "src/f32-igemm/gen/4x2-relu-scalar.c",
183 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700184 "src/f32-igemm/gen/1x4-minmax-scalar.c",
185 "src/f32-igemm/gen/2x4-minmax-scalar.c",
186 "src/f32-igemm/gen/4x2-minmax-scalar.c",
187 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700188 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
189 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
190 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700191 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
192 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
193 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
194 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800195 "src/f32-prelu/gen/scalar-2x1.c",
196 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800197 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
198 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
199 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
200 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
201 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
202 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
203 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700209 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800210 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
211 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
212 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
213 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
214 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
215 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
216 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
217 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
218 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700219 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
220 "src/f32-spmm/gen/1x1-minmax-scalar.c",
221 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
222 "src/f32-spmm/gen/2x1-minmax-scalar.c",
223 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
224 "src/f32-spmm/gen/4x1-minmax-scalar.c",
225 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
226 "src/f32-spmm/gen/8x1-minmax-scalar.c",
227 "src/f32-spmm/gen/8x2-minmax-scalar.c",
228 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700229 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
230 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
231 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
232 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
233 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
234 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
235 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
236 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
237 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
238 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
239 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
240 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800241 "src/f32-vbinary/gen/vmax-scalar-x1.c",
242 "src/f32-vbinary/gen/vmax-scalar-x2.c",
243 "src/f32-vbinary/gen/vmax-scalar-x4.c",
244 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
245 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
246 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
247 "src/f32-vbinary/gen/vmin-scalar-x1.c",
248 "src/f32-vbinary/gen/vmin-scalar-x2.c",
249 "src/f32-vbinary/gen/vmin-scalar-x4.c",
250 "src/f32-vbinary/gen/vminc-scalar-x1.c",
251 "src/f32-vbinary/gen/vminc-scalar-x2.c",
252 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700253 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
254 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
255 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
256 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
257 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
258 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
259 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
260 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
261 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
262 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
263 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
264 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700265 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
266 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
267 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
268 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
269 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
270 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700271 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
272 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
273 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
274 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
275 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
276 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700277 "src/f32-vlrelu/gen/scalar-x1.c",
278 "src/f32-vlrelu/gen/scalar-x2.c",
279 "src/f32-vlrelu/gen/scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700280 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
281 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
282 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700283 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
284 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
285 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
286 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
287 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
288 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
289 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
290 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
291 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
292 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
293 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
294 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700295 "src/f32-vunary/gen/vabs-scalar-x1.c",
296 "src/f32-vunary/gen/vabs-scalar-x2.c",
297 "src/f32-vunary/gen/vabs-scalar-x4.c",
298 "src/f32-vunary/gen/vneg-scalar-x1.c",
299 "src/f32-vunary/gen/vneg-scalar-x2.c",
300 "src/f32-vunary/gen/vneg-scalar-x4.c",
301 "src/f32-vunary/gen/vsqr-scalar-x1.c",
302 "src/f32-vunary/gen/vsqr-scalar-x2.c",
303 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800304 "src/math/expminus-scalar-lut2048-p1.c",
305 "src/math/expminus-scalar-lut64-p2.c",
306 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700307 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700308 "src/math/roundne-scalar-nearbyint.c",
309 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700310 "src/math/roundd-scalar-addsub.c",
311 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700312 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700313 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700314 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700315 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700316 "src/math/roundz-scalar-addsub.c",
317 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700318 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800319 "src/math/sigmoid-scalar-lut2048-p1-div.c",
320 "src/math/sigmoid-scalar-lut64-p2-div.c",
321 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700322 "src/q8-avgpool/9p8x-minmax-scalar-c1.c",
323 "src/q8-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700324 "src/q8-dwconv/up1x9-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700325 "src/q8-gavgpool/7p7x-minmax-scalar-c1.c",
326 "src/q8-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700327 "src/q8-gemm/2x2-minmax-scalar.c",
328 "src/q8-igemm/2x2-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700329 "src/q8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700330 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700331 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700332 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700333 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700334 "src/x32-fill/scalar-float.c",
335 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700336 "src/x32-packx/x2-scalar.c",
337 "src/x32-packx/x3-scalar.c",
338 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700339 "src/x32-pad/scalar-int.c",
340 "src/x32-pad/scalar-float.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700341 "src/x32-unpool/scalar.c",
342 "src/x32-zip/x2-scalar.c",
343 "src/x32-zip/x3-scalar.c",
344 "src/x32-zip/x4-scalar.c",
345 "src/x32-zip/xm-scalar.c",
346 "src/x8-lut/scalar.c",
347 "src/x8-zip/x2-scalar.c",
348 "src/x8-zip/x3-scalar.c",
349 "src/x8-zip/x4-scalar.c",
350 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700351 "src/requantization/precise-scalar.c",
352 "src/requantization/fp32-scalar.c",
353 "src/requantization/q31-scalar.c",
354 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700355]
356
Marat Dukhan436ebe62019-12-04 15:10:12 -0800357WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700358 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
359 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700360 "src/f32-clamp/gen/wasm-x1.c",
361 "src/f32-clamp/gen/wasm-x2.c",
362 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700363 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
364 "src/f32-dwconv/gen/up1x4-wasm.c",
365 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
366 "src/f32-dwconv/gen/up1x9-wasm.c",
367 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
368 "src/f32-dwconv/gen/up1x25-wasm.c",
369 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
370 "src/f32-dwconv/gen/up2x4-wasm.c",
371 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
372 "src/f32-dwconv/gen/up2x9-wasm.c",
373 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
374 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700375 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
376 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
377 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
378 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700379 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
380 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700381 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
382 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
383 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
384 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700385 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
386 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700387 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
388 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700389 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
390 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
391 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700392 "src/f32-gemm/gen/1x4-wasm.c",
393 "src/f32-gemm/gen/2x4-wasm.c",
394 "src/f32-gemm/gen/4x2-wasm.c",
395 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700396 "src/f32-gemm/gen/1x4-relu-wasm.c",
397 "src/f32-gemm/gen/2x4-relu-wasm.c",
398 "src/f32-gemm/gen/4x2-relu-wasm.c",
399 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700400 "src/f32-gemm/gen/1x4-minmax-wasm.c",
401 "src/f32-gemm/gen/2x4-minmax-wasm.c",
402 "src/f32-gemm/gen/4x2-minmax-wasm.c",
403 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800404 "src/f32-hswish/gen/wasm-x1.c",
405 "src/f32-hswish/gen/wasm-x2.c",
406 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700407 "src/f32-igemm/gen/1x4-wasm.c",
408 "src/f32-igemm/gen/2x4-wasm.c",
409 "src/f32-igemm/gen/4x2-wasm.c",
410 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700411 "src/f32-igemm/gen/1x4-relu-wasm.c",
412 "src/f32-igemm/gen/2x4-relu-wasm.c",
413 "src/f32-igemm/gen/4x2-relu-wasm.c",
414 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700415 "src/f32-igemm/gen/1x4-minmax-wasm.c",
416 "src/f32-igemm/gen/2x4-minmax-wasm.c",
417 "src/f32-igemm/gen/4x2-minmax-wasm.c",
418 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700419 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
420 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
421 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700422 "src/f32-prelu/gen/wasm-2x1.c",
423 "src/f32-prelu/gen/wasm-2x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700424 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
425 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
426 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
427 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
428 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
429 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
430 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
431 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
432 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
433 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
434 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
435 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800436 "src/f32-vbinary/gen/vmax-wasm-x1.c",
437 "src/f32-vbinary/gen/vmax-wasm-x2.c",
438 "src/f32-vbinary/gen/vmax-wasm-x4.c",
439 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
440 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
441 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
442 "src/f32-vbinary/gen/vmin-wasm-x1.c",
443 "src/f32-vbinary/gen/vmin-wasm-x2.c",
444 "src/f32-vbinary/gen/vmin-wasm-x4.c",
445 "src/f32-vbinary/gen/vminc-wasm-x1.c",
446 "src/f32-vbinary/gen/vminc-wasm-x2.c",
447 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700448 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
449 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
450 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
451 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
452 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
453 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
454 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
455 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
456 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
457 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
458 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
459 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
460 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
461 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
462 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
463 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
464 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
465 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700466 "src/f32-vlrelu/gen/wasm-x1.c",
467 "src/f32-vlrelu/gen/wasm-x2.c",
468 "src/f32-vlrelu/gen/wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700469 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
470 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
471 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800472]
473
Marat Dukhan290055c2020-06-09 12:24:29 -0700474WASMSIMD_UKERNELS = [
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700475 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
476 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
477 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
478 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
479 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
480 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700481 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
482 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
483 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
484 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
485 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
486 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
487 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
488 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
489 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
490 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
491 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
492 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
493 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
494 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
495 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
496 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
497 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
498 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
499 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
500 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
501 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
502 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
503 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
504 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700505 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
506 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
507 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
508 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat-arm.c",
509 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat-arm.c",
510 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat-x86.c",
511 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700512 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
513 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
514 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
515 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
516 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
517 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
518 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
519 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
520 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
521 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
522 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
523 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
524 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
525 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
526 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
527 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
528 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
529 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
530 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
531 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
532 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
533 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
534 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
535 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
536 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
537 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
538 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
539 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
540 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
541 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
542 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
543 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
544 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
545 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
546 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
547 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
548 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
549 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
550 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
551 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
552 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
553 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
554 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
555 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
556 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
557 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
558 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
559 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
560 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
561 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
562 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
563 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
564 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
565 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
566 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
567 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
568 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
569 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
570 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
571 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700572 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
573 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
574 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
575 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat-arm.c",
576 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat-arm.c",
577 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat-x86.c",
578 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700579 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
580 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
581 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
582 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
583 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
584 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
585 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
586 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
587 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
588 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
589 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
590 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
591 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
592 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
593 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
594 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
595 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
596 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
597 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
598 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
599 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
600 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
601 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
602 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
603 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
604 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
605 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
606 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
607 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
608 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhan9baec802020-06-25 21:34:35 -0700609 "src/f32-hswish/gen/wasmsimd-arm-x4.c",
610 "src/f32-hswish/gen/wasmsimd-arm-x8.c",
Marat Dukhanc303fe62020-06-26 10:09:25 -0700611 "src/f32-hswish/gen/wasmsimd-arm-x16.c",
Marat Dukhan9baec802020-06-25 21:34:35 -0700612 "src/f32-hswish/gen/wasmsimd-x86-x4.c",
613 "src/f32-hswish/gen/wasmsimd-x86-x8.c",
Marat Dukhanc303fe62020-06-26 10:09:25 -0700614 "src/f32-hswish/gen/wasmsimd-x86-x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700615 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
616 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
617 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
618 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700619 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
620 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
621 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
622 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700623 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
624 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
625 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
626 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700627 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
628 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
629 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
630 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700631 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
632 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
633 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
634 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700635 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
636 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
637 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
638 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700639 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
640 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
641 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
642 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700643 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
644 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
645 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
646 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700647 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
648 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
649 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
650 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700651 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
652 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
653 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
654 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700655 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
656 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
657 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
658 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
659 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
660 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
661 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
662 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700663 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
664 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
665 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
666 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700667 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
668 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700669 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
670 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700671 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
672 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
673 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
674 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700675 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
676 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
677 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
