blob: 2ee633f4b698db3a1842ccadb59450313a6e98e6 [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",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700262 "src/f32-vbinary/gen/vrsqrdiffc-scalar-x1.c",
263 "src/f32-vbinary/gen/vrsqrdiffc-scalar-x2.c",
264 "src/f32-vbinary/gen/vrsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700265 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
266 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
267 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700268 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
269 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
270 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
271 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
272 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
273 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700274 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
275 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
276 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
277 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
278 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
279 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700280 "src/f32-vlrelu/gen/scalar-x1.c",
281 "src/f32-vlrelu/gen/scalar-x2.c",
282 "src/f32-vlrelu/gen/scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700283 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
284 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
285 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700286 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
287 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
288 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
289 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
290 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
291 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
292 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
293 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
294 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
295 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
296 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
297 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700298 "src/f32-vunary/gen/vabs-scalar-x1.c",
299 "src/f32-vunary/gen/vabs-scalar-x2.c",
300 "src/f32-vunary/gen/vabs-scalar-x4.c",
301 "src/f32-vunary/gen/vneg-scalar-x1.c",
302 "src/f32-vunary/gen/vneg-scalar-x2.c",
303 "src/f32-vunary/gen/vneg-scalar-x4.c",
304 "src/f32-vunary/gen/vsqr-scalar-x1.c",
305 "src/f32-vunary/gen/vsqr-scalar-x2.c",
306 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800307 "src/math/expminus-scalar-lut2048-p1.c",
308 "src/math/expminus-scalar-lut64-p2.c",
309 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700310 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700311 "src/math/roundne-scalar-nearbyint.c",
312 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700313 "src/math/roundd-scalar-addsub.c",
314 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700315 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700316 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700317 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700318 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700319 "src/math/roundz-scalar-addsub.c",
320 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700321 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800322 "src/math/sigmoid-scalar-lut2048-p1-div.c",
323 "src/math/sigmoid-scalar-lut64-p2-div.c",
324 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700325 "src/q8-avgpool/9p8x-minmax-scalar-c1.c",
326 "src/q8-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700327 "src/q8-dwconv/up1x9-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700328 "src/q8-gavgpool/7p7x-minmax-scalar-c1.c",
329 "src/q8-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700330 "src/q8-gemm/2x2-minmax-scalar.c",
331 "src/q8-igemm/2x2-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700332 "src/q8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700333 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700334 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700335 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700336 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700337 "src/x32-fill/scalar-float.c",
338 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700339 "src/x32-packx/x2-scalar.c",
340 "src/x32-packx/x3-scalar.c",
341 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700342 "src/x32-pad/scalar-int.c",
343 "src/x32-pad/scalar-float.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700344 "src/x32-unpool/scalar.c",
345 "src/x32-zip/x2-scalar.c",
346 "src/x32-zip/x3-scalar.c",
347 "src/x32-zip/x4-scalar.c",
348 "src/x32-zip/xm-scalar.c",
349 "src/x8-lut/scalar.c",
350 "src/x8-zip/x2-scalar.c",
351 "src/x8-zip/x3-scalar.c",
352 "src/x8-zip/x4-scalar.c",
353 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700354 "src/requantization/precise-scalar.c",
355 "src/requantization/fp32-scalar.c",
356 "src/requantization/q31-scalar.c",
357 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700358]
359
Marat Dukhan436ebe62019-12-04 15:10:12 -0800360WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700361 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
362 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700363 "src/f32-clamp/gen/wasm-x1.c",
364 "src/f32-clamp/gen/wasm-x2.c",
365 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700366 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
367 "src/f32-dwconv/gen/up1x4-wasm.c",
368 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
369 "src/f32-dwconv/gen/up1x9-wasm.c",
370 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
371 "src/f32-dwconv/gen/up1x25-wasm.c",
372 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
373 "src/f32-dwconv/gen/up2x4-wasm.c",
374 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
375 "src/f32-dwconv/gen/up2x9-wasm.c",
376 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
377 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700378 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
379 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
380 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
381 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700382 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
383 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700384 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
385 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
386 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
387 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700388 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
389 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700390 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
391 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700392 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
393 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
394 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700395 "src/f32-gemm/gen/1x4-wasm.c",
396 "src/f32-gemm/gen/2x4-wasm.c",
397 "src/f32-gemm/gen/4x2-wasm.c",
398 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700399 "src/f32-gemm/gen/1x4-relu-wasm.c",
400 "src/f32-gemm/gen/2x4-relu-wasm.c",
401 "src/f32-gemm/gen/4x2-relu-wasm.c",
402 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700403 "src/f32-gemm/gen/1x4-minmax-wasm.c",
404 "src/f32-gemm/gen/2x4-minmax-wasm.c",
405 "src/f32-gemm/gen/4x2-minmax-wasm.c",
406 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800407 "src/f32-hswish/gen/wasm-x1.c",
408 "src/f32-hswish/gen/wasm-x2.c",
409 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700410 "src/f32-igemm/gen/1x4-wasm.c",
411 "src/f32-igemm/gen/2x4-wasm.c",
412 "src/f32-igemm/gen/4x2-wasm.c",
413 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700414 "src/f32-igemm/gen/1x4-relu-wasm.c",
415 "src/f32-igemm/gen/2x4-relu-wasm.c",
416 "src/f32-igemm/gen/4x2-relu-wasm.c",
417 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700418 "src/f32-igemm/gen/1x4-minmax-wasm.c",
419 "src/f32-igemm/gen/2x4-minmax-wasm.c",
420 "src/f32-igemm/gen/4x2-minmax-wasm.c",
421 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700422 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
423 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
424 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700425 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
426 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
427 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
428 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
429 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
430 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
431 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
432 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
433 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
434 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
435 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
436 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800437 "src/f32-vbinary/gen/vmax-wasm-x1.c",
438 "src/f32-vbinary/gen/vmax-wasm-x2.c",
439 "src/f32-vbinary/gen/vmax-wasm-x4.c",
440 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
441 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
442 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
443 "src/f32-vbinary/gen/vmin-wasm-x1.c",
444 "src/f32-vbinary/gen/vmin-wasm-x2.c",
445 "src/f32-vbinary/gen/vmin-wasm-x4.c",
446 "src/f32-vbinary/gen/vminc-wasm-x1.c",
447 "src/f32-vbinary/gen/vminc-wasm-x2.c",
448 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700449 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
450 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
451 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
452 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
453 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
454 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
455 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
456 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
457 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
458 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
459 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
460 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
461 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
462 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
463 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
464 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
465 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
466 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700467 "src/f32-vlrelu/gen/wasm-x1.c",
468 "src/f32-vlrelu/gen/wasm-x2.c",
469 "src/f32-vlrelu/gen/wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700470 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
471 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
472 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800473]
474
Marat Dukhan290055c2020-06-09 12:24:29 -0700475WASMSIMD_UKERNELS = [
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700476 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
477 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
478 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
479 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
480 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
481 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700482 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
483 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
484 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
485 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
486 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
487 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
488 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
489 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
490 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
491 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
492 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
493 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
494 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
495 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
496 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
497 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
498 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
499 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
500 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
501 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
502 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
503 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
504 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
505 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700506 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
507 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
508 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
509 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat-arm.c",
510 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat-arm.c",
511 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat-x86.c",
512 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700513 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
514 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
515 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
516 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
517 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
518 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
519 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
520 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
521 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
522 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
523 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
524 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
525 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
526 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
527 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
528 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
529 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
530 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
531 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
532 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
533 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
534 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
535 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
536 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
537 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
538 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
539 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
540 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
541 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
542 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
543 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
544 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
545 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
546 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
547 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
548 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
549 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
550 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
551 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
552 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
553 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
554 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
555 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
556 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
557 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
558 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
559 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
560 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
561 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
562 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
563 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
564 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
565 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
566 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
567 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
568 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
569 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
570 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
571 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
572 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700573 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
574 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
575 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
576 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat-arm.c",
577 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat-arm.c",
578 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat-x86.c",
579 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700580 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
581 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
582 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
583 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
584 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
585 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
586 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
587 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
588 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
589 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
590 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
591 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
592 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
593 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
594 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
595 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
596 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
597 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
598 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
599 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
600 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
601 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
602 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
603 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
604 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
605 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
606 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
607 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
608 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
609 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700610 "src/requantization/fp32-wasmsimd.c",
611]
612
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800613PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800614 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800615 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800616 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700617 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
618 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700619 "src/f32-clamp/gen/psimd-x4.c",
620 "src/f32-clamp/gen/psimd-x8.c",
Erich Elsen0a1970e2020-06-10 09:24:59 -0700621 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700622 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700623 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700624 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700625 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700626 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
627 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
628 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
629 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
630 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
631 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
632 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
633 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
634 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
635 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
636 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
637 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Erich Elsen6f278b52020-06-10 16:13:11 -0700638 "src/f32-gavgpool-cw/psimd-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700639 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
640 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700641 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
642 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
643 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
644 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
645 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
646 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
647 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
648 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
649 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
650 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
651 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
652 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
653 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
654 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
655 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
656 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
657 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
658 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
659 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800660 "src/f32-hswish/gen/psimd-x4.c",
661 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700662 "src/f32-ibilinear/gen/psimd-c4.c",
663 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700664 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
665 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
666 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
667 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
668 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
669 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
670 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
671 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
672 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
673 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700674 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700675 "src/f32-spmm/gen/4x1-minmax-psimd.c",
676 "src/f32-spmm/gen/8x1-minmax-psimd.c",
677 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700678 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
679 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700680 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800681 "src/f32-prelu/gen/psimd-2x4.c",
682 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800683 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700684 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
685 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
686 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
687 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
688 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
689 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
690 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
691 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800692 "src/f32-vbinary/gen/vmax-psimd-x4.c",
693 "src/f32-vbinary/gen/vmax-psimd-x8.c",
694 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
695 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
696 "src/f32-vbinary/gen/vmin-psimd-x4.c",
697 "src/f32-vbinary/gen/vmin-psimd-x8.c",
698 "src/f32-vbinary/gen/vminc-psimd-x4.c",
699 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700700 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
701 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
702 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
703 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
704 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
705 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700706 "src/f32-vbinary/gen/vrsqrdiffc-psimd-x4.c",
707 "src/f32-vbinary/gen/vrsqrdiffc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700708 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
709 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700710 "src/f32-vbinary/gen/vsqrdiff-psimd-x4.c",
711 "src/f32-vbinary/gen/vsqrdiff-psimd-x8.c",
712 "src/f32-vbinary/gen/vsqrdiffc-psimd-x4.c",
713 "src/f32-vbinary/gen/vsqrdiffc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700714 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
715 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
716 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
717 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700718 "src/f32-vlrelu/gen/psimd-x4.c",
719 "src/f32-vlrelu/gen/psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700720 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
721 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700722 "src/f32-vunary/gen/vabs-psimd-x4.c",
723 "src/f32-vunary/gen/vabs-psimd-x8.c",
724 "src/f32-vunary/gen/vneg-psimd-x4.c",
725 "src/f32-vunary/gen/vneg-psimd-x8.c",
726 "src/f32-vunary/gen/vsqr-psimd-x4.c",
727 "src/f32-vunary/gen/vsqr-psimd-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700728 "src/x32-fill/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700729 "src/x32-packx/x4-psimd.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700730 "src/x32-pad/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700731 "src/x32-unpool/psimd.c",
732 "src/x32-zip/x2-psimd.c",
733 "src/x32-zip/x3-psimd.c",
734 "src/x32-zip/x4-psimd.c",
735 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700736 "src/requantization/precise-psimd.c",
737 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700738]
739
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800740PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800741 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
742 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
743 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
744 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
745 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
746 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
747 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
748 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
749 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
750 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
751 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
752 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800753 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
754 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
755 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
756 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
757 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
758 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700759 "src/f32-vrnd/gen/vrndne-psimd-x4.c",
760 "src/f32-vrnd/gen/vrndne-psimd-x8.c",
761 "src/f32-vrnd/gen/vrndz-psimd-x4.c",
762 "src/f32-vrnd/gen/vrndz-psimd-x8.c",
763 "src/f32-vrnd/gen/vrndu-psimd-x4.c",
764 "src/f32-vrnd/gen/vrndu-psimd-x8.c",
765 "src/f32-vrnd/gen/vrndd-psimd-x4.c",
766 "src/f32-vrnd/gen/vrndd-psimd-x8.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700767 "src/math/roundne-psimd-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700768 "src/math/roundd-psimd-addsub.c",
769 "src/math/roundu-psimd-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700770 "src/math/roundz-psimd-addsub.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800771 "src/math/sigmoid-psimd-p5-div.c",
772]
773
Marat Dukhan08c4a432019-10-03 09:29:21 -0700774# ISA-specific micro-kernels
775NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700776 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
777 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700778 "src/f32-clamp/gen/neon-x4.c",
779 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700780 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
781 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700782 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
783 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700784 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
785 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700786 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
787 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700788 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
789 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
790 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
791 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700792 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
793 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
794 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
795 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700796 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
797 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
798 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
799 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700800 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700801 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
802 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700803 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
804 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
805 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
806 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
807 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
808 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
809 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
810 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
811 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
812 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
813 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
814 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
815 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
816 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
817 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
818 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
819 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
820 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
821 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
822 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
823 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
824 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
825 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
826 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
827 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
828 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
829 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
830 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