678 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Marat Dukhan72b399a2020-06-28 22:41:38 -0700679 "src/f32-vrnd/gen/vrndne-wasmsimd-x4.c",
680 "src/f32-vrnd/gen/vrndne-wasmsimd-x8.c",
681 "src/f32-vrnd/gen/vrndz-wasmsimd-x4.c",
682 "src/f32-vrnd/gen/vrndz-wasmsimd-x8.c",
683 "src/f32-vrnd/gen/vrndu-wasmsimd-x4.c",
684 "src/f32-vrnd/gen/vrndu-wasmsimd-x8.c",
685 "src/f32-vrnd/gen/vrndd-wasmsimd-x4.c",
686 "src/f32-vrnd/gen/vrndd-wasmsimd-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700687 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
688 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
689 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
690 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
691 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
692 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700693 "src/math/roundne-wasmsimd-addsub.c",
694 "src/math/roundd-wasmsimd-addsub.c",
695 "src/math/roundu-wasmsimd-addsub.c",
696 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700697 "src/requantization/fp32-wasmsimd.c",
698]
699
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800700PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800701 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800702 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800703 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700704 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
705 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700706 "src/f32-clamp/gen/psimd-x4.c",
707 "src/f32-clamp/gen/psimd-x8.c",
Erich Elsen0a1970e2020-06-10 09:24:59 -0700708 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700709 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700710 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700711 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700712 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700713 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
714 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
715 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
716 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
717 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
718 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
719 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
720 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
721 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
722 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
723 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
724 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Erich Elsen6f278b52020-06-10 16:13:11 -0700725 "src/f32-gavgpool-cw/psimd-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700726 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
727 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700728 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
729 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
730 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
731 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
732 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
733 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
734 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
735 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
736 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
737 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
738 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
739 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
740 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
741 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
742 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
743 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
744 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
745 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
746 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800747 "src/f32-hswish/gen/psimd-x4.c",
748 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700749 "src/f32-ibilinear/gen/psimd-c4.c",
750 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700751 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
752 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
753 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
754 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
755 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
756 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
757 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
758 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
759 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
760 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700761 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700762 "src/f32-spmm/gen/4x1-minmax-psimd.c",
763 "src/f32-spmm/gen/8x1-minmax-psimd.c",
764 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700765 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
766 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700767 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800768 "src/f32-prelu/gen/psimd-2x4.c",
769 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800770 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700771 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
772 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
773 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
774 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
775 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
776 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
777 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
778 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800779 "src/f32-vbinary/gen/vmax-psimd-x4.c",
780 "src/f32-vbinary/gen/vmax-psimd-x8.c",
781 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
782 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
783 "src/f32-vbinary/gen/vmin-psimd-x4.c",
784 "src/f32-vbinary/gen/vmin-psimd-x8.c",
785 "src/f32-vbinary/gen/vminc-psimd-x4.c",
786 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700787 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
788 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
789 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
790 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
791 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
792 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
793 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
794 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700795 "src/f32-vbinary/gen/vsqrdiff-psimd-x4.c",
796 "src/f32-vbinary/gen/vsqrdiff-psimd-x8.c",
797 "src/f32-vbinary/gen/vsqrdiffc-psimd-x4.c",
798 "src/f32-vbinary/gen/vsqrdiffc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700799 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
800 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
801 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
802 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700803 "src/f32-vlrelu/gen/psimd-x4.c",
804 "src/f32-vlrelu/gen/psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700805 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
806 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700807 "src/f32-vunary/gen/vabs-psimd-x4.c",
808 "src/f32-vunary/gen/vabs-psimd-x8.c",
809 "src/f32-vunary/gen/vneg-psimd-x4.c",
810 "src/f32-vunary/gen/vneg-psimd-x8.c",
811 "src/f32-vunary/gen/vsqr-psimd-x4.c",
812 "src/f32-vunary/gen/vsqr-psimd-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700813 "src/x32-fill/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700814 "src/x32-packx/x4-psimd.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700815 "src/x32-pad/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700816 "src/x32-unpool/psimd.c",
817 "src/x32-zip/x2-psimd.c",
818 "src/x32-zip/x3-psimd.c",
819 "src/x32-zip/x4-psimd.c",
820 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700821 "src/requantization/precise-psimd.c",
822 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700823]
824
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800825PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800826 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
827 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
828 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
829 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
830 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
831 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
832 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
833 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
834 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
835 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
836 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
837 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800838 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
839 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
840 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
841 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
842 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
843 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700844 "src/f32-vrnd/gen/vrndne-psimd-x4.c",
845 "src/f32-vrnd/gen/vrndne-psimd-x8.c",
846 "src/f32-vrnd/gen/vrndz-psimd-x4.c",
847 "src/f32-vrnd/gen/vrndz-psimd-x8.c",
848 "src/f32-vrnd/gen/vrndu-psimd-x4.c",
849 "src/f32-vrnd/gen/vrndu-psimd-x8.c",
850 "src/f32-vrnd/gen/vrndd-psimd-x4.c",
851 "src/f32-vrnd/gen/vrndd-psimd-x8.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700852 "src/math/roundne-psimd-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700853 "src/math/roundd-psimd-addsub.c",
854 "src/math/roundu-psimd-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700855 "src/math/roundz-psimd-addsub.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800856 "src/math/sigmoid-psimd-p5-div.c",
857]
858
Marat Dukhan08c4a432019-10-03 09:29:21 -0700859# ISA-specific micro-kernels
860NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700861 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
862 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700863 "src/f32-clamp/gen/neon-x4.c",
864 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700865 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
866 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700867 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
868 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700869 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
870 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700871 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
872 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700873 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
874 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
875 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
876 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700877 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
878 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
879 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
880 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700881 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
882 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
883 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
884 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700885 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700886 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
887 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700888 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
889 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
890 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
891 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
892 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
893 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
894 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
895 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
896 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
897 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
898 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
899 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
900 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
901 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
902 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
903 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
904 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
905 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
906 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
907 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
908 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
909 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
910 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
911 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
912 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
913 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
914 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
915 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
916 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
917 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
918 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800919 "src/f32-hswish/gen/neon-x4.c",
920 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700921 "src/f32-ibilinear/gen/neon-c4.c",
922 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700923 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
924 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
925 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
926 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
927 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
928 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
929 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
930 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
931 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
932 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
933 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
934 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
935 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
936 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
937 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
938 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700939 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
940 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
941 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700942 "src/f32-ppmm/gen/4x8-minmax-neon.c",
943 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800944 "src/f32-prelu/gen/neon-2x4.c",
945 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800946 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
947 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
948 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
949 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
950 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
951 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
952 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
953 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
954 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
955 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
956 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
957 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
958 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
959 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
960 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
961 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
962 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
963 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
964 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
965 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
966 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
967 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
968 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
969 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700970 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800971 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800972 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
973 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
974 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
975 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
976 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
977 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
978 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
979 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
980 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
981 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
982 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
983 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
984 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
985 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
986 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
987 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
988 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
989 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700990 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
991 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
992 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
993 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800994 "src/f32-vbinary/gen/vmax-neon-x4.c",
995 "src/f32-vbinary/gen/vmax-neon-x8.c",
996 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
997 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
998 "src/f32-vbinary/gen/vmin-neon-x4.c",
999 "src/f32-vbinary/gen/vmin-neon-x8.c",
1000 "src/f32-vbinary/gen/vminc-neon-x4.c",
1001 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001002 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1003 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1004 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1005 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1006 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1007 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001008 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1009 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1010 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1011 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001012 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1013 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1014 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1015 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001016 "src/f32-vlrelu/gen/neon-x4.c",
1017 "src/f32-vlrelu/gen/neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001018 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1019 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001020 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1021 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1022 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1023 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1024 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1025 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1026 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1027 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001028 "src/f32-vunary/gen/vabs-neon-x4.c",
1029 "src/f32-vunary/gen/vabs-neon-x8.c",
1030 "src/f32-vunary/gen/vneg-neon-x4.c",
1031 "src/f32-vunary/gen/vneg-neon-x8.c",
1032 "src/f32-vunary/gen/vsqr-neon-x4.c",
1033 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001034 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
1035 "src/q8-avgpool/9x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001036 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001037 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
1038 "src/q8-gavgpool/7x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001039 "src/q8-gemm/4x8-minmax-neon.c",
1040 "src/q8-gemm/8x8-minmax-neon.c",
1041 "src/q8-igemm/4x8-minmax-neon.c",
1042 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001043 "src/q8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001044 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001045 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001046 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001047 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001048 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001049 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001050 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001051 "src/x32-zip/x2-neon.c",
1052 "src/x32-zip/x3-neon.c",
1053 "src/x32-zip/x4-neon.c",
1054 "src/x32-zip/xm-neon.c",
1055 "src/x8-zip/x2-neon.c",
1056 "src/x8-zip/x3-neon.c",
1057 "src/x8-zip/x4-neon.c",
1058 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001059 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001060 "src/math/roundd-neon-addsub.c",
1061 "src/math/roundd-neon-cvt.c",
1062 "src/math/roundu-neon-addsub.c",
1063 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001064 "src/math/roundz-neon-addsub.c",
1065 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001066 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001067 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1068 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1069 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1070 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1071 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1072 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001073 "src/math/sqrt-neon-nr1rsqrts.c",
1074 "src/math/sqrt-neon-nr2rsqrts.c",
1075 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001076 "src/requantization/precise-neon.c",
1077 "src/requantization/fp32-neon.c",
1078 "src/requantization/q31-neon.c",
1079 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001080]
1081
1082NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001083 "src/f32-ibilinear/gen/neonfma-c4.c",
1084 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001085 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1086 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1087 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1088 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1089 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1090 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1091 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1092 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1093 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001094 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1095 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1096 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1097 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001098 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1099 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1100 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1101 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001102 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1103 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1104 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1105 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001106 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1107 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1108 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1109 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1110 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1111 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1112 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1113 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1114 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1115 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1116 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1117 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1118 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1119 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1120 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1121 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1122 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1123 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001124 "src/f32-hswish/gen/neonfma-x4.c",
1125 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001126 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1127 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001128 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1129 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1130 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1131 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1132 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1133 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1134 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1135 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1136 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1137 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1138 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1139 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1140 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1141 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1142 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1143 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1144 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1145 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1146 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1147 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1148 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1149 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1150 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1151 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001152 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1153 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1154 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1155 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1156 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1157 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1158 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1159 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1160 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1161 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1162 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1163 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1164 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1165 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1166 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1167 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1168 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1169 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1170 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1171 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1172 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1173 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1174 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1175 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1176 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1177 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1178 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1179 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1180 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1181 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1182 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1183 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1184 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1185 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1186 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1187 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1188 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1189 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1190 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1191 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1192 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1193 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1194 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1195 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1196 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1197 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1198 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1199 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1200 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1201 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1202 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1203 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1204 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1205 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001206 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1207 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001208 "src/math/exp-neonfma-lut64-p2.c",
1209 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001210 "src/math/expminus-neonfma-lut2048-p1.c",
1211 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001212 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001213 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1214 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1215 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1216 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1217 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1218 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1219 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1220 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1221 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1222 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1223 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1224 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1225 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1226 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1227 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1228 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1229 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1230 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001231 "src/math/sqrt-neonfma-nr1fma.c",
1232 "src/math/sqrt-neonfma-nr2fma.c",
1233 "src/math/sqrt-neonfma-nr3fma.c",
1234 "src/math/sqrt-neonfma-nr2fma1adj.c",
1235 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001236]
1237
1238AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001239 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1240 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1241 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1242 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1243 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1244 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001245 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1246 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1247 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1248 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1249 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1250 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1251 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1252 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1253 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1254 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1255 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1256 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1257 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1258 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1259 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1260 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1261 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1262 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1263 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1264 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001265 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1266 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001267 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1268 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001269 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1270 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001271 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1272 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001273 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1274 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1275 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1276 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1277 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001278 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1279 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1280 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1281 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1282 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1283 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1284 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1285 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1286 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1287 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1288 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1289 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1290 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1291 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1292 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1293 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1294 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1295 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001296 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1297 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1298 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1299 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1300 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1301 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1302 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1303 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1304 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1305 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1306 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1307 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1308 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1309 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1310 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1311 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1312 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1313 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001314 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1315 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1316 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1317 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1318 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1319 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001320]
1321
Marat Dukhan8853b822020-05-07 12:19:01 -07001322NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001323 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1324 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1325 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1326 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1327 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1328 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1329 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1330 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001331 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001332 "src/math/roundd-neonv8.c",
1333 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001334 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001335]
1336
Marat Dukhan08c4a432019-10-03 09:29:21 -07001337AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001338 "src/f16-clamp/gen/neonfp16arith-x16.c",
1339 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001340 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1341 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1342 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1343 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1344 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1345 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1346 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1347 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1348 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1349 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1350 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1351 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001352 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1353 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001354 "src/f16-hswish/gen/neonfp16arith-x16.c",
1355 "src/f16-hswish/gen/neonfp16arith-x8.c",
1356 "src/f16-prelu/gen/neonfp16arith-2x16.c",
1357 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001358 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1359 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1360 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1361 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1362 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1363 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1364 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1365 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1366 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1367 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1368 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1369 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1370 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1371 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1372 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1373 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1374 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1375 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1376 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1377 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1378 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1379 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1380 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1381 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1382 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1383 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1384 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1385 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001386 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1387 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001388 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1389 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1390 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1391 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1392 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1393 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1394 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1395 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001396 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1397 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1398 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1399 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001400 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1401 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1402 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1403 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1404 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1405 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1406 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1407 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1408 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1409 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1410 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1411 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001412 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1413 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1414 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1415 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1416 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1417 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1418 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1419 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001420]
1421
1422SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001423 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1424 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001425 "src/f32-clamp/gen/sse-x4.c",
1426 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001427 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1428 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001429 "src/f32-dwconv-chw/3x3p1-sse.c",
1430 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001431 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1432 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1433 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1434 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1435 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1436 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1437 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1438 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1439 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1440 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1441 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1442 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001443 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001444 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1445 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001446 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1447 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1448 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1449 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1450 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1451 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1452 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1453 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1454 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1455 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1456 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1457 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1458 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001459 "src/f32-hswish/gen/sse-x4.c",
1460 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001461 "src/f32-ibilinear/gen/sse-c4.c",
1462 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001463 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1464 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1465 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1466 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1467 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1468 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1469 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001470 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1471 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1472 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001473 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001474 "src/f32-prelu/gen/sse-2x4.c",
1475 "src/f32-prelu/gen/sse-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001476 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001477 "src/f32-spmm/gen/4x1-minmax-sse.c",
1478 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001479 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001480 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1481 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1482 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1483 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1484 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1485 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1486 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1487 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001488 "src/f32-vbinary/gen/vmax-sse-x4.c",
1489 "src/f32-vbinary/gen/vmax-sse-x8.c",
1490 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1491 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1492 "src/f32-vbinary/gen/vmin-sse-x4.c",
1493 "src/f32-vbinary/gen/vmin-sse-x8.c",
1494 "src/f32-vbinary/gen/vminc-sse-x4.c",
1495 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001496 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1497 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1498 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1499 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1500 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1501 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1502 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1503 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001504 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1505 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1506 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1507 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001508 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1509 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1510 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1511 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001512 "src/f32-vlrelu/gen/sse-x4.c",
1513 "src/f32-vlrelu/gen/sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001514 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1515 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001516 "src/f32-vunary/gen/vabs-sse-x4.c",
1517 "src/f32-vunary/gen/vabs-sse-x8.c",
1518 "src/f32-vunary/gen/vneg-sse-x4.c",
1519 "src/f32-vunary/gen/vneg-sse-x8.c",
1520 "src/f32-vunary/gen/vsqr-sse-x4.c",
1521 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001522 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001523 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001524 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001525 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001526 "src/math/roundd-sse-addsub.c",
1527 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001528 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001529 "src/math/sqrt-sse-hh1mac.c",
1530 "src/math/sqrt-sse-nr1mac.c",
1531 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001532]
1533
1534SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001535 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1536 "src/f32-argmaxpool/4x-sse2-c4.c",
1537 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001538 "src/f32-prelu/gen/sse2-2x4.c",
1539 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001540 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1541 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1542 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1543 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1544 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1545 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1546 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1547 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1548 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1549 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1550 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1551 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001552 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001553 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001554 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001555 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001556 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1557 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan0d3f4672020-06-25 16:42:58 -07001558 "src/f32-vlrelu/gen/sse2-x4.c",
1559 "src/f32-vlrelu/gen/sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001560 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1561 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1562 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1563 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1564 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1565 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1566 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1567 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001568 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1569 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001570 "src/q8-igemm/4x4c2-minmax-sse2.c",
1571 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001572 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1573 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001574 "src/q8-gemm/2x4c8-minmax-sse2.c",
1575 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001576 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001577 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001578 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001579 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001580 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001581 "src/x32-zip/x2-sse2.c",
1582 "src/x32-zip/x3-sse2.c",
1583 "src/x32-zip/x4-sse2.c",
1584 "src/x32-zip/xm-sse2.c",
1585 "src/x8-zip/x2-sse2.c",
1586 "src/x8-zip/x3-sse2.c",
1587 "src/x8-zip/x4-sse2.c",
1588 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001589 "src/math/exp-sse2-p5.c",
1590 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001591 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001592 "src/math/roundd-sse2-cvt.c",
1593 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001594 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001595 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001596 "src/requantization/precise-sse2.c",
1597 "src/requantization/fp32-sse2.c",
1598 "src/requantization/q31-sse2.c",
1599 "src/requantization/gemmlowp-sse2.c",
1600]
1601
1602SSSE3_UKERNELS = [
1603 "src/requantization/precise-ssse3.c",
1604 "src/requantization/q31-ssse3.c",
1605 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001606]
1607
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001608SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001609 "src/f32-prelu/gen/sse41-2x4.c",
1610 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001611 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001612 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001613 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001614 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001615 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1616 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan0d3f4672020-06-25 16:42:58 -07001617 "src/f32-vlrelu/gen/sse41-x4.c",
1618 "src/f32-vlrelu/gen/sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001619 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1620 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1621 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1622 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1623 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1624 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1625 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1626 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001627 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001628 "src/math/roundd-sse41.c",
1629 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001630 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001631 "src/requantization/precise-sse4.c",
1632 "src/requantization/q31-sse4.c",
1633 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001634]
1635
Marat Dukhan08c4a432019-10-03 09:29:21 -07001636AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001637 "src/f32-clamp/gen/avx-x8.c",
1638 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001639 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1640 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1641 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1642 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1643 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1644 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1645 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1646 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1647 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1648 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1649 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1650 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1651 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1652 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1653 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1654 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1655 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1656 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1657 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1658 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1659 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1660 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1661 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1662 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1663 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1664 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1665 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1666 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1667 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1668 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001669 "src/f32-hswish/gen/avx-x8.c",