831 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
832 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
833 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800834 "src/f32-hswish/gen/neon-x4.c",
835 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700836 "src/f32-ibilinear/gen/neon-c4.c",
837 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700838 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
839 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
840 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
841 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
842 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
843 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
844 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
845 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
846 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
847 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
848 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
849 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
850 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
851 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
852 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
853 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700854 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
855 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
856 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700857 "src/f32-ppmm/gen/4x8-minmax-neon.c",
858 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800859 "src/f32-prelu/gen/neon-2x4.c",
860 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800861 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
862 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
863 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
864 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
865 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
866 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
867 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
868 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
869 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
870 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
871 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
872 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
873 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
874 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
875 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
876 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
877 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
878 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
879 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
880 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
881 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
882 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
883 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
884 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700885 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800886 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800887 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
888 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
889 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
890 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
891 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
892 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
893 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
894 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
895 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
896 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
897 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
898 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
899 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
900 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
901 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
902 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
903 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
904 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700905 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
906 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
907 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
908 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800909 "src/f32-vbinary/gen/vmax-neon-x4.c",
910 "src/f32-vbinary/gen/vmax-neon-x8.c",
911 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
912 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
913 "src/f32-vbinary/gen/vmin-neon-x4.c",
914 "src/f32-vbinary/gen/vmin-neon-x8.c",
915 "src/f32-vbinary/gen/vminc-neon-x4.c",
916 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700917 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
918 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
919 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
920 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700921 "src/f32-vbinary/gen/vrsqrdiffc-neon-x4.c",
922 "src/f32-vbinary/gen/vrsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700923 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
924 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700925 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
926 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
927 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
928 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700929 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
930 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
931 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
932 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700933 "src/f32-vlrelu/gen/neon-x4.c",
934 "src/f32-vlrelu/gen/neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700935 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
936 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700937 "src/f32-vrnd/gen/vrndne-neon-x4.c",
938 "src/f32-vrnd/gen/vrndne-neon-x8.c",
939 "src/f32-vrnd/gen/vrndz-neon-x4.c",
940 "src/f32-vrnd/gen/vrndz-neon-x8.c",
941 "src/f32-vrnd/gen/vrndu-neon-x4.c",
942 "src/f32-vrnd/gen/vrndu-neon-x8.c",
943 "src/f32-vrnd/gen/vrndd-neon-x4.c",
944 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700945 "src/f32-vunary/gen/vabs-neon-x4.c",
946 "src/f32-vunary/gen/vabs-neon-x8.c",
947 "src/f32-vunary/gen/vneg-neon-x4.c",
948 "src/f32-vunary/gen/vneg-neon-x8.c",
949 "src/f32-vunary/gen/vsqr-neon-x4.c",
950 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700951 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
952 "src/q8-avgpool/9x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700953 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700954 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
955 "src/q8-gavgpool/7x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700956 "src/q8-gemm/4x8-minmax-neon.c",
957 "src/q8-gemm/8x8-minmax-neon.c",
958 "src/q8-igemm/4x8-minmax-neon.c",
959 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700960 "src/q8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700961 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700962 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700963 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700964 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700965 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700966 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -0700967 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700968 "src/x32-zip/x2-neon.c",
969 "src/x32-zip/x3-neon.c",
970 "src/x32-zip/x4-neon.c",
971 "src/x32-zip/xm-neon.c",
972 "src/x8-zip/x2-neon.c",
973 "src/x8-zip/x3-neon.c",
974 "src/x8-zip/x4-neon.c",
975 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700976 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700977 "src/math/roundd-neon-addsub.c",
978 "src/math/roundd-neon-cvt.c",
979 "src/math/roundu-neon-addsub.c",
980 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700981 "src/math/roundz-neon-addsub.c",
982 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -0800983 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -0800984 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
985 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
986 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
987 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
988 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
989 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700990 "src/requantization/precise-neon.c",
991 "src/requantization/fp32-neon.c",
992 "src/requantization/q31-neon.c",
993 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700994]
995
996NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -0700997 "src/f32-ibilinear/gen/neonfma-c4.c",
998 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700999 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1000 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1001 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1002 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1003 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1004 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1005 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1006 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1007 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001008 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1009 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1010 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1011 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001012 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1013 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1014 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1015 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001016 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1017 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1018 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1019 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001020 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1021 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1022 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1023 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1024 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1025 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1026 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1027 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1028 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1029 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1030 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1031 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1032 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1033 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1034 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1035 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1036 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1037 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001038 "src/f32-hswish/gen/neonfma-x4.c",
1039 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001040 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1041 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001042 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1043 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1044 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1045 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1046 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1047 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1048 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1049 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1050 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1051 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1052 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1053 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1054 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1055 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1056 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1057 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1058 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1059 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1060 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1061 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1062 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1063 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1064 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1065 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001066 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1067 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1068 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1069 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1070 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1071 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1072 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1073 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1074 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1075 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1076 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1077 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1078 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1079 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1080 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1081 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1082 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1083 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1084 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1085 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1086 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1087 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1088 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1089 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1090 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1091 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1092 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1093 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1094 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1095 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1096 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1097 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1098 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1099 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1100 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1101 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1102 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1103 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1104 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1105 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1106 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1107 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1108 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1109 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1110 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1111 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1112 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1113 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1114 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1115 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1116 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1117 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1118 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1119 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001120 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1121 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001122 "src/math/exp-neonfma-lut64-p2.c",
1123 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001124 "src/math/expminus-neonfma-lut2048-p1.c",
1125 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001126 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001127 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1128 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1129 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1130 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1131 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1132 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1133 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1134 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1135 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1136 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1137 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1138 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1139 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1140 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1141 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1142 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1143 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1144 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001145]
1146
1147AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001148 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1149 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1150 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1151 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1152 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1153 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001154 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1155 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1156 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1157 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1158 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1159 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1160 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1161 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1162 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1163 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1164 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1165 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1166 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1167 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1168 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1169 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1170 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1171 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1172 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1173 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001174 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1175 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001176 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1177 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001178 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1179 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001180 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1181 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001182 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1183 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1184 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1185 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1186 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001187 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1188 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1189 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1190 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1191 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1192 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1193 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1194 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1195 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1196 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1197 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1198 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1199 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1200 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1201 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1202 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1203 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1204 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001205 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1206 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1207 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1208 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1209 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1210 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1211 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1212 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1213 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1214 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1215 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1216 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1217 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1218 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1219 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1220 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1221 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1222 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001223 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1224 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1225 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1226 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1227 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1228 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001229]
1230
Marat Dukhan8853b822020-05-07 12:19:01 -07001231NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001232 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1233 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1234 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1235 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1236 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1237 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1238 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1239 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001240 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001241 "src/math/roundd-neonv8.c",
1242 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001243 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001244]
1245
Marat Dukhan08c4a432019-10-03 09:29:21 -07001246AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001247 "src/f16-clamp/gen/neonfp16arith-x16.c",
1248 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001249 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1250 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1251 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1252 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1253 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1254 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1255 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1256 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1257 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1258 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1259 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1260 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001261 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1262 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001263 "src/f16-hswish/gen/neonfp16arith-x16.c",
1264 "src/f16-hswish/gen/neonfp16arith-x8.c",
1265 "src/f16-prelu/gen/neonfp16arith-2x16.c",
1266 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001267 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1268 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1269 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1270 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1271 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1272 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1273 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1274 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1275 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1276 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1277 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1278 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1279 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1280 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1281 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1282 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1283 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1284 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1285 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1286 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1287 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1288 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1289 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1290 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1291 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1292 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1293 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1294 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001295 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1296 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001297 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1298 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1299 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1300 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1301 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1302 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1303 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1304 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001305 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1306 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1307 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1308 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001309 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1310 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1311 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1312 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1313 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1314 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1315 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1316 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1317 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1318 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1319 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1320 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001321 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1322 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1323 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1324 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1325 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1326 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1327 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1328 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001329]
1330
1331SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001332 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1333 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001334 "src/f32-clamp/gen/sse-x4.c",
1335 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001336 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1337 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001338 "src/f32-dwconv-chw/3x3p1-sse.c",
1339 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001340 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1341 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1342 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1343 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1344 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1345 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1346 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1347 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1348 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1349 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1350 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1351 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001352 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001353 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1354 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001355 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1356 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1357 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1358 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1359 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1360 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1361 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1362 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1363 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1364 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1365 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1366 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1367 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001368 "src/f32-hswish/gen/sse-x4.c",
1369 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001370 "src/f32-ibilinear/gen/sse-c4.c",
1371 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001372 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1373 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1374 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1375 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1376 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1377 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1378 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001379 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1380 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1381 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001382 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001383 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001384 "src/f32-spmm/gen/4x1-minmax-sse.c",
1385 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001386 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001387 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1388 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1389 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1390 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1391 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1392 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1393 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1394 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001395 "src/f32-vbinary/gen/vmax-sse-x4.c",
1396 "src/f32-vbinary/gen/vmax-sse-x8.c",