1670 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001671 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1672 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1673 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1674 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1675 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1676 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1677 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1678 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1679 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001680 "src/f32-prelu/gen/avx-2x8.c",
1681 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001682 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001683 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1684 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1685 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1686 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1687 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1688 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1689 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1690 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001691 "src/f32-vbinary/gen/vmax-avx-x8.c",
1692 "src/f32-vbinary/gen/vmax-avx-x16.c",
1693 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1694 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1695 "src/f32-vbinary/gen/vmin-avx-x8.c",
1696 "src/f32-vbinary/gen/vmin-avx-x16.c",
1697 "src/f32-vbinary/gen/vminc-avx-x8.c",
1698 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001699 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1700 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1701 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1702 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1703 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1704 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1705 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1706 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001707 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1708 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1709 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1710 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001711 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1712 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1713 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1714 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001715 "src/f32-vlrelu/gen/avx-x8.c",
1716 "src/f32-vlrelu/gen/avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001717 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001718 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1719 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1720 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1721 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1722 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1723 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1724 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1725 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001726 "src/f32-vunary/gen/vabs-avx-x8.c",
1727 "src/f32-vunary/gen/vabs-avx-x16.c",
1728 "src/f32-vunary/gen/vneg-avx-x8.c",
1729 "src/f32-vunary/gen/vneg-avx-x16.c",
1730 "src/f32-vunary/gen/vsqr-avx-x8.c",
1731 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001732]
1733
Marat Dukhanfda12b82019-11-21 12:27:59 -08001734FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001735 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1736 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1737 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1738 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1739 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1740 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1741 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1742 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1743 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1744 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1745 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1746 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1747 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1748 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1749 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1750 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1751 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1752 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1753 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1754 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1755 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1756 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1757 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1758 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1759 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1760 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1761 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1762 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1763 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1764 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1765 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1766 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1767 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1768 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1769 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1770 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1771 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1772 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1773 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1774 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001775 "src/f32-hswish/gen/fma3-x8.c",
1776 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001777 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1778 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1779 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1780 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1781 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1782 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1783 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1784 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1785 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1786 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1787 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1788 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1789 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1790 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001791 "src/math/sqrt-fma3-nr1fma.c",
1792 "src/math/sqrt-fma3-nr2fma.c",
1793 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001794]
1795
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001796AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001797 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1798 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1799 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1800 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1801 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1802 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1803 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1804 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1805 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1806 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1807 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1808 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1809 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1810 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1811 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1812 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1813 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1814 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1815 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1816 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1817 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1818 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1819 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1820 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1821 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1822 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1823 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1824 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1825 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1826 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1827 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1828 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1829 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1830 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1831 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1832 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001833 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1834 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1835 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1836 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1837 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1838 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1839 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1840 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1841 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1842 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1843 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1844 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1845 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1846 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1847 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1848 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1849 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1850 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1851 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1852 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1853 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1854 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1855 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1856 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1857 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1858 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1859 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1860 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1861 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1862 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001863 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1864 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1865 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1866 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1867 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1868 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1869 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1870 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1871 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1872 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1873 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1874 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1875 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1876 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1877 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1878 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1879 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1880 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1881 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1882 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1883 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1884 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1885 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1886 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001887 "src/math/exp-avx2-p5.c",
1888 "src/math/exp-avx2-perm-p3.c",
1889 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001890 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001891 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001892 "src/math/sigmoid-avx2-rr2-p5-div.c",
1893 "src/math/sigmoid-avx2-rr1-p5-div.c",
1894 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1895 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1896 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1897 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001898]
1899
Marat Dukhan08c4a432019-10-03 09:29:21 -07001900AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001901 "src/f32-clamp/gen/avx512f-x16.c",
1902 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001903 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1904 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1905 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1906 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1907 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1908 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1909 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1910 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1911 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1912 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1913 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1914 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1915 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1916 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1917 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1918 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1919 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1920 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1921 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1922 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1923 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1924 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1925 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1926 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001927 "src/f32-hswish/gen/avx512f-x16.c",
1928 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001929 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1930 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1931 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1932 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1933 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1934 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001935 "src/f32-prelu/gen/avx512f-2x16.c",
1936 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001937 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1938 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1939 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1940 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1941 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1942 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1943 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1944 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1945 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1946 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1947 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1948 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1949 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1950 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1951 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1952 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1953 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1954 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1955 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1956 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1957 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1958 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1959 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1960 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1961 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1962 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1963 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1964 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1965 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1966 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1967 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1968 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1969 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1970 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1971 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1972 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001973 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001974 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
1975 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
1976 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
1977 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
1978 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
1979 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
1980 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
1981 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001982 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1983 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1984 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1985 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1986 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1987 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1988 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1989 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001990 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
1991 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
1992 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
1993 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
1994 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
1995 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
1996 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
1997 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001998 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
1999 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2000 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2001 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002002 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2003 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2004 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2005 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07002006 "src/f32-vlrelu/gen/avx512f-x16.c",
2007 "src/f32-vlrelu/gen/avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002008 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002009 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2010 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2011 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2012 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2013 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2014 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2015 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2016 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2017 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2018 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2019 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2020 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2021 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2022 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2023 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2024 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2025 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2026 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2027 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2028 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2029 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2030 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2031 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2032 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002033 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2034 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2035 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2036 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2037 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2038 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2039 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2040 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002041 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2042 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2043 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2044 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2045 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2046 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002047 "src/math/exp-avx512f-p5-scalef.c",
2048 "src/math/exp-avx512f-p5.c",
2049 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002050 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002051 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002052 "src/math/sqrt-avx512f-nr1fma.c",
2053 "src/math/sqrt-avx512f-nr2fma.c",
2054 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002055]
2056
2057AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002058 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002059 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002060 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2061 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
2062 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2063 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002064 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2065 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S",
2066 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2067 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002068 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2069 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
2070 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2071 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002072]
2073
2074AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002075 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002076 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2077 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002078 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002079 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2080 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002081 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2082 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2083 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002084 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002085 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2086 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2087 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002088 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002089 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2090 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002091 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002092 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2093 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2094 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2095 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2096 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2097 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2098 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2099 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2100 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2101 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2102 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2103 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2104 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2105 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2106 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2107 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2108 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2109 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002110 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2111 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002112 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002113 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2114 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2115 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2116 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2117 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2118 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2119 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2120 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2121 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2122 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2123 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2124 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2125 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2126 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2127 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2128 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2129 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2130 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002131 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2132 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2133 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2134 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2135 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2136 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2137 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2138 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2139 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2140 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2141 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2142 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2143 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2144 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2145 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2146 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2147 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2148 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002149]
2150
Marat Dukhan1b354632020-03-23 12:50:22 -07002151INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002152 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002153 "src/xnnpack/argmaxpool.h",
2154 "src/xnnpack/avgpool.h",
2155 "src/xnnpack/clamp.h",
2156 "src/xnnpack/common.h",
2157 "src/xnnpack/conv.h",
2158 "src/xnnpack/dwconv.h",
2159 "src/xnnpack/gavgpool.h",
2160 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002161 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002162 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002163 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002164 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002165 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002166 "src/xnnpack/lut.h",
2167 "src/xnnpack/math.h",
2168 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002169 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002170 "src/xnnpack/packx.h",
2171 "src/xnnpack/pad.h",
2172 "src/xnnpack/params.h",
2173 "src/xnnpack/pavgpool.h",
2174 "src/xnnpack/ppmm.h",
2175 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002176 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002177 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002178 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002179 "src/xnnpack/rmax.h",
2180 "src/xnnpack/scalar-utils.h",
2181 "src/xnnpack/spmm.h",
2182 "src/xnnpack/unpool.h",
2183 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002184 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002185 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002186 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002187 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002188 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002189 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002190 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002191]
2192
2193INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002194 "include/xnnpack.h",
2195 "src/xnnpack/allocator.h",
2196 "src/xnnpack/compute.h",
2197 "src/xnnpack/im2col.h",
2198 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002199 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002200 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002201 "src/xnnpack/operator.h",
2202 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002203 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002204 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002205 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002206 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002207]
2208
Marat Dukhan1b354632020-03-23 12:50:22 -07002209ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002210 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002211]
2212
Marat Dukhan1b354632020-03-23 12:50:22 -07002213MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002214 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002215 "include/xnnpack.h",
2216]
2217
Marat Dukhan1b354632020-03-23 12:50:22 -07002218MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002219 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002220 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002221 "src/xnnpack/requantization.h",
2222 "include/xnnpack.h",
2223]
2224
2225OPERATOR_TEST_PARAMS_HDRS = [
2226 "src/xnnpack/params.h",
2227 "src/xnnpack/common.h",
2228]
2229
2230WEIGHTS_PACK_HDRS = [
2231 "src/xnnpack/pack.h",
2232 "src/xnnpack/operator.h",
2233 "src/xnnpack/compute.h",
2234]
2235
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002236LOGGING_COPTS = select({
2237 # No logging in optimized mode
2238 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2239 # Full logging in debug mode
2240 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2241 # Error-only logging in default (fastbuild) mode
2242 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2243})
2244
Marat Dukhan3b59de22020-06-03 20:15:19 -07002245LOGGING_SRCS = select({
2246 # No logging in optimized mode
2247 ":optimized_build": [],
2248 "//conditions:default": [
2249 "src/operator-strings.c",
2250 "src/subgraph-strings.c",
2251 ],
2252})
2253
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002254LOGGING_HDRS = [
2255 "src/xnnpack/log.h",
2256]
2257
Marat Dukhan08c4a432019-10-03 09:29:21 -07002258xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002259 name = "tables",
2260 srcs = TABLE_SRCS,
2261 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002262 gcc_copts = xnnpack_gcc_std_copts(),
2263 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002264)
2265
2266xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002267 name = "scalar_ukernels",
2268 srcs = SCALAR_UKERNELS,
2269 hdrs = INTERNAL_HDRS,
2270 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002271 gcc_copts = xnnpack_gcc_std_copts(),
2272 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002273 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002274 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002275 "@FP16",
2276 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002277 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002278 ],
2279)
2280
2281xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002282 name = "scalar_ukernels_test_mode",
2283 srcs = SCALAR_UKERNELS,
2284 hdrs = INTERNAL_HDRS,
2285 aarch32_copts = ["-marm"],
2286 copts = [
2287 "-UNDEBUG",
2288 "-DXNN_TEST_MODE=1",
2289 ],
2290 gcc_copts = xnnpack_gcc_std_copts(),
2291 msvc_copts = xnnpack_msvc_std_copts(),
2292 deps = [
2293 ":tables",
2294 "@FP16",
2295 "@FXdiv",
2296 "@pthreadpool",
2297 ],
2298)
2299
2300xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002301 name = "wasm_ukernels",
2302 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002303 gcc_copts = xnnpack_gcc_std_copts(),
2304 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002305 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002306 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002307 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002308 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002309 "@FP16",
2310 "@FXdiv",
2311 "@pthreadpool",
2312 ],
2313)
2314
2315xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002316 name = "wasm_ukernels_test_mode",
2317 hdrs = INTERNAL_HDRS,
2318 copts = [
2319 "-UNDEBUG",
2320 "-DXNN_TEST_MODE=1",
2321 ],
2322 gcc_copts = xnnpack_gcc_std_copts(),
2323 msvc_copts = xnnpack_msvc_std_copts(),
2324 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002325 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002326 deps = [
2327 ":tables",
2328 "@FP16",
2329 "@FXdiv",
2330 "@pthreadpool",
2331 ],
2332)
2333
2334xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002335 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002336 hdrs = INTERNAL_HDRS,
2337 aarch32_copts = [
2338 "-marm",
2339 "-mfpu=neon",
2340 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002341 gcc_copts = xnnpack_gcc_std_copts(),
2342 gcc_x86_copts = ["-msse2"],
2343 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002344 optimized_copts = [
2345 "-O3",
2346 "-ffast-math",
2347 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002348 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002349 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002350 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002351 "@FP16",
2352 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002353 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002354 ],
2355)
2356
2357xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002358 name = "psimd_fastmath_ukernels_test_mode",
2359 hdrs = INTERNAL_HDRS,
2360 aarch32_copts = [
2361 "-marm",
2362 "-mfpu=neon",
2363 ],
2364 copts = [
2365 "-UNDEBUG",
2366 "-DXNN_TEST_MODE=1",
2367 ],
2368 gcc_copts = xnnpack_gcc_std_copts(),
2369 gcc_x86_copts = ["-msse2"],
2370 msvc_copts = xnnpack_msvc_std_copts(),
2371 optimized_copts = [
2372 "-O3",
2373 "-ffast-math",
2374 ],
2375 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2376 deps = [
2377 ":tables",
2378 "@FP16",
2379 "@psimd",
2380 "@pthreadpool",
2381 ],
2382)
2383
2384xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002385 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002386 hdrs = INTERNAL_HDRS,
2387 aarch32_copts = [
2388 "-marm",
2389 "-mfpu=neon",
2390 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002391 gcc_copts = xnnpack_gcc_std_copts(),
2392 gcc_x86_copts = ["-msse2"],
2393 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002394 optimized_copts = [
2395 "-O3",
2396 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002397 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002398 deps = [
2399 ":tables",
2400 "@FP16",
2401 "@psimd",
2402 "@pthreadpool",
2403 ],
2404)
2405
2406xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002407 name = "psimd_accmath_ukernels_test_mode",
2408 hdrs = INTERNAL_HDRS,
2409 aarch32_copts = [
2410 "-marm",
2411 "-mfpu=neon",
2412 ],
2413 copts = [
2414 "-UNDEBUG",
2415 "-DXNN_TEST_MODE=1",
2416 ],
2417 gcc_copts = xnnpack_gcc_std_copts(),
2418 gcc_x86_copts = ["-msse2"],
2419 msvc_copts = xnnpack_msvc_std_copts(),
2420 optimized_copts = [
2421 "-O3",
2422 ],
2423 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2424 deps = [
2425 ":tables",
2426 "@FP16",
2427 "@psimd",
2428 "@pthreadpool",
2429 ],
2430)
2431
2432xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002433 name = "neon_ukernels",
2434 hdrs = INTERNAL_HDRS,
2435 aarch32_copts = [
2436 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002437 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002438 "-mfpu=neon",
2439 ],
2440 aarch32_srcs = NEON_UKERNELS,
2441 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002442 gcc_copts = xnnpack_gcc_std_copts(),
2443 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002444 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002445 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002446 "@FP16",
2447 "@pthreadpool",
2448 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002449)
2450
2451xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002452 name = "neon_ukernels_test_mode",
2453 hdrs = INTERNAL_HDRS,
2454 aarch32_copts = [
2455 "-marm",
2456 "-march=armv7-a",
2457 "-mfpu=neon",
2458 ],
2459 aarch32_srcs = NEON_UKERNELS,
2460 aarch64_srcs = NEON_UKERNELS,
2461 copts = [
2462 "-UNDEBUG",
2463 "-DXNN_TEST_MODE=1",
2464 ],
2465 gcc_copts = xnnpack_gcc_std_copts(),
2466 msvc_copts = xnnpack_msvc_std_copts(),
2467 deps = [
2468 ":tables",
2469 "@FP16",
2470 "@pthreadpool",
2471 ],
2472)
2473
2474xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002475 name = "neonfma_ukernels",
2476 hdrs = INTERNAL_HDRS,
2477 aarch32_copts = [
2478 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002479 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002480 "-mfpu=neon-vfpv4",
2481 ],
2482 aarch32_srcs = NEONFMA_UKERNELS,
2483 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002484 apple_aarch32_copts = [
2485 "-mcpu=swift",
2486 "-mtune=generic",
2487 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002488 gcc_copts = xnnpack_gcc_std_copts(),
2489 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002490 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002491 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002492 "@FP16",
2493 "@pthreadpool",
2494 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002495)
2496
2497xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002498 name = "neonfma_ukernels_test_mode",
2499 hdrs = INTERNAL_HDRS,
2500 aarch32_copts = [
2501 "-marm",
2502 "-march=armv7-a",
2503 "-mfpu=neon-vfpv4",
2504 ],
2505 aarch32_srcs = NEONFMA_UKERNELS,
2506 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002507 apple_aarch32_copts = [
2508 "-mcpu=swift",
2509 "-mtune=generic",
2510 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002511 copts = [
2512 "-UNDEBUG",
2513 "-DXNN_TEST_MODE=1",
2514 ],
2515 gcc_copts = xnnpack_gcc_std_copts(),
2516 msvc_copts = xnnpack_msvc_std_copts(),
2517 deps = [
2518 ":tables",
2519 "@FP16",
2520 "@pthreadpool",
2521 ],
2522)
2523
2524xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002525 name = "neonv8_ukernels",
2526 hdrs = INTERNAL_HDRS,
2527 aarch32_copts = [
2528 "-marm",
2529 "-march=armv8-a",
2530 "-mfpu=neon-fp-armv8",
2531 ],
2532 aarch32_srcs = NEONV8_UKERNELS,
2533 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002534 apple_aarch32_copts = [
2535 "-mcpu=cyclone",
2536 "-mtune=generic",
2537 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002538 gcc_copts = xnnpack_gcc_std_copts(),
2539 msvc_copts = xnnpack_msvc_std_copts(),
2540 deps = [
2541 ":tables",
2542 "@FP16",
2543 "@pthreadpool",
2544 ],
2545)
2546
2547xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002548 name = "neonv8_ukernels_test_mode",
2549 hdrs = INTERNAL_HDRS,
2550 aarch32_copts = [
2551 "-marm",
2552 "-march=armv8-a",
2553 "-mfpu=neon-fp-armv8",
2554 ],
2555 aarch32_srcs = NEONV8_UKERNELS,
2556 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002557 apple_aarch32_copts = [
2558 "-mcpu=cyclone",
2559 "-mtune=generic",
2560 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002561 copts = [
2562 "-UNDEBUG",
2563 "-DXNN_TEST_MODE=1",
2564 ],
2565 gcc_copts = xnnpack_gcc_std_copts(),
2566 msvc_copts = xnnpack_msvc_std_copts(),
2567 deps = [
2568 ":tables",
2569 "@FP16",
2570 "@pthreadpool",
2571 ],
2572)
2573
2574xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002575 name = "neonfp16arith_ukernels",
2576 hdrs = INTERNAL_HDRS,
2577 aarch64_copts = ["-march=armv8.2-a+fp16"],
2578 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002579 gcc_copts = xnnpack_gcc_std_copts(),
2580 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002581 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002582 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002583 "@FP16",
2584 "@pthreadpool",
2585 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002586)
2587
2588xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002589 name = "neonfp16arith_ukernels_test_mode",
2590 hdrs = INTERNAL_HDRS,
2591 aarch64_copts = ["-march=armv8.2-a+fp16"],
2592 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2593 copts = [
2594 "-UNDEBUG",
2595 "-DXNN_TEST_MODE=1",
2596 ],
2597 gcc_copts = xnnpack_gcc_std_copts(),
2598 msvc_copts = xnnpack_msvc_std_copts(),
2599 deps = [
2600 ":tables",
2601 "@FP16",
2602 "@pthreadpool",
2603 ],
2604)
2605
2606xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002607 name = "sse2_ukernels",
2608 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002609 gcc_copts = xnnpack_gcc_std_copts(),
2610 gcc_x86_copts = ["-msse2"],
2611 msvc_copts = xnnpack_msvc_std_copts(),
2612 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002613 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002614 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002615 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002616 "@FP16",
2617 "@pthreadpool",
2618 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002619)
2620
2621xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002622 name = "sse2_ukernels_test_mode",
2623 hdrs = INTERNAL_HDRS,
2624 copts = [
2625 "-UNDEBUG",
2626 "-DXNN_TEST_MODE=1",
2627 ],
2628 gcc_copts = xnnpack_gcc_std_copts(),
2629 gcc_x86_copts = ["-msse2"],
2630 msvc_copts = xnnpack_msvc_std_copts(),
2631 msvc_x86_32_copts = ["/arch:SSE2"],
2632 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2633 deps = [
2634 ":tables",
2635 "@FP16",
2636 "@pthreadpool",
2637 ],
2638)
2639
2640xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002641 name = "ssse3_ukernels",
2642 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002643 gcc_copts = xnnpack_gcc_std_copts(),
2644 gcc_x86_copts = ["-mssse3"],
2645 msvc_copts = xnnpack_msvc_std_copts(),
2646 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002647 x86_srcs = SSSE3_UKERNELS,
2648 deps = [
2649 ":tables",
2650 "@FP16",
2651 "@pthreadpool",
2652 ],
2653)
2654
2655xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002656 name = "ssse3_ukernels_test_mode",
2657 hdrs = INTERNAL_HDRS,
2658 copts = [
2659 "-UNDEBUG",
2660 "-DXNN_TEST_MODE=1",
2661 ],
2662 gcc_copts = xnnpack_gcc_std_copts(),
2663 gcc_x86_copts = ["-mssse3"],
2664 msvc_copts = xnnpack_msvc_std_copts(),
2665 msvc_x86_32_copts = ["/arch:SSE2"],
2666 x86_srcs = SSSE3_UKERNELS,
2667 deps = [
2668 ":tables",
2669 "@FP16",
2670 "@pthreadpool",
2671 ],
2672)
2673
2674xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002675 name = "sse41_ukernels",
2676 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002677 gcc_copts = xnnpack_gcc_std_copts(),
2678 gcc_x86_copts = ["-msse4.1"],
2679 msvc_copts = xnnpack_msvc_std_copts(),
2680 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002681 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002682 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002683 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002684 "@FP16",
2685 "@pthreadpool",
2686 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002687)
2688
2689xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002690 name = "sse41_ukernels_test_mode",
2691 hdrs = INTERNAL_HDRS,
2692 copts = [
2693 "-UNDEBUG",
2694 "-DXNN_TEST_MODE=1",
2695 ],
2696 gcc_copts = xnnpack_gcc_std_copts(),
2697 gcc_x86_copts = ["-msse4.1"],
2698 msvc_copts = xnnpack_msvc_std_copts(),
2699 msvc_x86_32_copts = ["/arch:SSE2"],
2700 x86_srcs = SSE41_UKERNELS,
2701 deps = [
2702 ":tables",
2703 "@FP16",
2704 "@pthreadpool",
2705 ],
2706)
2707
2708xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002709 name = "avx_ukernels",
2710 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002711 gcc_copts = xnnpack_gcc_std_copts(),
2712 gcc_x86_copts = ["-mavx"],
2713 msvc_copts = xnnpack_msvc_std_copts(),
2714 msvc_x86_32_copts = ["/arch:AVX"],
2715 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002716 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002717 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002718 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002719 "@FP16",
2720 "@pthreadpool",
2721 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002722)
2723
2724xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002725 name = "avx_ukernels_test_mode",
2726 hdrs = INTERNAL_HDRS,
2727 copts = [
2728 "-UNDEBUG",
2729 "-DXNN_TEST_MODE=1",
2730 ],
2731 gcc_copts = xnnpack_gcc_std_copts(),
2732 gcc_x86_copts = ["-mavx"],
2733 msvc_copts = xnnpack_msvc_std_copts(),
2734 msvc_x86_32_copts = ["/arch:AVX"],
2735 msvc_x86_64_copts = ["/arch:AVX"],
2736 x86_srcs = AVX_UKERNELS,
2737 deps = [
2738 ":tables",
2739 "@FP16",
2740 "@pthreadpool",
2741 ],
2742)
2743
2744xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002745 name = "fma3_ukernels",
2746 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002747 gcc_copts = xnnpack_gcc_std_copts(),
2748 gcc_x86_copts = ["-mfma"],
2749 msvc_copts = xnnpack_msvc_std_copts(),
2750 msvc_x86_32_copts = ["/arch:AVX"],
2751 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002752 x86_srcs = FMA3_UKERNELS,
2753 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002754 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002755 "@FP16",
2756 "@pthreadpool",
2757 ],
2758)
2759
2760xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002761 name = "fma3_ukernels_test_mode",
2762 hdrs = INTERNAL_HDRS,
2763 copts = [
2764 "-UNDEBUG",
2765 "-DXNN_TEST_MODE=1",
2766 ],
2767 gcc_copts = xnnpack_gcc_std_copts(),
2768 gcc_x86_copts = ["-mfma"],
2769 msvc_copts = xnnpack_msvc_std_copts(),
2770 msvc_x86_32_copts = ["/arch:AVX"],
2771 msvc_x86_64_copts = ["/arch:AVX"],
2772 x86_srcs = FMA3_UKERNELS,
2773 deps = [
2774 ":tables",
2775 "@FP16",
2776 "@pthreadpool",
2777 ],
2778)
2779
2780xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002781 name = "avx2_ukernels",
2782 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002783 gcc_copts = xnnpack_gcc_std_copts(),
2784 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002785 "-mfma",
2786 "-mavx2",
2787 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002788 msvc_copts = xnnpack_msvc_std_copts(),
2789 msvc_x86_32_copts = ["/arch:AVX2"],
2790 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002791 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002792 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002793 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002794 "@FP16",
2795 "@pthreadpool",
2796 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002797)
2798
2799xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002800 name = "avx2_ukernels_test_mode",
2801 hdrs = INTERNAL_HDRS,
2802 copts = [
2803 "-UNDEBUG",
2804 "-DXNN_TEST_MODE=1",
2805 ],
2806 gcc_copts = xnnpack_gcc_std_copts(),
2807 gcc_x86_copts = [
2808 "-mfma",
2809 "-mavx2",
2810 ],
2811 msvc_copts = xnnpack_msvc_std_copts(),
2812 msvc_x86_32_copts = ["/arch:AVX2"],
2813 msvc_x86_64_copts = ["/arch:AVX2"],
2814 x86_srcs = AVX2_UKERNELS,
2815 deps = [
2816 ":tables",
2817 "@FP16",
2818 "@pthreadpool",
2819 ],
2820)
2821
2822xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002823 name = "avx512f_ukernels",
2824 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002825 gcc_copts = xnnpack_gcc_std_copts(),
2826 gcc_x86_copts = ["-mavx512f"],
2827 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2828 msvc_copts = xnnpack_msvc_std_copts(),
2829 msvc_x86_32_copts = ["/arch:AVX512"],
2830 msvc_x86_64_copts = ["/arch:AVX512"],
2831 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002832 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002833 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002834 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002835 "@FP16",
2836 "@pthreadpool",
2837 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002838)
2839
2840xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002841 name = "avx512f_ukernels_test_mode",
2842 hdrs = INTERNAL_HDRS,
2843 copts = [
2844 "-UNDEBUG",
2845 "-DXNN_TEST_MODE=1",
2846 ],
2847 gcc_copts = xnnpack_gcc_std_copts(),
2848 gcc_x86_copts = ["-mavx512f"],
2849 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2850 msvc_copts = xnnpack_msvc_std_copts(),
2851 msvc_x86_32_copts = ["/arch:AVX512"],
2852 msvc_x86_64_copts = ["/arch:AVX512"],
2853 msys_copts = ["-fno-asynchronous-unwind-tables"],
2854 x86_srcs = AVX512F_UKERNELS,
2855 deps = [
2856 ":tables",
2857 "@FP16",
2858 "@pthreadpool",
2859 ],
2860)
2861
2862xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002863 name = "asm_ukernels",
2864 hdrs = ["src/xnnpack/assembly.h"],
2865 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002866 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002867 aarch64_srcs = AARCH64_ASM_UKERNELS,
2868)
2869
Marat Dukhan3b59de22020-06-03 20:15:19 -07002870xnnpack_cc_library(
2871 name = "logging_utils",
2872 srcs = LOGGING_SRCS,
2873 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2874 copts = LOGGING_COPTS + [
2875 "-Isrc",
2876 "-Iinclude",
2877 ] + select({
2878 ":debug_build": [],
2879 "//conditions:default": xnnpack_min_size_copts(),
2880 }),
2881 gcc_copts = xnnpack_gcc_std_copts(),
2882 msvc_copts = xnnpack_msvc_std_copts(),
2883 visibility = xnnpack_visibility(),
2884 deps = [
2885 "@FP16",
2886 "@clog",
2887 "@pthreadpool",
2888 ],
2889)
2890
Marat Dukhan08c4a432019-10-03 09:29:21 -07002891xnnpack_aggregate_library(
2892 name = "ukernels",
2893 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002894 ":neon_ukernels",
2895 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002896 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002897 ":asm_ukernels",
2898 ],
2899 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002900 ":neon_ukernels",
2901 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002902 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002903 ":neonfp16arith_ukernels",
2904 ":asm_ukernels",
2905 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002906 generic_deps = [
2907 ":scalar_ukernels",
2908 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002909 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002910 ":psimd_fastmath_ukernels",
2911 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002912 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002913 wasm_deps = [
2914 ":wasm_ukernels",
2915 ],
2916 wasmsimd_deps = [
2917 ":wasm_ukernels",
2918 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002919 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002920 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002921 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002922 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002923 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002924 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002925 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002926 ":avx512f_ukernels",
2927 ],
2928)
2929
Marat Dukhan33fcf782020-05-24 14:27:15 -07002930xnnpack_aggregate_library(
2931 name = "ukernels_test_mode",
2932 aarch32_deps = [
2933 ":neon_ukernels_test_mode",
2934 ":neonfma_ukernels_test_mode",
2935 ":neonv8_ukernels_test_mode",
2936 ":asm_ukernels",
2937 ],
2938 aarch64_deps = [
2939 ":neon_ukernels_test_mode",
2940 ":neonfma_ukernels_test_mode",
2941 ":neonv8_ukernels_test_mode",
2942 ":neonfp16arith_ukernels_test_mode",
2943 ":asm_ukernels",
2944 ],
2945 generic_deps = [
2946 ":scalar_ukernels_test_mode",
2947 ],
2948 psimd_deps = [
2949 ":psimd_fastmath_ukernels_test_mode",
2950 ":psimd_accmath_ukernels_test_mode",
2951 ],
2952 wasm_deps = [
2953 ":wasm_ukernels_test_mode",
2954 ],
2955 wasmsimd_deps = [
2956 ":wasm_ukernels_test_mode",
2957 ],
2958 x86_deps = [
2959 ":sse2_ukernels_test_mode",
2960 ":ssse3_ukernels_test_mode",
2961 ":sse41_ukernels_test_mode",
2962 ":avx_ukernels_test_mode",
2963 ":fma3_ukernels_test_mode",
2964 ":avx2_ukernels_test_mode",
2965 ":avx512f_ukernels_test_mode",
2966 ],
2967)
2968
Marat Dukhan86df3842020-06-17 14:03:31 -07002969# This target must be compatible with Android NDK r17 and compile with Clang 6.
2970# It is similar to :ukernels target, but excludes microkernels which require
2971# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07002972xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07002973 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07002974 aarch32_deps = [
2975 ":neon_ukernels",
2976 ":neonfma_ukernels",
2977 ":neonv8_ukernels",
2978 ":asm_ukernels",
2979 ],
2980 aarch64_deps = [
2981 ":neon_ukernels",
2982 ":neonfma_ukernels",
2983 ":neonv8_ukernels",
2984 ":asm_ukernels",
2985 ],
2986 generic_deps = [
2987 ":scalar_ukernels",
2988 ],
2989 psimd_deps = [
2990 ":psimd_fastmath_ukernels",
2991 ":psimd_accmath_ukernels",
2992 ],
2993 wasm_deps = [
2994 ":wasm_ukernels",