1397 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1398 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1399 "src/f32-vbinary/gen/vmin-sse-x4.c",
1400 "src/f32-vbinary/gen/vmin-sse-x8.c",
1401 "src/f32-vbinary/gen/vminc-sse-x4.c",
1402 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001403 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1404 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1405 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1406 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1407 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1408 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001409 "src/f32-vbinary/gen/vrsqrdiffc-sse-x4.c",
1410 "src/f32-vbinary/gen/vrsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001411 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1412 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001413 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1414 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1415 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1416 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001417 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1418 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1419 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1420 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001421 "src/f32-vlrelu/gen/sse-x4.c",
1422 "src/f32-vlrelu/gen/sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001423 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1424 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001425 "src/f32-vunary/gen/vabs-sse-x4.c",
1426 "src/f32-vunary/gen/vabs-sse-x8.c",
1427 "src/f32-vunary/gen/vneg-sse-x4.c",
1428 "src/f32-vunary/gen/vneg-sse-x8.c",
1429 "src/f32-vunary/gen/vsqr-sse-x4.c",
1430 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001431 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001432 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001433 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001434 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001435 "src/math/roundd-sse-addsub.c",
1436 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001437 "src/math/roundz-sse-addsub.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001438]
1439
1440SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001441 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1442 "src/f32-argmaxpool/4x-sse2-c4.c",
1443 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001444 "src/f32-prelu/gen/sse2-2x4.c",
1445 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001446 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1447 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1448 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1449 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1450 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1451 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1452 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1453 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1454 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1455 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1456 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1457 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001458 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001459 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001460 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001461 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001462 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1463 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001464 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1465 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1466 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1467 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1468 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1469 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1470 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1471 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001472 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1473 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001474 "src/q8-igemm/4x4c2-minmax-sse2.c",
1475 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001476 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1477 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001478 "src/q8-gemm/2x4c8-minmax-sse2.c",
1479 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001480 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001481 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001482 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001483 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001484 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001485 "src/x32-zip/x2-sse2.c",
1486 "src/x32-zip/x3-sse2.c",
1487 "src/x32-zip/x4-sse2.c",
1488 "src/x32-zip/xm-sse2.c",
1489 "src/x8-zip/x2-sse2.c",
1490 "src/x8-zip/x3-sse2.c",
1491 "src/x8-zip/x4-sse2.c",
1492 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001493 "src/math/exp-sse2-p5.c",
1494 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001495 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001496 "src/math/roundd-sse2-cvt.c",
1497 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001498 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001499 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001500 "src/requantization/precise-sse2.c",
1501 "src/requantization/fp32-sse2.c",
1502 "src/requantization/q31-sse2.c",
1503 "src/requantization/gemmlowp-sse2.c",
1504]
1505
1506SSSE3_UKERNELS = [
1507 "src/requantization/precise-ssse3.c",
1508 "src/requantization/q31-ssse3.c",
1509 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001510]
1511
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001512SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001513 "src/f32-prelu/gen/sse41-2x4.c",
1514 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001515 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001516 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001517 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001518 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001519 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1520 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001521 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1522 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1523 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1524 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1525 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1526 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1527 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1528 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001529 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001530 "src/math/roundd-sse41.c",
1531 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001532 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001533 "src/requantization/precise-sse4.c",
1534 "src/requantization/q31-sse4.c",
1535 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001536]
1537
Marat Dukhan08c4a432019-10-03 09:29:21 -07001538AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001539 "src/f32-clamp/gen/avx-x8.c",
1540 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001541 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1542 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1543 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1544 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1545 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1546 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1547 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1548 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1549 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1550 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1551 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1552 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1553 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1554 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1555 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1556 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1557 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1558 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1559 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1560 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1561 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1562 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1563 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1564 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1565 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1566 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1567 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1568 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1569 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1570 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001571 "src/f32-hswish/gen/avx-x8.c",
1572 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001573 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1574 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1575 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1576 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1577 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1578 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1579 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1580 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1581 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001582 "src/f32-prelu/gen/avx-2x8.c",
1583 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001584 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001585 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1586 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1587 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1588 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1589 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1590 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1591 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1592 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001593 "src/f32-vbinary/gen/vmax-avx-x8.c",
1594 "src/f32-vbinary/gen/vmax-avx-x16.c",
1595 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1596 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1597 "src/f32-vbinary/gen/vmin-avx-x8.c",
1598 "src/f32-vbinary/gen/vmin-avx-x16.c",
1599 "src/f32-vbinary/gen/vminc-avx-x8.c",
1600 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001601 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1602 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1603 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1604 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1605 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1606 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001607 "src/f32-vbinary/gen/vrsqrdiffc-avx-x8.c",
1608 "src/f32-vbinary/gen/vrsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001609 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1610 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001611 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1612 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1613 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1614 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001615 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1616 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1617 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1618 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001619 "src/f32-vlrelu/gen/avx-x8.c",
1620 "src/f32-vlrelu/gen/avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001621 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001622 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1623 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1624 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1625 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1626 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1627 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1628 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1629 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001630 "src/f32-vunary/gen/vabs-avx-x8.c",
1631 "src/f32-vunary/gen/vabs-avx-x16.c",
1632 "src/f32-vunary/gen/vneg-avx-x8.c",
1633 "src/f32-vunary/gen/vneg-avx-x16.c",
1634 "src/f32-vunary/gen/vsqr-avx-x8.c",
1635 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001636]
1637
Marat Dukhanfda12b82019-11-21 12:27:59 -08001638FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001639 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1640 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1641 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1642 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1643 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1644 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1645 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1646 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1647 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1648 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1649 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1650 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1651 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1652 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1653 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1654 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1655 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1656 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1657 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1658 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1659 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1660 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1661 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1662 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1663 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1664 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1665 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1666 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1667 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1668 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1669 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1670 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1671 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1672 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1673 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1674 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1675 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1676 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1677 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1678 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001679 "src/f32-hswish/gen/fma3-x8.c",
1680 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001681 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1682 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1683 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1684 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1685 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1686 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1687 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1688 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1689 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1690 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1691 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1692 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1693 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1694 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001695]