2995 ],
2996 wasmsimd_deps = [
2997 ":wasm_ukernels",
2998 ],
2999 x86_deps = [
3000 ":sse2_ukernels",
3001 ":ssse3_ukernels",
3002 ":sse41_ukernels",
3003 ":avx_ukernels",
3004 ":fma3_ukernels",
3005 ":avx2_ukernels",
3006 ":avx512f_ukernels",
3007 ],
3008)
3009
Marat Dukhan08c4a432019-10-03 09:29:21 -07003010xnnpack_cc_library(
3011 name = "im2col",
3012 srcs = ["src/im2col.c"],
3013 hdrs = [
3014 "src/xnnpack/common.h",
3015 "src/xnnpack/im2col.h",
3016 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003017 gcc_copts = xnnpack_gcc_std_copts(),
3018 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003019)
3020
3021xnnpack_cc_library(
3022 name = "indirection",
3023 srcs = ["src/indirection.c"],
3024 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003025 gcc_copts = xnnpack_gcc_std_copts(),
3026 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003027 deps = [
3028 "@FP16",
3029 "@FXdiv",
3030 "@pthreadpool",
3031 ],
3032)
3033
3034xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003035 name = "indirection_test_mode",
3036 srcs = ["src/indirection.c"],
3037 hdrs = INTERNAL_HDRS,
3038 copts = [
3039 "-UNDEBUG",
3040 "-DXNN_TEST_MODE=1",
3041 ],
3042 gcc_copts = xnnpack_gcc_std_copts(),
3043 msvc_copts = xnnpack_msvc_std_copts(),
3044 deps = [
3045 "@FP16",
3046 "@FXdiv",
3047 "@pthreadpool",
3048 ],
3049)
3050
3051xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003052 name = "operator_run",
3053 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003054 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003055 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003056 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3057 "//conditions:default": [],
3058 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003059 gcc_copts = xnnpack_gcc_std_copts(),
3060 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003061 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003062 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003063 "@FP16",
3064 "@FXdiv",
3065 "@clog",
3066 "@pthreadpool",
3067 ],
3068)
3069
Chao Mei6ddfc602020-05-13 22:29:36 -07003070xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003071 name = "operator_run_test_mode",
3072 srcs = ["src/operator-run.c"],
3073 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3074 copts = LOGGING_COPTS + [
3075 "-UNDEBUG",
3076 "-DXNN_TEST_MODE=1",
3077 ] + select({
3078 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3079 "//conditions:default": [],
3080 }),
3081 gcc_copts = xnnpack_gcc_std_copts(),
3082 msvc_copts = xnnpack_msvc_std_copts(),
3083 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003084 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003085 "@FP16",
3086 "@FXdiv",
3087 "@clog",
3088 "@pthreadpool",
3089 ],
3090)
3091
3092xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003093 name = "memory_planner",
3094 srcs = ["src/memory-planner.c"],
3095 hdrs = INTERNAL_HDRS,
3096 defines = select({
3097 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3098 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3099 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3100 }),
3101 gcc_copts = xnnpack_gcc_std_copts(),
3102 msvc_copts = xnnpack_msvc_std_copts(),
3103 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003104 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003105 "@pthreadpool",
3106 ],
3107)
3108
Marat Dukhan33fcf782020-05-24 14:27:15 -07003109xnnpack_cc_library(
3110 name = "memory_planner_test_mode",
3111 srcs = ["src/memory-planner.c"],
3112 hdrs = INTERNAL_HDRS,
3113 copts = [
3114 "-UNDEBUG",
3115 "-DXNN_TEST_MODE=1",
3116 ],
3117 defines = select({
3118 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3119 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3120 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3121 }),
3122 gcc_copts = xnnpack_gcc_std_copts(),
3123 msvc_copts = xnnpack_msvc_std_copts(),
3124 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003125 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003126 "@pthreadpool",
3127 ],
3128)
3129
Marat Dukhan08c4a432019-10-03 09:29:21 -07003130cc_library(
3131 name = "enable_assembly",
3132 defines = select({
3133 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3134 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003135 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003136 }),
3137)
3138
Marat Dukhan9de90e02020-06-18 16:04:12 -07003139cc_library(
3140 name = "enable_sparse",
3141 defines = select({
3142 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3143 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003144 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003145 }),
3146)
3147
Marat Dukhancf056b22019-10-07 10:26:29 -07003148xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003149 name = "operators",
3150 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003151 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003152 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003153 ],
3154 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003155 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003156 "-Isrc",
3157 "-Iinclude",
3158 ] + select({
3159 ":debug_build": [],
3160 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003161 }) + select({
3162 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3163 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003164 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003165 gcc_copts = xnnpack_gcc_std_copts(),
3166 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003167 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003168 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003169 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003170 "@FP16",
3171 "@FXdiv",
3172 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003173 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003174 ],
3175)
3176
Marat Dukhan10a38082020-04-17 03:58:35 -07003177xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003178 name = "operators_test_mode",
3179 srcs = OPERATOR_SRCS + [
3180 "src/memory.c",
3181 "src/operator-delete.c",
3182 ],
3183 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3184 copts = LOGGING_COPTS + [
3185 "-Isrc",
3186 "-Iinclude",
3187 "-UNDEBUG",
3188 "-DXNN_TEST_MODE=1",
3189 ] + select({
3190 ":debug_build": [],
3191 "//conditions:default": xnnpack_min_size_copts(),
3192 }) + select({
3193 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3194 "//conditions:default": [],
3195 }),
3196 gcc_copts = xnnpack_gcc_std_copts(),
3197 msvc_copts = xnnpack_msvc_std_copts(),
3198 deps = [
3199 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003200 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003201 "@FP16",
3202 "@FXdiv",
3203 "@clog",
3204 "@pthreadpool",
3205 ],
3206)
3207
3208xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003209 name = "XNNPACK",
3210 srcs = [
3211 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003212 "src/runtime.c",
3213 "src/subgraph.c",
3214 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003215 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003216 hdrs = ["include/xnnpack.h"],
3217 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003218 "-Isrc",
3219 "-Iinclude",
3220 ] + select({
3221 ":debug_build": [],
3222 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003223 }) + select({
3224 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3225 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003226 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003227 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003228 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003229 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003230 visibility = xnnpack_visibility(),
3231 deps = [
3232 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003233 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003234 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003235 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003236 ":operator_run",
3237 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003238 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003239 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003240 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003241 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003242 ] + select({
3243 ":emscripten": [],
3244 "//conditions:default": ["@cpuinfo"],
3245 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003246)
3247
Marat Dukhan10a38082020-04-17 03:58:35 -07003248xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003249 name = "XNNPACK_test_mode",
3250 srcs = [
3251 "src/init.c",
3252 "src/runtime.c",
3253 "src/subgraph.c",
3254 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003255 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003256 hdrs = ["include/xnnpack.h"],
3257 copts = LOGGING_COPTS + [
3258 "-Isrc",
3259 "-Iinclude",
3260 "-UNDEBUG",
3261 "-DXNN_TEST_MODE=1",
3262 ] + select({
3263 ":debug_build": [],
3264 "//conditions:default": xnnpack_min_size_copts(),
3265 }) + select({
3266 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3267 "//conditions:default": [],
3268 }),
3269 gcc_copts = xnnpack_gcc_std_copts(),
3270 includes = ["include"],
3271 msvc_copts = xnnpack_msvc_std_copts(),
3272 visibility = xnnpack_visibility(),
3273 deps = [
3274 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003275 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003276 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003277 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003278 ":operator_run_test_mode",
3279 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003280 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003281 "@clog",
3282 "@FP16",
3283 "@pthreadpool",
3284 ] + select({
3285 ":emscripten": [],
3286 "//conditions:default": ["@cpuinfo"],
3287 }),
3288)
3289
3290xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003291 name = "xnnpack_f32",
3292 srcs = [
3293 "src/init.c",
3294 "src/runtime.c",
3295 "src/subgraph.c",
3296 "src/tensor.c",
3297 ] + SUBGRAPH_SRCS,
3298 hdrs = ["include/xnnpack.h"],
3299 copts = LOGGING_COPTS + [
3300 "-Isrc",
3301 "-Iinclude",
3302 ] + select({
3303 ":debug_build": [],
3304 "//conditions:default": xnnpack_min_size_copts(),
3305 }) + select({
3306 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3307 "//conditions:default": [],
3308 }),
3309 defines = [
3310 "XNN_NO_Q8_OPERATORS",
3311 "XNN_NO_U8_OPERATORS",
3312 "XNN_NO_X8_OPERATORS",
3313 "XNN_NO_F16_OPERATORS",
3314 "XNN_NO_X16_OPERATORS",
3315 ],
3316 gcc_copts = xnnpack_gcc_std_copts(),
3317 includes = ["include"],
3318 msvc_copts = xnnpack_msvc_std_copts(),
3319 visibility = xnnpack_visibility(),
3320 deps = [
3321 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003322 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003323 ":logging_utils",
3324 ":memory_planner",
3325 ":operator_run",
3326 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003327 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003328 "@clog",
3329 "@FP16",
3330 "@pthreadpool",
3331 ] + select({
3332 ":emscripten": [],
3333 "//conditions:default": ["@cpuinfo"],
3334 }),
3335)
3336
3337xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003338 name = "xnnpack_operators_nhwc_f32",
3339 srcs = [
3340 "src/init.c",
3341 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003342 hdrs = ["include/xnnpack.h"],
3343 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003344 "-Isrc",
3345 "-Iinclude",
3346 ] + select({
3347 ":debug_build": [],
3348 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003349 }) + select({
3350 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3351 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003352 }),
3353 defines = [
3354 "XNN_NO_Q8_OPERATORS",
3355 "XNN_NO_U8_OPERATORS",
3356 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003357 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003358 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003359 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003360 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003361 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003362 visibility = xnnpack_visibility(),
3363 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003364 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003365 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003366 ":operator_run",
3367 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003368 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003369 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003370 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003371 ] + select({
3372 ":emscripten": [],
3373 "//conditions:default": ["@cpuinfo"],
3374 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003375)
3376
Marat Dukhancf056b22019-10-07 10:26:29 -07003377xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003378 name = "bench_utils",
3379 srcs = ["bench/utils.cc"],
3380 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003381 deps = [
3382 "@com_google_benchmark//:benchmark",
3383 "@cpuinfo",
3384 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003385)
3386
Frank Barchard7e955972019-10-11 10:34:25 -07003387######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003388
3389xnnpack_benchmark(
3390 name = "q8_gemm_bench",
3391 srcs = [
3392 "bench/gemm.h",
3393 "bench/q8-gemm.cc",
3394 "src/xnnpack/AlignedAllocator.h",
3395 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003396 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003397 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
3398)
3399
3400xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003401 name = "f16_igemm_bench",
3402 srcs = [
3403 "bench/f16-igemm.cc",
3404 "bench/conv.h",
3405 "bench/google/conv.h",
3406 "src/xnnpack/AlignedAllocator.h",
3407 ] + MICROKERNEL_BENCHMARK_HDRS,
3408 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3409)
3410
3411xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003412 name = "f16_gemm_bench",
3413 srcs = [
3414 "bench/f16-gemm.cc",
3415 "bench/gemm.h",
3416 "src/xnnpack/AlignedAllocator.h",
3417 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003418 deps = MICROKERNEL_BENCHMARK_DEPS,
3419)
3420
3421xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003422 name = "f16_spmm_bench",
3423 srcs = [
3424 "bench/f16-spmm.cc",
3425 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003426 "src/xnnpack/AlignedAllocator.h",
3427 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003428 deps = MICROKERNEL_BENCHMARK_DEPS,
3429)
3430
3431xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003432 name = "f32_igemm_bench",
3433 srcs = [
3434 "bench/f32-igemm.cc",
3435 "bench/conv.h",
3436 "src/xnnpack/AlignedAllocator.h",
3437 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07003438 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003439)
3440
3441xnnpack_benchmark(
3442 name = "f32_conv_hwc_bench",
3443 srcs = [
3444 "bench/f32-conv-hwc.cc",
3445 "bench/dconv.h",
3446 "src/xnnpack/AlignedAllocator.h",
3447 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003448 deps = MICROKERNEL_BENCHMARK_DEPS,
3449)
3450
3451xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003452 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003453 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003454 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003455 "bench/dconv.h",
3456 "src/xnnpack/AlignedAllocator.h",
3457 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Erich Elsen563df5f2019-10-23 08:02:21 -07003458 deps = MICROKERNEL_BENCHMARK_DEPS,
3459)
3460
3461xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003462 name = "f16_dwconv_bench",
3463 srcs = [
3464 "bench/f16-dwconv.cc",
3465 "bench/dwconv.h",
3466 "bench/google/dwconv.h",
3467 "src/xnnpack/AlignedAllocator.h",
3468 ] + MICROKERNEL_BENCHMARK_HDRS,
3469 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3470)
3471
3472xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003473 name = "f32_dwconv_bench",
3474 srcs = [
3475 "bench/f32-dwconv.cc",
3476 "bench/dwconv.h",
3477 "src/xnnpack/AlignedAllocator.h",
3478 ] + MICROKERNEL_BENCHMARK_HDRS,
3479 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3480)
3481
3482xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003483 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003484 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003485 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003486 "bench/dwconv.h",
3487 "src/xnnpack/AlignedAllocator.h",
3488 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3489 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3490)
3491
3492xnnpack_benchmark(
3493 name = "f32_gemm_bench",
3494 srcs = [
3495 "bench/f32-gemm.cc",
3496 "bench/gemm.h",
3497 "src/xnnpack/AlignedAllocator.h",
3498 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003499 copts = xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07003500 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003501)
3502
3503xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003504 name = "f32_hswish_bench",
3505 srcs = [
3506 "bench/f32-hswish.cc",
3507 "src/xnnpack/AlignedAllocator.h",
3508 ] + MICROKERNEL_BENCHMARK_HDRS,
3509 deps = MICROKERNEL_BENCHMARK_DEPS,
3510)
3511
3512xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003513 name = "f32_raddexpminusmax_bench",
3514 srcs = [
3515 "bench/f32-raddexpminusmax.cc",
3516 "src/xnnpack/AlignedAllocator.h",
3517 ] + MICROKERNEL_BENCHMARK_HDRS,
3518 deps = MICROKERNEL_BENCHMARK_DEPS,
3519)
3520
3521xnnpack_benchmark(
3522 name = "f32_raddextexp_bench",
3523 srcs = [
3524 "bench/f32-raddextexp.cc",
3525 "src/xnnpack/AlignedAllocator.h",
3526 ] + MICROKERNEL_BENCHMARK_HDRS,
3527 deps = MICROKERNEL_BENCHMARK_DEPS,
3528)
3529
3530xnnpack_benchmark(
3531 name = "f32_raddstoreexpminusmax_bench",
3532 srcs = [
3533 "bench/f32-raddstoreexpminusmax.cc",
3534 "src/xnnpack/AlignedAllocator.h",
3535 ] + MICROKERNEL_BENCHMARK_HDRS,
3536 deps = MICROKERNEL_BENCHMARK_DEPS,
3537)
3538
3539xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003540 name = "f32_rmax_bench",
3541 srcs = [
3542 "bench/f32-rmax.cc",
3543 "src/xnnpack/AlignedAllocator.h",
3544 ] + MICROKERNEL_BENCHMARK_HDRS,
3545 deps = MICROKERNEL_BENCHMARK_DEPS,
3546)
3547
3548xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003549 name = "f32_sigmoid_bench",
3550 srcs = [
3551 "bench/f32-sigmoid.cc",
3552 "src/xnnpack/AlignedAllocator.h",
3553 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003554 deps = MICROKERNEL_BENCHMARK_DEPS,
3555)
3556
3557xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003558 name = "f32_spmm_bench",
3559 srcs = [
3560 "bench/f32-spmm.cc",
3561 "bench/gemm.h",
3562 "src/xnnpack/AlignedAllocator.h",
3563 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003564 deps = MICROKERNEL_BENCHMARK_DEPS,
3565)
3566
3567xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003568 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003569 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003570 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003571 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003572 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003573 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003574)
3575
3576xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003577 name = "f32_vscaleexpminusmax_bench",
3578 srcs = [
3579 "bench/f32-vscaleexpminusmax.cc",
3580 "src/xnnpack/AlignedAllocator.h",
3581 ] + MICROKERNEL_BENCHMARK_HDRS,
3582 deps = MICROKERNEL_BENCHMARK_DEPS,
3583)
3584
3585xnnpack_benchmark(
3586 name = "f32_vscaleextexp_bench",
3587 srcs = [
3588 "bench/f32-vscaleextexp.cc",
3589 "src/xnnpack/AlignedAllocator.h",
3590 ] + MICROKERNEL_BENCHMARK_HDRS,
3591 deps = MICROKERNEL_BENCHMARK_DEPS,
3592)
3593
3594xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003595 name = "f32_im2col_gemm_bench",
3596 srcs = [
3597 "bench/f32-im2col-gemm.cc",
3598 "bench/conv.h",
3599 "src/xnnpack/AlignedAllocator.h",
3600 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3601 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
3602)
3603
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003604xnnpack_benchmark(
3605 name = "requantization_bench",
3606 srcs = [
3607 "bench/requantization.cc",
3608 "src/xnnpack/requantization-stubs.h",
3609 "src/xnnpack/AlignedAllocator.h",
3610 ] + MICROKERNEL_BENCHMARK_HDRS,
3611 deps = MICROKERNEL_BENCHMARK_DEPS,
3612)
3613
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003614xnnpack_benchmark(
3615 name = "rounding_bench",
3616 srcs = [
3617 "bench/rounding.cc",
3618 "src/xnnpack/math-stubs.h",
3619 "src/xnnpack/AlignedAllocator.h",
3620 ] + MICROKERNEL_BENCHMARK_HDRS,
3621 deps = MICROKERNEL_BENCHMARK_DEPS,
3622)
3623
Marat Dukhan08c4a432019-10-03 09:29:21 -07003624########################### Benchmarks for operators ###########################
3625
3626xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003627 name = "average_pooling_bench",
3628 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003629 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003630 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003631 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003632)
3633
3634xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003635 name = "bankers_rounding_bench",
3636 srcs = ["bench/bankers-rounding.cc"],
3637 copts = xnnpack_optional_tflite_copts(),
3638 tags = ["nowin32"],
3639 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3640)
3641
3642xnnpack_benchmark(
3643 name = "ceiling_bench",
3644 srcs = ["bench/ceiling.cc"],
3645 copts = xnnpack_optional_tflite_copts(),
3646 tags = ["nowin32"],
3647 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3648)
3649
3650xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003651 name = "channel_shuffle_bench",
3652 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003653 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003654)
3655
3656xnnpack_benchmark(
3657 name = "convolution_bench",
3658 srcs = ["bench/convolution.cc"],
3659 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003660 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003661 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003662)
3663
3664xnnpack_benchmark(
3665 name = "deconvolution_bench",
3666 srcs = ["bench/deconvolution.cc"],
3667 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003668 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003669 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003670)
3671
3672xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003673 name = "floor_bench",
3674 srcs = ["bench/floor.cc"],
3675 copts = xnnpack_optional_tflite_copts(),
3676 tags = ["nowin32"],
3677 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3678)
3679
3680xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003681 name = "global_average_pooling_bench",
3682 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003683 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003684)
3685
3686xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07003687 name = "hardswish_bench",
3688 srcs = ["bench/hardswish.cc"],
3689 copts = xnnpack_optional_tflite_copts(),
3690 tags = ["nowin32"],
3691 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3692)
3693
3694xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003695 name = "max_pooling_bench",
3696 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003697 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003698)
3699
3700xnnpack_benchmark(
3701 name = "sigmoid_bench",
3702 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003703 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003704 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003705 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003706)
3707
3708xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003709 name = "prelu_bench",
3710 srcs = ["bench/prelu.cc"],
3711 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003712 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003713 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003714)
3715
3716xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003717 name = "softmax_bench",
3718 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003719 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003720 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003721 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003722)
3723
Marat Dukhan87727142020-06-24 15:24:10 -07003724xnnpack_benchmark(
3725 name = "truncation_bench",
3726 srcs = ["bench/truncation.cc"],
3727 deps = OPERATOR_BENCHMARK_DEPS,
3728)
3729
Marat Dukhanc068bb62019-10-04 13:24:39 -07003730############################# End-to-end benchmarks ############################
3731
3732cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003733 name = "fp32_mobilenet_v1",
3734 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003735 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003736 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003737 linkstatic = True,
3738 deps = [
3739 ":XNNPACK",
3740 "@pthreadpool",
3741 ],
3742)
3743
3744cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003745 name = "fp16_mobilenet_v1",
3746 srcs = ["models/fp16-mobilenet-v1.cc"],
3747 hdrs = ["models/models.h"],
3748 copts = xnnpack_std_cxxopts(),
3749 linkstatic = True,
3750 deps = [
3751 ":XNNPACK",
3752 "@FP16",
3753 "@pthreadpool",
3754 ],
3755)
3756
3757cc_library(
3758 name = "fp32_mobilenet_v2",
3759 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003760 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003761 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003762 linkstatic = True,
3763 deps = [
3764 ":XNNPACK",
3765 "@pthreadpool",
3766 ],
3767)
3768
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003769cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003770 name = "fp16_mobilenet_v2",
3771 srcs = ["models/fp16-mobilenet-v2.cc"],
3772 hdrs = ["models/models.h"],
3773 copts = xnnpack_std_cxxopts(),
3774 linkstatic = True,
3775 deps = [
3776 ":XNNPACK",
3777 "@FP16",
3778 "@pthreadpool",
3779 ],
3780)
3781
3782cc_library(
3783 name = "fp32_mobilenet_v3_large",
3784 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003785 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003786 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003787 linkstatic = True,
3788 deps = [
3789 ":XNNPACK",
3790 "@pthreadpool",
3791 ],
3792)
3793
3794cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003795 name = "fp32_mobilenet_v3_small",
3796 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003797 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003798 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003799 linkstatic = True,
3800 deps = [
3801 ":XNNPACK",
3802 "@pthreadpool",
3803 ],
3804)
3805
Marat Dukhanc068bb62019-10-04 13:24:39 -07003806xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07003807 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003808 srcs = [