1696
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001697AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001698 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1699 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1700 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1701 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1702 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1703 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1704 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1705 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1706 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1707 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1708 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1709 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1710 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1711 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1712 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1713 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1714 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1715 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1716 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1717 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1718 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1719 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1720 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1721 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1722 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1723 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1724 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1725 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1726 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1727 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1728 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1729 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1730 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1731 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1732 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1733 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001734 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1735 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1736 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1737 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1738 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1739 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1740 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1741 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1742 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1743 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1744 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1745 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1746 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1747 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1748 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1749 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1750 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1751 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1752 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1753 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1754 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1755 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1756 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1757 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1758 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1759 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1760 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1761 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1762 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1763 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001764 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1765 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1766 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1767 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1768 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1769 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1770 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1771 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1772 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1773 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1774 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1775 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1776 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1777 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1778 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1779 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1780 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1781 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1782 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1783 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1784 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1785 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1786 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1787 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001788 "src/math/exp-avx2-p5.c",
1789 "src/math/exp-avx2-perm-p3.c",
1790 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001791 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001792 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001793 "src/math/sigmoid-avx2-rr2-p5-div.c",
1794 "src/math/sigmoid-avx2-rr1-p5-div.c",
1795 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1796 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1797 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1798 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001799]
1800
Marat Dukhan08c4a432019-10-03 09:29:21 -07001801AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001802 "src/f32-clamp/gen/avx512f-x16.c",
1803 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001804 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1805 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1806 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1807 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1808 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1809 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1810 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1811 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1812 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1813 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1814 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1815 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1816 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1817 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1818 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1819 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1820 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1821 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1822 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1823 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1824 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1825 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1826 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1827 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001828 "src/f32-hswish/gen/avx512f-x16.c",
1829 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001830 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1831 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1832 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1833 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1834 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1835 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001836 "src/f32-prelu/gen/avx512f-2x16.c",
1837 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001838 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1839 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1840 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1841 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
1842 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1843 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
1844 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1845 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1846 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
1847 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1848 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1849 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
1850 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
1851 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
1852 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
1853 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
1854 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
1855 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
1856 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
1857 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
1858 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
1859 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
1860 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
1861 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
1862 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
1863 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1864 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1865 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
1866 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
1867 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
1868 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
1869 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
1870 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
1871 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
1872 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
1873 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001874 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001875 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
1876 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
1877 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
1878 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
1879 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
1880 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
1881 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
1882 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001883 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
1884 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
1885 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
1886 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
1887 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
1888 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
1889 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
1890 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001891 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
1892 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
1893 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
1894 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
1895 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
1896 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001897 "src/f32-vbinary/gen/vrsqrdiffc-avx512f-x16.c",
1898 "src/f32-vbinary/gen/vrsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001899 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
1900 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001901 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
1902 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
1903 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
1904 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001905 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
1906 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
1907 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
1908 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001909 "src/f32-vlrelu/gen/avx512f-x16.c",
1910 "src/f32-vlrelu/gen/avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001911 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001912 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
1913 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
1914 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
1915 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
1916 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
1917 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
1918 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
1919 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
1920 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
1921 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
1922 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
1923 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
1924 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
1925 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
1926 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
1927 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
1928 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
1929 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
1930 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
1931 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
1932 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
1933 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
1934 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
1935 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001936 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
1937 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
1938 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
1939 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
1940 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
1941 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
1942 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
1943 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001944 "src/f32-vunary/gen/vabs-avx512f-x16.c",
1945 "src/f32-vunary/gen/vabs-avx512f-x32.c",
1946 "src/f32-vunary/gen/vneg-avx512f-x16.c",
1947 "src/f32-vunary/gen/vneg-avx512f-x32.c",
1948 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
1949 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001950 "src/math/exp-avx512f-p5-scalef.c",
1951 "src/math/exp-avx512f-p5.c",
1952 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07001953 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001954 "src/math/extexp-avx512f-p5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001955]
1956
1957AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07001958 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07001959 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001960 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1961 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
1962 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1963 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07001964 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
1965 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S",