3809 "bench/f32-dwconv-e2e.cc",
3810 "bench/end2end.h",
3811 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07003812 deps = MICROKERNEL_BENCHMARK_DEPS + [
3813 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003814 ":fp32_mobilenet_v1",
3815 ":fp32_mobilenet_v2",
3816 ":fp32_mobilenet_v3_large",
3817 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07003818 ],
3819)
3820
3821xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07003822 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003823 srcs = [
3824 "bench/f32-gemm-e2e.cc",
3825 "bench/end2end.h",
3826 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07003827 deps = MICROKERNEL_BENCHMARK_DEPS + [
3828 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003829 ":fp32_mobilenet_v1",
3830 ":fp32_mobilenet_v2",
3831 ":fp32_mobilenet_v3_large",
3832 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07003833 ],
3834)
3835
3836xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07003837 name = "end2end_bench",
3838 srcs = ["bench/end2end.cc"],
3839 deps = [
3840 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07003841 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003842 ":fp16_mobilenet_v1",
3843 ":fp16_mobilenet_v2",
3844 ":fp32_mobilenet_v1",
3845 ":fp32_mobilenet_v2",
3846 ":fp32_mobilenet_v3_large",
3847 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07003848 "@pthreadpool",
3849 ],
3850)
3851
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003852#################### Accuracy evaluation for math functions ####################
3853
3854xnnpack_benchmark(
3855 name = "f32_exp_eval",
3856 srcs = [
3857 "eval/f32-exp.cc",
3858 "src/xnnpack/AlignedAllocator.h",
3859 ] + ACCURACY_EVAL_HDRS,
3860 deps = ACCURACY_EVAL_DEPS,
3861)
3862
Marat Dukhan515c9772019-10-17 18:07:57 -07003863xnnpack_benchmark(
3864 name = "f32_expminus_eval",
3865 srcs = [
3866 "eval/f32-expminus.cc",
3867 "src/xnnpack/AlignedAllocator.h",
3868 ] + ACCURACY_EVAL_HDRS,
3869 deps = ACCURACY_EVAL_DEPS,
3870)
3871
Marat Dukhan98ba4412019-10-23 02:14:28 -07003872xnnpack_benchmark(
3873 name = "f32_extexp_eval",
3874 srcs = [
3875 "eval/f32-extexp.cc",
3876 "src/xnnpack/AlignedAllocator.h",
3877 ] + ACCURACY_EVAL_HDRS,
3878 deps = ACCURACY_EVAL_DEPS,
3879)
3880
Marat Dukhan8853b822020-05-07 12:19:01 -07003881xnnpack_unit_test(
3882 name = "f32_roundne_eval",
3883 srcs = [
3884 "eval/f32-roundne.cc",
3885 "src/xnnpack/AlignedAllocator.h",
3886 "src/xnnpack/math-stubs.h",
3887 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07003888 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07003889 deps = MICROKERNEL_TEST_DEPS,
3890)
3891
Marat Dukhan2dbb9442020-05-12 20:43:43 -07003892xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07003893 name = "f32_roundd_eval",
3894 srcs = [
3895 "eval/f32-roundd.cc",
3896 "src/xnnpack/AlignedAllocator.h",
3897 "src/xnnpack/math-stubs.h",
3898 ] + MICROKERNEL_TEST_HDRS,
3899 automatic = False,
3900 deps = MICROKERNEL_TEST_DEPS,
3901)
3902
3903xnnpack_unit_test(
3904 name = "f32_roundu_eval",
3905 srcs = [
3906 "eval/f32-roundu.cc",
3907 "src/xnnpack/AlignedAllocator.h",
3908 "src/xnnpack/math-stubs.h",
3909 ] + MICROKERNEL_TEST_HDRS,
3910 automatic = False,
3911 deps = MICROKERNEL_TEST_DEPS,
3912)
3913
3914xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07003915 name = "f32_roundz_eval",
3916 srcs = [
3917 "eval/f32-roundz.cc",
3918 "src/xnnpack/AlignedAllocator.h",
3919 "src/xnnpack/math-stubs.h",
3920 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07003921 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07003922 deps = MICROKERNEL_TEST_DEPS,
3923)
3924
Marat Dukhan346a9e52019-11-15 09:06:30 -08003925xnnpack_benchmark(
3926 name = "f32_sigmoid_eval",
3927 srcs = [
3928 "eval/f32-sigmoid.cc",
3929 "src/xnnpack/AlignedAllocator.h",
3930 ] + ACCURACY_EVAL_HDRS,
3931 deps = ACCURACY_EVAL_DEPS,
3932)
3933
Marat Dukhan84000762020-06-29 18:38:43 -07003934xnnpack_benchmark(
3935 name = "f32_sqrt_eval",
3936 srcs = [
3937 "eval/f32-sqrt.cc",
3938 "src/xnnpack/AlignedAllocator.h",
3939 ] + ACCURACY_EVAL_HDRS,
3940 deps = ACCURACY_EVAL_DEPS + [
3941 ":bench_utils",
3942 ],
3943)
3944
Marat Dukhan08c4a432019-10-03 09:29:21 -07003945######################### Unit tests for micro-kernels #########################
3946
3947xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003948 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003949 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003950 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003951 "test/gemm-microkernel-tester.h",
3952 "src/xnnpack/AlignedAllocator.h",
3953 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3954 deps = MICROKERNEL_TEST_DEPS,
3955)
3956
3957xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07003958 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003959 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07003960 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003961 "test/spmm-microkernel-tester.h",
3962 "src/xnnpack/AlignedAllocator.h",
3963 ] + MICROKERNEL_TEST_HDRS,
3964 deps = MICROKERNEL_TEST_DEPS,
3965)
3966
3967xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003968 name = "f32_argmaxpool_test",
3969 srcs = [
3970 "test/f32-argmaxpool.cc",
3971 "test/argmaxpool-microkernel-tester.h",
3972 "src/xnnpack/AlignedAllocator.h",
3973 ] + MICROKERNEL_TEST_HDRS,
3974 deps = MICROKERNEL_TEST_DEPS,
3975)
3976
3977xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003978 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003979 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003980 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003981 "test/avgpool-microkernel-tester.h",
3982 "src/xnnpack/AlignedAllocator.h",
3983 ] + MICROKERNEL_TEST_HDRS,
3984 deps = MICROKERNEL_TEST_DEPS,
3985)
3986
3987xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07003988 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08003989 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07003990 "test/f32-ibilinear.cc",
3991 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08003992 "src/xnnpack/AlignedAllocator.h",
3993 ] + MICROKERNEL_TEST_HDRS,
3994 deps = MICROKERNEL_TEST_DEPS,
3995)
3996
3997xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07003998 name = "f16_clamp_test",
3999 srcs = [
4000 "test/f16-clamp.cc",
4001 "test/clamp-microkernel-tester.h",
4002 ] + MICROKERNEL_TEST_HDRS,
4003 deps = MICROKERNEL_TEST_DEPS,
4004)
4005
4006xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004007 name = "f32_clamp_test",
4008 srcs = [
4009 "test/f32-clamp.cc",
4010 "test/clamp-microkernel-tester.h",
4011 ] + MICROKERNEL_TEST_HDRS,
4012 deps = MICROKERNEL_TEST_DEPS,
4013)
4014
4015xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004016 name = "f32_igemm_test",
4017 srcs = [
4018 "test/f32-igemm.cc",
4019 "test/gemm-microkernel-tester.h",
4020 "src/xnnpack/AlignedAllocator.h",
4021 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4022 deps = MICROKERNEL_TEST_DEPS,
4023)
4024
4025xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004026 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004027 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004028 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004029 "test/gemm-microkernel-tester.h",
4030 "src/xnnpack/AlignedAllocator.h",
4031 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4032 deps = MICROKERNEL_TEST_DEPS,
4033)
4034
4035xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004036 name = "f32_igemm_minmax_test",
4037 srcs = [
4038 "test/f32-igemm-minmax.cc",
4039 "test/gemm-microkernel-tester.h",
4040 "src/xnnpack/AlignedAllocator.h",
4041 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4042 deps = MICROKERNEL_TEST_DEPS,
4043)
4044
4045xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004046 name = "f16_igemm_minmax_test",
4047 srcs = [
4048 "test/f16-igemm-minmax.cc",
4049 "test/gemm-microkernel-tester.h",
4050 "src/xnnpack/AlignedAllocator.h",
4051 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4052 deps = MICROKERNEL_TEST_DEPS,
4053)
4054
4055xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004056 name = "f32_conv_hwc_test",
4057 srcs = [
4058 "test/f32-conv-hwc.cc",
4059 "test/conv-hwc-microkernel-tester.h",
4060 "src/xnnpack/AlignedAllocator.h",
4061 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4062 deps = MICROKERNEL_TEST_DEPS,
4063)
4064
4065xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004066 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004067 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004068 "test/f32-conv-hwc2chw.cc",
4069 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004070 "src/xnnpack/AlignedAllocator.h",
4071 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4072 deps = MICROKERNEL_TEST_DEPS,
4073)
4074
4075xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004076 name = "f32_dwconv_test",
4077 srcs = [
4078 "test/f32-dwconv.cc",
4079 "test/dwconv-microkernel-tester.h",
4080 "src/xnnpack/AlignedAllocator.h",
4081 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4082 deps = MICROKERNEL_TEST_DEPS,
4083)
4084
4085xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004086 name = "f16_dwconv_minmax_test",
4087 srcs = [
4088 "test/f16-dwconv-minmax.cc",
4089 "test/dwconv-microkernel-tester.h",
4090 "src/xnnpack/AlignedAllocator.h",
4091 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4092 deps = MICROKERNEL_TEST_DEPS,
4093)
4094
4095xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004096 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004097 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004098 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004099 "test/dwconv-microkernel-tester.h",
4100 "src/xnnpack/AlignedAllocator.h",
4101 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4102 deps = MICROKERNEL_TEST_DEPS,
4103)
4104
4105xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004106 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004107 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004108 "test/f32-dwconv-chw.cc",
4109 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004110 "src/xnnpack/AlignedAllocator.h",
4111 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4112 deps = MICROKERNEL_TEST_DEPS,
4113)
4114
4115xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004116 name = "f16_gavgpool_minmax_test",
4117 srcs = [
4118 "test/f16-gavgpool-minmax.cc",
4119 "test/gavgpool-microkernel-tester.h",
4120 "src/xnnpack/AlignedAllocator.h",
4121 ] + MICROKERNEL_TEST_HDRS,
4122 deps = MICROKERNEL_TEST_DEPS,
4123)
4124
4125xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004126 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004127 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004128 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004129 "test/gavgpool-microkernel-tester.h",
4130 "src/xnnpack/AlignedAllocator.h",
4131 ] + MICROKERNEL_TEST_HDRS,
4132 deps = MICROKERNEL_TEST_DEPS,
4133)
4134
4135xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004136 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004137 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004138 "test/f32-gavgpool-cw.cc",
4139 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004140 "src/xnnpack/AlignedAllocator.h",
4141 ] + MICROKERNEL_TEST_HDRS,
4142 deps = MICROKERNEL_TEST_DEPS,
4143)
4144
4145xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004146 name = "f32_gemm_test",
4147 srcs = [
4148 "test/f32-gemm.cc",
4149 "test/gemm-microkernel-tester.h",
4150 "src/xnnpack/AlignedAllocator.h",
4151 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4152 deps = MICROKERNEL_TEST_DEPS,
4153)
4154
4155xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004156 name = "f32_gemm_relu_test",
4157 srcs = [
4158 "test/f32-gemm-relu.cc",
4159 "test/gemm-microkernel-tester.h",
4160 "src/xnnpack/AlignedAllocator.h",
4161 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4162 deps = MICROKERNEL_TEST_DEPS,
4163)
4164
4165xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004166 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004167 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004168 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004169 "test/gemm-microkernel-tester.h",
4170 "src/xnnpack/AlignedAllocator.h",
4171 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4172 deps = MICROKERNEL_TEST_DEPS,
4173)
4174
4175xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004176 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004177 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004178 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004179 "test/gemm-microkernel-tester.h",
4180 "src/xnnpack/AlignedAllocator.h",
4181 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4182 deps = MICROKERNEL_TEST_DEPS,
4183)
4184
4185xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004186 name = "f16_hswish_test",
4187 srcs = [
4188 "test/f16-hswish.cc",
4189 "test/hswish-microkernel-tester.h",
4190 ] + MICROKERNEL_TEST_HDRS,
4191 deps = MICROKERNEL_TEST_DEPS,
4192)
4193
4194xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004195 name = "f32_hswish_test",
4196 srcs = [
4197 "test/f32-hswish.cc",
4198 "test/hswish-microkernel-tester.h",
4199 ] + MICROKERNEL_TEST_HDRS,
4200 deps = MICROKERNEL_TEST_DEPS,
4201)
4202
4203xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004204 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004205 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004206 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004207 "test/maxpool-microkernel-tester.h",
4208 ] + MICROKERNEL_TEST_HDRS,
4209 deps = MICROKERNEL_TEST_DEPS,
4210)
4211
4212xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004213 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004214 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004215 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004216 "test/avgpool-microkernel-tester.h",
4217 "src/xnnpack/AlignedAllocator.h",
4218 ] + MICROKERNEL_TEST_HDRS,
4219 deps = MICROKERNEL_TEST_DEPS,
4220)
4221
4222xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004223 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004224 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004225 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004226 "test/gemm-microkernel-tester.h",
4227 "src/xnnpack/AlignedAllocator.h",
4228 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4229 deps = MICROKERNEL_TEST_DEPS,
4230)
4231
4232xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004233 name = "f16_prelu_test",
4234 srcs = [
4235 "test/f16-prelu.cc",
4236 "test/prelu-microkernel-tester.h",
4237 "src/xnnpack/AlignedAllocator.h",
4238 ] + MICROKERNEL_TEST_HDRS,
4239 deps = MICROKERNEL_TEST_DEPS,
4240)
4241
4242xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004243 name = "f32_prelu_test",
4244 srcs = [
4245 "test/f32-prelu.cc",
4246 "test/prelu-microkernel-tester.h",
4247 "src/xnnpack/AlignedAllocator.h",
4248 ] + MICROKERNEL_TEST_HDRS,
4249 deps = MICROKERNEL_TEST_DEPS,
4250)
4251
4252xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004253 name = "f32_raddexpminusmax_test",
4254 srcs = [
4255 "test/f32-raddexpminusmax.cc",
4256 "test/raddexpminusmax-microkernel-tester.h",
4257 ] + MICROKERNEL_TEST_HDRS,
4258 deps = MICROKERNEL_TEST_DEPS,
4259)
4260
4261xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004262 name = "f32_raddextexp_test",
4263 srcs = [
4264 "test/f32-raddextexp.cc",
4265 "test/raddextexp-microkernel-tester.h",
4266 ] + MICROKERNEL_TEST_HDRS,
4267 deps = MICROKERNEL_TEST_DEPS,
4268)
4269
4270xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004271 name = "f32_raddstoreexpminusmax_test",
4272 srcs = [
4273 "test/f32-raddstoreexpminusmax.cc",
4274 "test/raddstoreexpminusmax-microkernel-tester.h",
4275 ] + MICROKERNEL_TEST_HDRS,
4276 deps = MICROKERNEL_TEST_DEPS,
4277)
4278
4279xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004280 name = "f32_rmax_test",
4281 srcs = [
4282 "test/f32-rmax.cc",
4283 "test/rmax-microkernel-tester.h",
4284 ] + MICROKERNEL_TEST_HDRS,
4285 deps = MICROKERNEL_TEST_DEPS,
4286)
4287
4288xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004289 name = "f32_sigmoid_test",
4290 srcs = [
4291 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004292 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004293 ] + MICROKERNEL_TEST_HDRS,
4294 deps = MICROKERNEL_TEST_DEPS,
4295)
4296
4297xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004298 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004299 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004300 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004301 "test/spmm-microkernel-tester.h",
4302 "src/xnnpack/AlignedAllocator.h",
4303 ] + MICROKERNEL_TEST_HDRS,
4304 deps = MICROKERNEL_TEST_DEPS,
4305)
4306
4307xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004308 name = "f16_vadd_minmax_test",
4309 srcs = [
4310 "test/f16-vadd-minmax.cc",
4311 "test/vbinary-microkernel-tester.h",
4312 ] + MICROKERNEL_TEST_HDRS,
4313 deps = MICROKERNEL_TEST_DEPS,
4314)
4315
4316xnnpack_unit_test(
4317 name = "f16_vaddc_minmax_test",
4318 srcs = [
4319 "test/f16-vaddc-minmax.cc",
4320 "test/vbinaryc-microkernel-tester.h",
4321 ] + MICROKERNEL_TEST_HDRS,
4322 deps = MICROKERNEL_TEST_DEPS,
4323)
4324
4325xnnpack_unit_test(
4326 name = "f16_vdiv_minmax_test",
4327 srcs = [
4328 "test/f16-vdiv-minmax.cc",
4329 "test/vbinary-microkernel-tester.h",
4330 ] + MICROKERNEL_TEST_HDRS,
4331 deps = MICROKERNEL_TEST_DEPS,
4332)
4333
4334xnnpack_unit_test(
4335 name = "f16_vdivc_minmax_test",
4336 srcs = [
4337 "test/f16-vdivc-minmax.cc",
4338 "test/vbinaryc-microkernel-tester.h",
4339 ] + MICROKERNEL_TEST_HDRS,
4340 deps = MICROKERNEL_TEST_DEPS,
4341)
4342
4343xnnpack_unit_test(
4344 name = "f16_vrdivc_minmax_test",
4345 srcs = [
4346 "test/f16-vrdivc-minmax.cc",
4347 "test/vbinaryc-microkernel-tester.h",
4348 ] + MICROKERNEL_TEST_HDRS,
4349 deps = MICROKERNEL_TEST_DEPS,
4350)
4351
4352xnnpack_unit_test(
4353 name = "f16_vmax_test",
4354 srcs = [
4355 "test/f16-vmax.cc",
4356 "test/vbinary-microkernel-tester.h",
4357 ] + MICROKERNEL_TEST_HDRS,
4358 deps = MICROKERNEL_TEST_DEPS,
4359)
4360
4361xnnpack_unit_test(
4362 name = "f16_vmaxc_test",
4363 srcs = [
4364 "test/f16-vmaxc.cc",
4365 "test/vbinaryc-microkernel-tester.h",
4366 ] + MICROKERNEL_TEST_HDRS,
4367 deps = MICROKERNEL_TEST_DEPS,
4368)
4369
4370xnnpack_unit_test(
4371 name = "f16_vmin_test",
4372 srcs = [
4373 "test/f16-vmin.cc",
4374 "test/vbinary-microkernel-tester.h",
4375 ] + MICROKERNEL_TEST_HDRS,
4376 deps = MICROKERNEL_TEST_DEPS,
4377)
4378
4379xnnpack_unit_test(
4380 name = "f16_vminc_test",
4381 srcs = [
4382 "test/f16-vminc.cc",
4383 "test/vbinaryc-microkernel-tester.h",
4384 ] + MICROKERNEL_TEST_HDRS,
4385 deps = MICROKERNEL_TEST_DEPS,
4386)
4387
4388xnnpack_unit_test(
4389 name = "f16_vmul_minmax_test",
4390 srcs = [
4391 "test/f16-vmul-minmax.cc",
4392 "test/vbinary-microkernel-tester.h",
4393 ] + MICROKERNEL_TEST_HDRS,
4394 deps = MICROKERNEL_TEST_DEPS,
4395)
4396
4397xnnpack_unit_test(
4398 name = "f16_vmulc_minmax_test",
4399 srcs = [
4400 "test/f16-vmulc-minmax.cc",
4401 "test/vbinaryc-microkernel-tester.h",
4402 ] + MICROKERNEL_TEST_HDRS,
4403 deps = MICROKERNEL_TEST_DEPS,
4404)
4405
4406xnnpack_unit_test(
4407 name = "f16_vsub_minmax_test",
4408 srcs = [
4409 "test/f16-vsub-minmax.cc",
4410 "test/vbinary-microkernel-tester.h",
4411 ] + MICROKERNEL_TEST_HDRS,
4412 deps = MICROKERNEL_TEST_DEPS,
4413)
4414
4415xnnpack_unit_test(
4416 name = "f16_vsubc_minmax_test",
4417 srcs = [
4418 "test/f16-vsubc-minmax.cc",
4419 "test/vbinaryc-microkernel-tester.h",
4420 ] + MICROKERNEL_TEST_HDRS,
4421 deps = MICROKERNEL_TEST_DEPS,
4422)
4423
4424xnnpack_unit_test(
4425 name = "f16_vrsubc_minmax_test",
4426 srcs = [
4427 "test/f16-vrsubc-minmax.cc",
4428 "test/vbinaryc-microkernel-tester.h",
4429 ] + MICROKERNEL_TEST_HDRS,
4430 deps = MICROKERNEL_TEST_DEPS,
4431)
4432
4433xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004434 name = "f32_vabs_test",
4435 srcs = [
4436 "test/f32-vabs.cc",
4437 "test/vunary-microkernel-tester.h",
4438 ] + MICROKERNEL_TEST_HDRS,
4439 deps = MICROKERNEL_TEST_DEPS,
4440)
4441
4442xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004443 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004444 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004445 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004446 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004447 ] + MICROKERNEL_TEST_HDRS,
4448 deps = MICROKERNEL_TEST_DEPS,
4449)
4450
4451xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004452 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004453 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004454 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004455 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004456 ] + MICROKERNEL_TEST_HDRS,
4457 deps = MICROKERNEL_TEST_DEPS,
4458)
4459
4460xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004461 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004462 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004463 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004464 "test/vbinary-microkernel-tester.h",
4465 ] + MICROKERNEL_TEST_HDRS,
4466 deps = MICROKERNEL_TEST_DEPS,
4467)
4468
4469xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004470 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004471 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004472 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004473 "test/vbinaryc-microkernel-tester.h",
4474 ] + MICROKERNEL_TEST_HDRS,
4475 deps = MICROKERNEL_TEST_DEPS,
4476)
4477
4478xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004479 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004480 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004481 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004482 "test/vbinaryc-microkernel-tester.h",
4483 ] + MICROKERNEL_TEST_HDRS,
4484 deps = MICROKERNEL_TEST_DEPS,
4485)
4486
4487xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004488 name = "f32_vmax_test",
4489 srcs = [
4490 "test/f32-vmax.cc",
4491 "test/vbinary-microkernel-tester.h",
4492 ] + MICROKERNEL_TEST_HDRS,
4493 deps = MICROKERNEL_TEST_DEPS,
4494)
4495
4496xnnpack_unit_test(
4497 name = "f32_vmaxc_test",
4498 srcs = [
4499 "test/f32-vmaxc.cc",
4500 "test/vbinaryc-microkernel-tester.h",
4501 ] + MICROKERNEL_TEST_HDRS,
4502 deps = MICROKERNEL_TEST_DEPS,
4503)
4504
4505xnnpack_unit_test(
4506 name = "f32_vmin_test",
4507 srcs = [
4508 "test/f32-vmin.cc",
4509 "test/vbinary-microkernel-tester.h",
4510 ] + MICROKERNEL_TEST_HDRS,
4511 deps = MICROKERNEL_TEST_DEPS,
4512)
4513
4514xnnpack_unit_test(
4515 name = "f32_vminc_test",
4516 srcs = [
4517 "test/f32-vminc.cc",
4518 "test/vbinaryc-microkernel-tester.h",
4519 ] + MICROKERNEL_TEST_HDRS,
4520 deps = MICROKERNEL_TEST_DEPS,
4521)
4522
4523xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004524 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004525 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004526 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004527 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004528 ] + MICROKERNEL_TEST_HDRS,
4529 deps = MICROKERNEL_TEST_DEPS,
4530)
4531
4532xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004533 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004534 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004535 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004536 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004537 ] + MICROKERNEL_TEST_HDRS,
4538 deps = MICROKERNEL_TEST_DEPS,
4539)
4540
4541xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004542 name = "f16_vmulcaddc_minmax_test",
4543 srcs = [
4544 "test/f16-vmulcaddc-minmax.cc",
4545 "test/vmulcaddc-microkernel-tester.h",
4546 "src/xnnpack/AlignedAllocator.h",
4547 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4548 deps = MICROKERNEL_TEST_DEPS,
4549)
4550
4551xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004552 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004553 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004554 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004555 "test/vmulcaddc-microkernel-tester.h",
4556 "src/xnnpack/AlignedAllocator.h",
4557 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4558 deps = MICROKERNEL_TEST_DEPS,
4559)
4560
4561xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004562 name = "f32_vlrelu_test",
4563 srcs = [
4564 "test/f32-vlrelu.cc",
4565 "test/vunary-microkernel-tester.h",
4566 ] + MICROKERNEL_TEST_HDRS,
4567 deps = MICROKERNEL_TEST_DEPS,
4568)
4569
4570xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004571 name = "f32_vneg_test",
4572 srcs = [
4573 "test/f32-vneg.cc",
4574 "test/vunary-microkernel-tester.h",
4575 ] + MICROKERNEL_TEST_HDRS,
4576 deps = MICROKERNEL_TEST_DEPS,
4577)
4578
4579xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004580 name = "f32_vrndne_test",
4581 srcs = [
4582 "test/f32-vrndne.cc",
4583 "test/vunary-microkernel-tester.h",
4584 ] + MICROKERNEL_TEST_HDRS,
4585 deps = MICROKERNEL_TEST_DEPS,
4586)
4587
4588xnnpack_unit_test(
4589 name = "f32_vrndz_test",
4590 srcs = [
4591 "test/f32-vrndz.cc",
4592 "test/vunary-microkernel-tester.h",
4593 ] + MICROKERNEL_TEST_HDRS,
4594 deps = MICROKERNEL_TEST_DEPS,
4595)
4596
4597xnnpack_unit_test(
4598 name = "f32_vrndu_test",
4599 srcs = [
4600 "test/f32-vrndu.cc",
4601 "test/vunary-microkernel-tester.h",
4602 ] + MICROKERNEL_TEST_HDRS,
4603 deps = MICROKERNEL_TEST_DEPS,
4604)
4605
4606xnnpack_unit_test(
4607 name = "f32_vrndd_test",
4608 srcs = [
4609 "test/f32-vrndd.cc",
4610 "test/vunary-microkernel-tester.h",
4611 ] + MICROKERNEL_TEST_HDRS,
4612 deps = MICROKERNEL_TEST_DEPS,
4613)
4614
4615xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004616 name = "f32_vscale_test",
4617 srcs = [
4618 "test/f32-vscale.cc",
4619 "test/vscale-microkernel-tester.h",
4620 ] + MICROKERNEL_TEST_HDRS,
4621 deps = MICROKERNEL_TEST_DEPS,
4622)
4623
4624xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004625 name = "f32_vscaleexpminusmax_test",
4626 srcs = [
4627 "test/f32-vscaleexpminusmax.cc",
4628 "test/vscaleexpminusmax-microkernel-tester.h",
4629 ] + MICROKERNEL_TEST_HDRS,
4630 deps = MICROKERNEL_TEST_DEPS,
4631)
4632
4633xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004634 name = "f32_vscaleextexp_test",
4635 srcs = [
4636 "test/f32-vscaleextexp.cc",
4637 "test/vscaleextexp-microkernel-tester.h",
4638 ] + MICROKERNEL_TEST_HDRS,
4639 deps = MICROKERNEL_TEST_DEPS,
4640)
4641
4642xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004643 name = "f32_vsqr_test",
4644 srcs = [
4645 "test/f32-vsqr.cc",
4646 "test/vunary-microkernel-tester.h",
4647 ] + MICROKERNEL_TEST_HDRS,
4648 deps = MICROKERNEL_TEST_DEPS,
4649)
4650
4651xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004652 name = "f32_vsqrdiff_test",
4653 srcs = [
4654 "test/f32-vsqrdiff.cc",