1966 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
1967 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001968 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
1969 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
1970 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
1971 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001972]
1973
1974AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07001975 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001976 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
1977 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07001978 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07001979 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
1980 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001981 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
1982 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
1983 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001984 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07001985 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1986 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
1987 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07001988 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001989 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
1990 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07001991 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07001992 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
1993 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
1994 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
1995 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
1996 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
1997 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
1998 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
1999 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2000 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2001 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2002 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2003 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2004 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2005 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2006 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2007 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2008 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2009 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002010 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2011 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002012 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002013 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2014 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2015 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2016 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2017 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2018 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2019 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2020 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2021 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2022 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2023 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2024 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2025 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2026 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2027 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2028 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2029 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2030 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002031 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2032 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2033 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2034 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2035 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2036 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2037 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2038 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2039 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2040 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2041 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2042 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2043 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2044 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2045 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2046 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2047 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2048 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002049]
2050
Marat Dukhan1b354632020-03-23 12:50:22 -07002051INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002052 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002053 "src/xnnpack/argmaxpool.h",
2054 "src/xnnpack/avgpool.h",
2055 "src/xnnpack/clamp.h",
2056 "src/xnnpack/common.h",
2057 "src/xnnpack/conv.h",
2058 "src/xnnpack/dwconv.h",
2059 "src/xnnpack/gavgpool.h",
2060 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002061 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002062 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002063 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002064 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002065 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002066 "src/xnnpack/lut.h",
2067 "src/xnnpack/math.h",
2068 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002069 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002070 "src/xnnpack/packx.h",
2071 "src/xnnpack/pad.h",
2072 "src/xnnpack/params.h",
2073 "src/xnnpack/pavgpool.h",
2074 "src/xnnpack/ppmm.h",
2075 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002076 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002077 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002078 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002079 "src/xnnpack/rmax.h",
2080 "src/xnnpack/scalar-utils.h",
2081 "src/xnnpack/spmm.h",
2082 "src/xnnpack/unpool.h",
2083 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002084 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002085 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002086 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002087 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002088 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002089 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002090 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002091]
2092
2093INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002094 "include/xnnpack.h",
2095 "src/xnnpack/allocator.h",
2096 "src/xnnpack/compute.h",
2097 "src/xnnpack/im2col.h",
2098 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002099 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002100 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002101 "src/xnnpack/operator.h",
2102 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002103 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002104 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002105 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002106 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002107]
2108
Marat Dukhan1b354632020-03-23 12:50:22 -07002109ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002110 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002111]
2112
Marat Dukhan1b354632020-03-23 12:50:22 -07002113MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002114 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002115 "include/xnnpack.h",
2116]
2117
Marat Dukhan1b354632020-03-23 12:50:22 -07002118MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002119 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002120 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002121 "src/xnnpack/requantization.h",
2122 "include/xnnpack.h",
2123]
2124
2125OPERATOR_TEST_PARAMS_HDRS = [
2126 "src/xnnpack/params.h",
2127 "src/xnnpack/common.h",
2128]
2129
2130WEIGHTS_PACK_HDRS = [
2131 "src/xnnpack/pack.h",
2132 "src/xnnpack/operator.h",
2133 "src/xnnpack/compute.h",
2134]
2135
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002136LOGGING_COPTS = select({
2137 # No logging in optimized mode
2138 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2139 # Full logging in debug mode
2140 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2141 # Error-only logging in default (fastbuild) mode
2142 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2143})
2144
Marat Dukhan3b59de22020-06-03 20:15:19 -07002145LOGGING_SRCS = select({
2146 # No logging in optimized mode
2147 ":optimized_build": [],
2148 "//conditions:default": [
2149 "src/operator-strings.c",
2150 "src/subgraph-strings.c",
2151 ],
2152})
2153
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002154LOGGING_HDRS = [
2155 "src/xnnpack/log.h",
2156]
2157
Marat Dukhan08c4a432019-10-03 09:29:21 -07002158xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002159 name = "tables",
2160 srcs = TABLE_SRCS,
2161 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002162 gcc_copts = xnnpack_gcc_std_copts(),
2163 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002164)
2165
2166xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002167 name = "scalar_ukernels",
2168 srcs = SCALAR_UKERNELS,
2169 hdrs = INTERNAL_HDRS,
2170 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002171 gcc_copts = xnnpack_gcc_std_copts(),
2172 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002173 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002174 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002175 "@FP16",
2176 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002177 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002178 ],
2179)
2180
2181xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002182 name = "scalar_ukernels_test_mode",
2183 srcs = SCALAR_UKERNELS,
2184 hdrs = INTERNAL_HDRS,
2185 aarch32_copts = ["-marm"],
2186 copts = [
2187 "-UNDEBUG",
2188 "-DXNN_TEST_MODE=1",
2189 ],
2190 gcc_copts = xnnpack_gcc_std_copts(),
2191 msvc_copts = xnnpack_msvc_std_copts(),
2192 deps = [
2193 ":tables",
2194 "@FP16",
2195 "@FXdiv",
2196 "@pthreadpool",
2197 ],
2198)
2199
2200xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002201 name = "wasm_ukernels",
2202 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002203 gcc_copts = xnnpack_gcc_std_copts(),
2204 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002205 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1bbf96b2020-06-15 23:01:20 -07002206 wasmsimd_copts = [
2207 "-msimd128",
2208 "-munimplemented-simd128",
2209 ],
Marat Dukhan290055c2020-06-09 12:24:29 -07002210 wasmsimd_srcs = WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002211 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002212 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002213 "@FP16",
2214 "@FXdiv",
2215 "@pthreadpool",
2216 ],
2217)
2218
2219xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002220 name = "wasm_ukernels_test_mode",
2221 hdrs = INTERNAL_HDRS,
2222 copts = [
2223 "-UNDEBUG",
2224 "-DXNN_TEST_MODE=1",
2225 ],
2226 gcc_copts = xnnpack_gcc_std_copts(),
2227 msvc_copts = xnnpack_msvc_std_copts(),
2228 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1bbf96b2020-06-15 23:01:20 -07002229 wasmsimd_copts = [
2230 "-msimd128",
2231 "-munimplemented-simd128",
2232 ],
Marat Dukhan290055c2020-06-09 12:24:29 -07002233 wasmsimd_srcs = WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002234 deps = [
2235 ":tables",
2236 "@FP16",
2237 "@FXdiv",
2238 "@pthreadpool",
2239 ],
2240)
2241
2242xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002243 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002244 hdrs = INTERNAL_HDRS,
2245 aarch32_copts = [
2246 "-marm",
2247 "-mfpu=neon",
2248 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002249 gcc_copts = xnnpack_gcc_std_copts(),
2250 gcc_x86_copts = ["-msse2"],
2251 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002252 optimized_copts = [
2253 "-O3",
2254 "-ffast-math",
2255 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002256 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002257 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002258 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002259 "@FP16",
2260 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002261 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002262 ],
2263)
2264
2265xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002266 name = "psimd_fastmath_ukernels_test_mode",
2267 hdrs = INTERNAL_HDRS,
2268 aarch32_copts = [
2269 "-marm",
2270 "-mfpu=neon",
2271 ],
2272 copts = [
2273 "-UNDEBUG",
2274 "-DXNN_TEST_MODE=1",
2275 ],
2276 gcc_copts = xnnpack_gcc_std_copts(),
2277 gcc_x86_copts = ["-msse2"],
2278 msvc_copts = xnnpack_msvc_std_copts(),
2279 optimized_copts = [
2280 "-O3",
2281 "-ffast-math",
2282 ],
2283 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2284 deps = [
2285 ":tables",
2286 "@FP16",
2287 "@psimd",
2288 "@pthreadpool",
2289 ],
2290)
2291
2292xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002293 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002294 hdrs = INTERNAL_HDRS,
2295 aarch32_copts = [
2296 "-marm",
2297 "-mfpu=neon",
2298 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002299 gcc_copts = xnnpack_gcc_std_copts(),
2300 gcc_x86_copts = ["-msse2"],
2301 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002302 optimized_copts = [
2303 "-O3",
2304 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002305 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002306 deps = [
2307 ":tables",
2308 "@FP16",
2309 "@psimd",
2310 "@pthreadpool",
2311 ],
2312)
2313
2314xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002315 name = "psimd_accmath_ukernels_test_mode",
2316 hdrs = INTERNAL_HDRS,
2317 aarch32_copts = [
2318 "-marm",
2319 "-mfpu=neon",
2320 ],
2321 copts = [
2322 "-UNDEBUG",
2323 "-DXNN_TEST_MODE=1",
2324 ],
2325 gcc_copts = xnnpack_gcc_std_copts(),
2326 gcc_x86_copts = ["-msse2"],
2327 msvc_copts = xnnpack_msvc_std_copts(),
2328 optimized_copts = [
2329 "-O3",
2330 ],
2331 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2332 deps = [
2333 ":tables",
2334 "@FP16",
2335 "@psimd",
2336 "@pthreadpool",
2337 ],
2338)
2339
2340xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002341 name = "neon_ukernels",
2342 hdrs = INTERNAL_HDRS,
2343 aarch32_copts = [
2344 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002345 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002346 "-mfpu=neon",
2347 ],
2348 aarch32_srcs = NEON_UKERNELS,
2349 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002350 gcc_copts = xnnpack_gcc_std_copts(),
2351 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002352 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002353 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002354 "@FP16",
2355 "@pthreadpool",
2356 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002357)
2358
2359xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002360 name = "neon_ukernels_test_mode",
2361 hdrs = INTERNAL_HDRS,
2362 aarch32_copts = [
2363 "-marm",
2364 "-march=armv7-a",
2365 "-mfpu=neon",
2366 ],
2367 aarch32_srcs = NEON_UKERNELS,
2368 aarch64_srcs = NEON_UKERNELS,
2369 copts = [
2370 "-UNDEBUG",
2371 "-DXNN_TEST_MODE=1",
2372 ],
2373 gcc_copts = xnnpack_gcc_std_copts(),
2374 msvc_copts = xnnpack_msvc_std_copts(),
2375 deps = [
2376 ":tables",
2377 "@FP16",
2378 "@pthreadpool",
2379 ],
2380)
2381
2382xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002383 name = "neonfma_ukernels",
2384 hdrs = INTERNAL_HDRS,