4655 "test/vbinary-microkernel-tester.h",
4656 ] + MICROKERNEL_TEST_HDRS,
4657 deps = MICROKERNEL_TEST_DEPS,
4658)
4659
4660xnnpack_unit_test(
4661 name = "f32_vsqrdiffc_test",
4662 srcs = [
4663 "test/f32-vsqrdiffc.cc",
4664 "test/vbinaryc-microkernel-tester.h",
4665 ] + MICROKERNEL_TEST_HDRS,
4666 deps = MICROKERNEL_TEST_DEPS,
4667)
4668
4669xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004670 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07004671 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004672 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004673 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004674 ] + MICROKERNEL_TEST_HDRS,
4675 deps = MICROKERNEL_TEST_DEPS,
4676)
4677
4678xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004679 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004680 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004681 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004682 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004683 ] + MICROKERNEL_TEST_HDRS,
4684 deps = MICROKERNEL_TEST_DEPS,
4685)
4686
4687xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004688 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004689 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004690 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004691 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07004692 ] + MICROKERNEL_TEST_HDRS,
4693 deps = MICROKERNEL_TEST_DEPS,
4694)
4695
4696xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004697 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004698 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004699 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004700 "test/avgpool-microkernel-tester.h",
4701 "src/xnnpack/AlignedAllocator.h",
4702 ] + MICROKERNEL_TEST_HDRS,
4703 deps = MICROKERNEL_TEST_DEPS,
4704)
4705
4706xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004707 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004708 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004709 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004710 "test/gemm-microkernel-tester.h",
4711 "src/xnnpack/AlignedAllocator.h",
4712 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4713 deps = MICROKERNEL_TEST_DEPS,
4714)
4715
4716xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004717 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004718 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004719 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004720 "test/dwconv-microkernel-tester.h",
4721 "src/xnnpack/AlignedAllocator.h",
4722 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4723 deps = MICROKERNEL_TEST_DEPS,
4724)
4725
4726xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004727 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004728 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004729 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004730 "test/gavgpool-microkernel-tester.h",
4731 "src/xnnpack/AlignedAllocator.h",
4732 ] + MICROKERNEL_TEST_HDRS,
4733 deps = MICROKERNEL_TEST_DEPS,
4734)
4735
4736xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004737 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004738 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004739 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004740 "test/gemm-microkernel-tester.h",
4741 "src/xnnpack/AlignedAllocator.h",
4742 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4743 deps = MICROKERNEL_TEST_DEPS,
4744)
4745
4746xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004747 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004748 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004749 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004750 "test/vadd-microkernel-tester.h",
4751 ] + MICROKERNEL_TEST_HDRS,
4752 deps = MICROKERNEL_TEST_DEPS,
4753)
4754
4755xnnpack_unit_test(
4756 name = "u8_clamp_test",
4757 srcs = [
4758 "test/u8-clamp.cc",
4759 "test/clamp-microkernel-tester.h",
4760 ] + MICROKERNEL_TEST_HDRS,
4761 deps = MICROKERNEL_TEST_DEPS,
4762)
4763
4764xnnpack_unit_test(
4765 name = "u8_lut32norm_test",
4766 srcs = [
4767 "test/u8-lut32norm.cc",
4768 "test/lut-norm-microkernel-tester.h",
4769 ] + MICROKERNEL_TEST_HDRS,
4770 deps = MICROKERNEL_TEST_DEPS,
4771)
4772
4773xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004774 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004775 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004776 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004777 "test/maxpool-microkernel-tester.h",
4778 ] + MICROKERNEL_TEST_HDRS,
4779 deps = MICROKERNEL_TEST_DEPS,
4780)
4781
4782xnnpack_unit_test(
4783 name = "u8_rmax_test",
4784 srcs = [
4785 "test/u8-rmax.cc",
4786 "test/rmax-microkernel-tester.h",
4787 ] + MICROKERNEL_TEST_HDRS,
4788 deps = MICROKERNEL_TEST_DEPS,
4789)
4790
4791xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07004792 name = "x32_fill_test",
4793 srcs = [
4794 "test/x32-fill.cc",
4795 "test/fill-microkernel-tester.h",
4796 ] + MICROKERNEL_TEST_HDRS,
4797 deps = MICROKERNEL_TEST_DEPS,
4798)
4799
4800xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004801 name = "x32_packx_test",
4802 srcs = [
4803 "test/x32-packx.cc",
4804 "test/pack-microkernel-tester.h",
4805 "src/xnnpack/AlignedAllocator.h",
4806 ] + MICROKERNEL_TEST_HDRS,
4807 deps = MICROKERNEL_TEST_DEPS,
4808)
4809
4810xnnpack_unit_test(
4811 name = "x32_pad_test",
4812 srcs = [
4813 "test/x32-pad.cc",
4814 "test/pad-microkernel-tester.h",
4815 ] + MICROKERNEL_TEST_HDRS,
4816 deps = MICROKERNEL_TEST_DEPS,
4817)
4818
4819xnnpack_unit_test(
4820 name = "x32_unpool_test",
4821 srcs = [
4822 "test/x32-unpool.cc",
4823 "test/unpool-microkernel-tester.h",
4824 ] + MICROKERNEL_TEST_HDRS,
4825 deps = MICROKERNEL_TEST_DEPS,
4826)
4827
4828xnnpack_unit_test(
4829 name = "x32_zip_test",
4830 srcs = [
4831 "test/x32-zip.cc",
4832 "test/zip-microkernel-tester.h",
4833 ] + MICROKERNEL_TEST_HDRS,
4834 deps = MICROKERNEL_TEST_DEPS,
4835)
4836
4837xnnpack_unit_test(
4838 name = "x8_lut_test",
4839 srcs = [
4840 "test/x8-lut.cc",
4841 "test/lut-microkernel-tester.h",
4842 ] + MICROKERNEL_TEST_HDRS,
4843 deps = MICROKERNEL_TEST_DEPS,
4844)
4845
4846xnnpack_unit_test(
4847 name = "x8_zip_test",
4848 srcs = [
4849 "test/x8-zip.cc",
4850 "test/zip-microkernel-tester.h",
4851 ] + MICROKERNEL_TEST_HDRS,
4852 deps = MICROKERNEL_TEST_DEPS,
4853)
4854
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004855xnnpack_unit_test(
4856 name = "requantization_test",
4857 srcs = [
4858 "src/xnnpack/requantization-stubs.h",
4859 "test/requantization.cc",
4860 "test/requantization-tester.h",
4861 ] + MICROKERNEL_TEST_HDRS,
4862 deps = MICROKERNEL_TEST_DEPS,
4863)
4864
Marat Dukhan20c3b922020-03-10 03:45:06 -07004865########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004866
4867xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07004868 name = "operator_size_test",
4869 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004870 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004871)
4872
Marat Dukhan20c3b922020-03-10 03:45:06 -07004873xnnpack_binary(
4874 name = "subgraph_size_test",
4875 srcs = ["test/subgraph-size.c"],
4876 deps = [":XNNPACK"],
4877)
4878
4879########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004880
4881xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004882 name = "abs_nc_test",
4883 srcs = [
4884 "test/abs-nc.cc",
4885 "test/abs-operator-tester.h",
4886 ],
4887 deps = OPERATOR_TEST_DEPS,
4888)
4889
4890xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08004891 name = "add_nd_test",
4892 srcs = [
4893 "test/add-nd.cc",
4894 "test/binary-elementwise-operator-tester.h",
4895 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004896 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08004897)
4898
4899xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004900 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004901 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004902 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004903 "test/argmax-pooling-operator-tester.h",
4904 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004905 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004906)
4907
4908xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004909 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004910 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004911 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004912 "test/average-pooling-operator-tester.h",
4913 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004914 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004915)
4916
4917xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07004918 name = "bankers_rounding_nc_test",
4919 srcs = [
4920 "test/bankers-rounding-nc.cc",
4921 "test/bankers-rounding-operator-tester.h",
4922 ],
4923 deps = OPERATOR_TEST_DEPS,
4924)
4925
4926xnnpack_unit_test(
4927 name = "ceiling_nc_test",
4928 srcs = [
4929 "test/ceiling-nc.cc",
4930 "test/ceiling-operator-tester.h",
4931 ],
4932 deps = OPERATOR_TEST_DEPS,
4933)
4934
4935xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004936 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004937 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004938 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004939 "test/channel-shuffle-operator-tester.h",
4940 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004941 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004942)
4943
4944xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004945 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004946 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004947 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004948 "test/clamp-operator-tester.h",
4949 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004950 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004951)
4952
4953xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07004954 name = "constant_pad_nd_test",
4955 srcs = [
4956 "test/constant-pad-nd.cc",
4957 "test/constant-pad-operator-tester.h",
4958 ],
4959 deps = OPERATOR_TEST_DEPS,
4960)
4961
4962xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004963 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004964 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004965 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004966 "test/convolution-operator-tester.h",
4967 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004968 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004969)
4970
4971xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004972 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004973 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004974 "test/convolution-nchw.cc",
4975 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004976 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004977 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004978)
4979
4980xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07004981 name = "copy_nc_test",
4982 srcs = [
4983 "test/copy-nc.cc",
4984 "test/copy-operator-tester.h",
4985 ],
4986 deps = OPERATOR_TEST_DEPS,
4987)
4988
4989xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004990 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004991 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004992 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004993 "test/deconvolution-operator-tester.h",
4994 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004995 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004996)
4997
4998xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08004999 name = "divide_nd_test",
5000 srcs = [
5001 "test/binary-elementwise-operator-tester.h",
5002 "test/divide-nd.cc",
5003 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005004 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005005)
5006
5007xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005008 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005009 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005010 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005011 "test/fully-connected-operator-tester.h",
5012 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005013 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005014)
5015
5016xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005017 name = "floor_nc_test",
5018 srcs = [
5019 "test/floor-nc.cc",
5020 "test/floor-operator-tester.h",
5021 ],
5022 deps = OPERATOR_TEST_DEPS,
5023)
5024
5025xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005026 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005027 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005028 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005029 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005030 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005031 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005032)
5033
5034xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005035 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005036 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005037 "test/global-average-pooling-ncw.cc",
5038 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005039 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005040 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005041)
5042
5043xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005044 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005045 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005046 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005047 "test/hardswish-operator-tester.h",
5048 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005049 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005050)
5051
5052xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005053 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005054 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005055 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005056 "test/leaky-relu-operator-tester.h",
5057 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005058 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005059)
5060
5061xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005062 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005063 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005064 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005065 "test/max-pooling-operator-tester.h",
5066 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005067 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005068)
5069
5070xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005071 name = "maximum_nd_test",
5072 srcs = [
5073 "test/binary-elementwise-operator-tester.h",
5074 "test/maximum-nd.cc",
5075 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005076 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005077)
5078
5079xnnpack_unit_test(
5080 name = "minimum_nd_test",
5081 srcs = [
5082 "test/binary-elementwise-operator-tester.h",
5083 "test/minimum-nd.cc",
5084 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005085 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005086)
5087
5088xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005089 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005090 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005091 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005092 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005093 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005094 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005095)
5096
5097xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005098 name = "negate_nc_test",
5099 srcs = [
5100 "test/negate-nc.cc",
5101 "test/negate-operator-tester.h",
5102 ],
5103 deps = OPERATOR_TEST_DEPS,
5104)
5105
5106xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005107 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005108 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005109 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005110 "test/prelu-operator-tester.h",
5111 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005112 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005113)
5114
5115xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005116 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005117 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005118 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005119 "test/resize-bilinear-operator-tester.h",
5120 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005121 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005122)
5123
5124xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005125 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005126 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005127 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005128 "test/sigmoid-operator-tester.h",
5129 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005130 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005131)
5132
5133xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005134 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005135 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005136 "test/softmax-nc.cc",
5137 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005138 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005139 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005140)
5141
5142xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005143 name = "square_nc_test",
5144 srcs = [
5145 "test/square-nc.cc",
5146 "test/square-operator-tester.h",
5147 ],
5148 deps = OPERATOR_TEST_DEPS,
5149)
5150
5151xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005152 name = "squared_difference_nd_test",
5153 srcs = [
5154 "test/binary-elementwise-operator-tester.h",
5155 "test/squared-difference-nd.cc",
5156 ],
5157 deps = OPERATOR_TEST_DEPS,
5158)
5159
5160xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005161 name = "subtract_nd_test",
5162 srcs = [
5163 "test/binary-elementwise-operator-tester.h",
5164 "test/subtract-nd.cc",
5165 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005166 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005167)
5168
5169xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005170 name = "truncation_nc_test",
5171 srcs = [
5172 "test/truncation-nc.cc",
5173 "test/truncation-operator-tester.h",
5174 ],
5175 deps = OPERATOR_TEST_DEPS,
5176)
5177
5178xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005179 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005180 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005181 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005182 "test/unpooling-operator-tester.h",
5183 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005184 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005185)
5186
Chao Mei6ddfc602020-05-13 22:29:36 -07005187############################### Misc unit tests ###############################
5188
5189xnnpack_unit_test(
5190 name = "memory_planner_test",
5191 srcs = [
5192 "test/memory-planner-test.cc",
5193 ],
5194 deps = [
5195 ":XNNPACK",
5196 ":memory_planner",
5197 ],
5198)
5199
Marat Dukhan08c4a432019-10-03 09:29:21 -07005200############################# Build configurations #############################
5201
Marat Dukhanb8642352019-10-30 15:43:02 -07005202# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005203config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005204 name = "xnn_enable_assembly_explicit_true",
5205 define_values = {"xnn_enable_assembly": "true"},
5206)
5207
5208# Disables usage of assembly kernels.
5209config_setting(
5210 name = "xnn_enable_assembly_explicit_false",
5211 define_values = {"xnn_enable_assembly": "false"},
5212)
5213
Marat Dukhan9de90e02020-06-18 16:04:12 -07005214# Enables usage of sparse inference.
5215config_setting(
5216 name = "xnn_enable_sparse_explicit_true",
5217 define_values = {"xnn_enable_sparse": "true"},
5218)
5219
5220# Disables usage of sparse inference.
5221config_setting(
5222 name = "xnn_enable_sparse_explicit_false",
5223 define_values = {"xnn_enable_sparse": "false"},
5224)
5225
Marat Dukhan05702cf2020-03-26 15:41:33 -07005226# Disables usage of HMP-aware optimizations.
5227config_setting(
5228 name = "xnn_enable_hmp_explicit_false",
5229 define_values = {"xnn_enable_hmp": "false"},
5230)
5231
Chao Mei6ddfc602020-05-13 22:29:36 -07005232# Enable usage of optimized memory allocation
5233config_setting(
5234 name = "xnn_enable_memopt_explicit_true",
5235 define_values = {"xnnpack_enable_memopt": "true"},
5236)
5237
5238# Disable usage of optimized memory allocation
5239config_setting(
5240 name = "xnn_enable_memopt_explicit_false",
5241 define_values = {"xnnpack_enable_memopt": "false"},
5242)
5243
Marat Dukhanb8642352019-10-30 15:43:02 -07005244# Builds with -c dbg
5245config_setting(
5246 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005247 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005248 "compilation_mode": "dbg",
5249 },
5250)
5251
5252# Builds with -c opt
5253config_setting(
5254 name = "optimized_build",
5255 values = {
5256 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005257 },
5258)
5259
5260config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005261 name = "linux_k8",
5262 values = {"cpu": "k8"},
5263)
5264
5265config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005266 name = "linux_arm",
5267 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005268)
5269
5270config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005271 name = "linux_armeabi",
5272 values = {"cpu": "armeabi"},
5273)
5274
5275config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005276 name = "linux_armhf",
5277 values = {"cpu": "armhf"},
5278)
5279
5280config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005281 name = "linux_armv7a",
5282 values = {"cpu": "armv7a"},
5283)
5284
5285config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005286 name = "linux_aarch64",
5287 values = {"cpu": "aarch64"},
5288)
5289
5290config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005291 name = "android",
5292 values = {"crosstool_top": "//external:android/crosstool"},
5293)
5294
5295config_setting(
5296 name = "android_armv7",
5297 values = {
5298 "crosstool_top": "//external:android/crosstool",
5299 "cpu": "armeabi-v7a",
5300 },
5301)
5302
5303config_setting(
5304 name = "android_arm64",
5305 values = {
5306 "crosstool_top": "//external:android/crosstool",
5307 "cpu": "arm64-v8a",
5308 },
5309)
5310
5311config_setting(
5312 name = "android_x86",
5313 values = {
5314 "crosstool_top": "//external:android/crosstool",
5315 "cpu": "x86",
5316 },
5317)
5318
5319config_setting(
5320 name = "android_x86_64",
5321 values = {
5322 "crosstool_top": "//external:android/crosstool",
5323 "cpu": "x86_64",
5324 },
5325)
5326
5327config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005328 name = "windows_x86_64",
5329 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005330)
5331
5332config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005333 name = "windows_x86_64_clang",
5334 values = {
5335 "compiler": "clang-cl",
5336 "cpu": "x64_windows",
5337 },
5338)
5339
5340config_setting(
5341 name = "windows_x86_64_mingw",
5342 values = {
5343 "compiler": "mingw-gcc",
5344 "cpu": "x64_windows",
5345 },
5346)
5347
5348config_setting(
5349 name = "windows_x86_64_msys",
5350 values = {
5351 "compiler": "msys-gcc",
5352 "cpu": "x64_windows",
5353 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005354)
5355
5356config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005357 name = "macos_x86_64",
5358 values = {
5359 "apple_platform_type": "macos",
5360 "cpu": "darwin",
5361 },
5362)
5363
5364config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005365 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005366 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005367)
5368
5369config_setting(
5370 name = "emscripten_wasm",
5371 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005372 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005373 "cpu": "wasm",
5374 },
5375)
5376
5377config_setting(
5378 name = "emscripten_wasmsimd",
5379 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005380 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005381 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005382 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005383 },
5384)
5385
5386config_setting(
5387 name = "emscripten_asmjs",
5388 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005389 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005390 "cpu": "asmjs",
5391 },
5392)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005393
5394config_setting(
5395 name = "ios_armv7",
5396 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005397 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005398 "cpu": "ios_armv7",
5399 },
5400)
5401
5402config_setting(
5403 name = "ios_arm64",
5404 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005405 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005406 "cpu": "ios_arm64",
5407 },
5408)
5409
5410config_setting(
5411 name = "ios_arm64e",
5412 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005413 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005414 "cpu": "ios_arm64e",
5415 },
5416)
5417
5418config_setting(
5419 name = "ios_x86",
5420 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005421 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005422 "cpu": "ios_i386",
5423 },
5424)
5425
5426config_setting(
5427 name = "ios_x86_64",
5428 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005429 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005430 "cpu": "ios_x86_64",
5431 },
5432)
5433
5434config_setting(
5435 name = "watchos_armv7k",
5436 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005437 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005438 "cpu": "watchos_armv7k",
5439 },
5440)
5441
5442config_setting(
5443 name = "watchos_arm64_32",
5444 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005445 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005446 "cpu": "watchos_arm64_32",
5447 },
5448)
5449
5450config_setting(
5451 name = "watchos_x86",
5452 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005453 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005454 "cpu": "watchos_i386",
5455 },
5456)
5457
5458config_setting(
5459 name = "watchos_x86_64",
5460 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005461 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005462 "cpu": "watchos_x86_64",
5463 },
5464)
5465
5466config_setting(
5467 name = "tvos_arm64",
5468 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005469 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005470 "cpu": "tvos_arm64",
5471 },
5472)
5473
5474config_setting(
5475 name = "tvos_x86_64",
5476 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005477 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005478 "cpu": "tvos_x86_64",
5479 },
5480)