2385 aarch32_copts = [
2386 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002387 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002388 "-mfpu=neon-vfpv4",
2389 ],
2390 aarch32_srcs = NEONFMA_UKERNELS,
2391 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002392 apple_aarch32_copts = [
2393 "-mcpu=swift",
2394 "-mtune=generic",
2395 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002396 gcc_copts = xnnpack_gcc_std_copts(),
2397 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002398 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002399 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002400 "@FP16",
2401 "@pthreadpool",
2402 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002403)
2404
2405xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002406 name = "neonfma_ukernels_test_mode",
2407 hdrs = INTERNAL_HDRS,
2408 aarch32_copts = [
2409 "-marm",
2410 "-march=armv7-a",
2411 "-mfpu=neon-vfpv4",
2412 ],
2413 aarch32_srcs = NEONFMA_UKERNELS,
2414 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002415 apple_aarch32_copts = [
2416 "-mcpu=swift",
2417 "-mtune=generic",
2418 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002419 copts = [
2420 "-UNDEBUG",
2421 "-DXNN_TEST_MODE=1",
2422 ],
2423 gcc_copts = xnnpack_gcc_std_copts(),
2424 msvc_copts = xnnpack_msvc_std_copts(),
2425 deps = [
2426 ":tables",
2427 "@FP16",
2428 "@pthreadpool",
2429 ],
2430)
2431
2432xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002433 name = "neonv8_ukernels",
2434 hdrs = INTERNAL_HDRS,
2435 aarch32_copts = [
2436 "-marm",
2437 "-march=armv8-a",
2438 "-mfpu=neon-fp-armv8",
2439 ],
2440 aarch32_srcs = NEONV8_UKERNELS,
2441 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002442 apple_aarch32_copts = [
2443 "-mcpu=cyclone",
2444 "-mtune=generic",
2445 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002446 gcc_copts = xnnpack_gcc_std_copts(),
2447 msvc_copts = xnnpack_msvc_std_copts(),
2448 deps = [
2449 ":tables",
2450 "@FP16",
2451 "@pthreadpool",
2452 ],
2453)
2454
2455xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002456 name = "neonv8_ukernels_test_mode",
2457 hdrs = INTERNAL_HDRS,
2458 aarch32_copts = [
2459 "-marm",
2460 "-march=armv8-a",
2461 "-mfpu=neon-fp-armv8",
2462 ],
2463 aarch32_srcs = NEONV8_UKERNELS,
2464 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002465 apple_aarch32_copts = [
2466 "-mcpu=cyclone",
2467 "-mtune=generic",
2468 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002469 copts = [
2470 "-UNDEBUG",
2471 "-DXNN_TEST_MODE=1",
2472 ],
2473 gcc_copts = xnnpack_gcc_std_copts(),
2474 msvc_copts = xnnpack_msvc_std_copts(),
2475 deps = [
2476 ":tables",
2477 "@FP16",
2478 "@pthreadpool",
2479 ],
2480)
2481
2482xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002483 name = "neonfp16arith_ukernels",
2484 hdrs = INTERNAL_HDRS,
2485 aarch64_copts = ["-march=armv8.2-a+fp16"],
2486 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002487 gcc_copts = xnnpack_gcc_std_copts(),
2488 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002489 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002490 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002491 "@FP16",
2492 "@pthreadpool",
2493 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002494)
2495
2496xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002497 name = "neonfp16arith_ukernels_test_mode",
2498 hdrs = INTERNAL_HDRS,
2499 aarch64_copts = ["-march=armv8.2-a+fp16"],
2500 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2501 copts = [
2502 "-UNDEBUG",
2503 "-DXNN_TEST_MODE=1",
2504 ],
2505 gcc_copts = xnnpack_gcc_std_copts(),
2506 msvc_copts = xnnpack_msvc_std_copts(),
2507 deps = [
2508 ":tables",
2509 "@FP16",
2510 "@pthreadpool",
2511 ],
2512)
2513
2514xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002515 name = "sse2_ukernels",
2516 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002517 gcc_copts = xnnpack_gcc_std_copts(),
2518 gcc_x86_copts = ["-msse2"],
2519 msvc_copts = xnnpack_msvc_std_copts(),
2520 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002521 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002522 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002523 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002524 "@FP16",
2525 "@pthreadpool",
2526 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002527)
2528
2529xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002530 name = "sse2_ukernels_test_mode",
2531 hdrs = INTERNAL_HDRS,
2532 copts = [
2533 "-UNDEBUG",
2534 "-DXNN_TEST_MODE=1",
2535 ],
2536 gcc_copts = xnnpack_gcc_std_copts(),
2537 gcc_x86_copts = ["-msse2"],
2538 msvc_copts = xnnpack_msvc_std_copts(),
2539 msvc_x86_32_copts = ["/arch:SSE2"],
2540 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2541 deps = [
2542 ":tables",
2543 "@FP16",
2544 "@pthreadpool",
2545 ],
2546)
2547
2548xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002549 name = "ssse3_ukernels",
2550 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002551 gcc_copts = xnnpack_gcc_std_copts(),
2552 gcc_x86_copts = ["-mssse3"],
2553 msvc_copts = xnnpack_msvc_std_copts(),
2554 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002555 x86_srcs = SSSE3_UKERNELS,
2556 deps = [
2557 ":tables",
2558 "@FP16",
2559 "@pthreadpool",
2560 ],
2561)
2562
2563xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002564 name = "ssse3_ukernels_test_mode",
2565 hdrs = INTERNAL_HDRS,
2566 copts = [
2567 "-UNDEBUG",
2568 "-DXNN_TEST_MODE=1",
2569 ],
2570 gcc_copts = xnnpack_gcc_std_copts(),
2571 gcc_x86_copts = ["-mssse3"],
2572 msvc_copts = xnnpack_msvc_std_copts(),
2573 msvc_x86_32_copts = ["/arch:SSE2"],
2574 x86_srcs = SSSE3_UKERNELS,
2575 deps = [
2576 ":tables",
2577 "@FP16",
2578 "@pthreadpool",
2579 ],
2580)
2581
2582xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002583 name = "sse41_ukernels",
2584 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002585 gcc_copts = xnnpack_gcc_std_copts(),
2586 gcc_x86_copts = ["-msse4.1"],
2587 msvc_copts = xnnpack_msvc_std_copts(),
2588 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002589 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002590 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002591 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002592 "@FP16",
2593 "@pthreadpool",
2594 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002595)
2596
2597xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002598 name = "sse41_ukernels_test_mode",
2599 hdrs = INTERNAL_HDRS,
2600 copts = [
2601 "-UNDEBUG",
2602 "-DXNN_TEST_MODE=1",
2603 ],
2604 gcc_copts = xnnpack_gcc_std_copts(),
2605 gcc_x86_copts = ["-msse4.1"],
2606 msvc_copts = xnnpack_msvc_std_copts(),
2607 msvc_x86_32_copts = ["/arch:SSE2"],
2608 x86_srcs = SSE41_UKERNELS,
2609 deps = [
2610 ":tables",
2611 "@FP16",
2612 "@pthreadpool",
2613 ],
2614)
2615
2616xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002617 name = "avx_ukernels",
2618 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002619 gcc_copts = xnnpack_gcc_std_copts(),
2620 gcc_x86_copts = ["-mavx"],
2621 msvc_copts = xnnpack_msvc_std_copts(),
2622 msvc_x86_32_copts = ["/arch:AVX"],
2623 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002624 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002625 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002626 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002627 "@FP16",
2628 "@pthreadpool",
2629 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002630)
2631
2632xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002633 name = "avx_ukernels_test_mode",
2634 hdrs = INTERNAL_HDRS,
2635 copts = [
2636 "-UNDEBUG",
2637 "-DXNN_TEST_MODE=1",
2638 ],
2639 gcc_copts = xnnpack_gcc_std_copts(),
2640 gcc_x86_copts = ["-mavx"],
2641 msvc_copts = xnnpack_msvc_std_copts(),
2642 msvc_x86_32_copts = ["/arch:AVX"],
2643 msvc_x86_64_copts = ["/arch:AVX"],
2644 x86_srcs = AVX_UKERNELS,
2645 deps = [
2646 ":tables",
2647 "@FP16",
2648 "@pthreadpool",
2649 ],
2650)
2651
2652xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002653 name = "fma3_ukernels",
2654 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002655 gcc_copts = xnnpack_gcc_std_copts(),
2656 gcc_x86_copts = ["-mfma"],
2657 msvc_copts = xnnpack_msvc_std_copts(),
2658 msvc_x86_32_copts = ["/arch:AVX"],
2659 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002660 x86_srcs = FMA3_UKERNELS,
2661 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002662 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002663 "@FP16",
2664 "@pthreadpool",
2665 ],
2666)
2667
2668xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002669 name = "fma3_ukernels_test_mode",
2670 hdrs = INTERNAL_HDRS,
2671 copts = [
2672 "-UNDEBUG",
2673 "-DXNN_TEST_MODE=1",
2674 ],
2675 gcc_copts = xnnpack_gcc_std_copts(),
2676 gcc_x86_copts = ["-mfma"],
2677 msvc_copts = xnnpack_msvc_std_copts(),
2678 msvc_x86_32_copts = ["/arch:AVX"],
2679 msvc_x86_64_copts = ["/arch:AVX"],
2680 x86_srcs = FMA3_UKERNELS,
2681 deps = [
2682 ":tables",
2683 "@FP16",
2684 "@pthreadpool",
2685 ],
2686)
2687
2688xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002689 name = "avx2_ukernels",
2690 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002691 gcc_copts = xnnpack_gcc_std_copts(),
2692 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002693 "-mfma",
2694 "-mavx2",
2695 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002696 msvc_copts = xnnpack_msvc_std_copts(),
2697 msvc_x86_32_copts = ["/arch:AVX2"],
2698 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002699 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002700 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002701 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002702 "@FP16",
2703 "@pthreadpool",
2704 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002705)
2706
2707xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002708 name = "avx2_ukernels_test_mode",
2709 hdrs = INTERNAL_HDRS,
2710 copts = [
2711 "-UNDEBUG",
2712 "-DXNN_TEST_MODE=1",
2713 ],
2714 gcc_copts = xnnpack_gcc_std_copts(),
2715 gcc_x86_copts = [
2716 "-mfma",
2717 "-mavx2",
2718 ],
2719 msvc_copts = xnnpack_msvc_std_copts(),
2720 msvc_x86_32_copts = ["/arch:AVX2"],
2721 msvc_x86_64_copts = ["/arch:AVX2"],
2722 x86_srcs = AVX2_UKERNELS,
2723 deps = [
2724 ":tables",
2725 "@FP16",
2726 "@pthreadpool",
2727 ],
2728)
2729
2730xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002731 name = "avx512f_ukernels",
2732 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002733 gcc_copts = xnnpack_gcc_std_copts(),
2734 gcc_x86_copts = ["-mavx512f"],
2735 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2736 msvc_copts = xnnpack_msvc_std_copts(),
2737 msvc_x86_32_copts = ["/arch:AVX512"],
2738 msvc_x86_64_copts = ["/arch:AVX512"],
2739 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002740 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002741 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002742 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002743 "@FP16",
2744 "@pthreadpool",
2745 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002746)
2747
2748xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002749 name = "avx512f_ukernels_test_mode",
2750 hdrs = INTERNAL_HDRS,
2751 copts = [
2752 "-UNDEBUG",
2753 "-DXNN_TEST_MODE=1",
2754 ],
2755 gcc_copts = xnnpack_gcc_std_copts(),
2756 gcc_x86_copts = ["-mavx512f"],
2757 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2758 msvc_copts = xnnpack_msvc_std_copts(),
2759 msvc_x86_32_copts = ["/arch:AVX512"],
2760 msvc_x86_64_copts = ["/arch:AVX512"],
2761 msys_copts = ["-fno-asynchronous-unwind-tables"],
2762 x86_srcs = AVX512F_UKERNELS,
2763 deps = [
2764 ":tables",
2765 "@FP16",
2766 "@pthreadpool",
2767 ],
2768)
2769
2770xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002771 name = "asm_ukernels",
2772 hdrs = ["src/xnnpack/assembly.h"],
2773 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002774 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002775 aarch64_srcs = AARCH64_ASM_UKERNELS,
2776)
2777
Marat Dukhan3b59de22020-06-03 20:15:19 -07002778xnnpack_cc_library(
2779 name = "logging_utils",
2780 srcs = LOGGING_SRCS,
2781 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2782 copts = LOGGING_COPTS + [
2783 "-Isrc",
2784 "-Iinclude",
2785 ] + select({
2786 ":debug_build": [],
2787 "//conditions:default": xnnpack_min_size_copts(),
2788 }),
2789 gcc_copts = xnnpack_gcc_std_copts(),
2790 msvc_copts = xnnpack_msvc_std_copts(),
2791 visibility = xnnpack_visibility(),
2792 deps = [
2793 "@FP16",
2794 "@clog",
2795 "@pthreadpool",
2796 ],
2797)
2798
Marat Dukhan08c4a432019-10-03 09:29:21 -07002799xnnpack_aggregate_library(
2800 name = "ukernels",
2801 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002802 ":neon_ukernels",
2803 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002804 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002805 ":asm_ukernels",
2806 ],
2807 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002808 ":neon_ukernels",
2809 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002810 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002811 ":neonfp16arith_ukernels",
2812 ":asm_ukernels",
2813 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002814 generic_deps = [
2815 ":scalar_ukernels",
2816 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002817 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002818 ":psimd_fastmath_ukernels",
2819 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002820 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002821 wasm_deps = [
2822 ":wasm_ukernels",
2823 ],
2824 wasmsimd_deps = [
2825 ":wasm_ukernels",
2826 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002827 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002828 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002829 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002830 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002831 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002832 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002833 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002834 ":avx512f_ukernels",
2835 ],
2836)
2837
Marat Dukhan33fcf782020-05-24 14:27:15 -07002838xnnpack_aggregate_library(
2839 name = "ukernels_test_mode",
2840 aarch32_deps = [
2841 ":neon_ukernels_test_mode",
2842 ":neonfma_ukernels_test_mode",
2843 ":neonv8_ukernels_test_mode",
2844 ":asm_ukernels",
2845 ],
2846 aarch64_deps = [
2847 ":neon_ukernels_test_mode",
2848 ":neonfma_ukernels_test_mode",
2849 ":neonv8_ukernels_test_mode",
2850 ":neonfp16arith_ukernels_test_mode",
2851 ":asm_ukernels",
2852 ],
2853 generic_deps = [
2854 ":scalar_ukernels_test_mode",
2855 ],
2856 psimd_deps = [
2857 ":psimd_fastmath_ukernels_test_mode",
2858 ":psimd_accmath_ukernels_test_mode",
2859 ],
2860 wasm_deps = [
2861 ":wasm_ukernels_test_mode",
2862 ],
2863 wasmsimd_deps = [
2864 ":wasm_ukernels_test_mode",
2865 ],
2866 x86_deps = [
2867 ":sse2_ukernels_test_mode",
2868 ":ssse3_ukernels_test_mode",
2869 ":sse41_ukernels_test_mode",
2870 ":avx_ukernels_test_mode",
2871 ":fma3_ukernels_test_mode",
2872 ":avx2_ukernels_test_mode",
2873 ":avx512f_ukernels_test_mode",
2874 ],
2875)
2876
Marat Dukhanae046f52020-06-15 13:16:14 -07002877xnnpack_aggregate_library(
2878 name = "ukernels_no_f16",
2879 aarch32_deps = [
2880 ":neon_ukernels",
2881 ":neonfma_ukernels",
2882 ":neonv8_ukernels",
2883 ":asm_ukernels",
2884 ],
2885 aarch64_deps = [
2886 ":neon_ukernels",
2887 ":neonfma_ukernels",
2888 ":neonv8_ukernels",
2889 ":asm_ukernels",
2890 ],
2891 generic_deps = [
2892 ":scalar_ukernels",
2893 ],
2894 psimd_deps = [
2895 ":psimd_fastmath_ukernels",
2896 ":psimd_accmath_ukernels",
2897 ],
2898 wasm_deps = [
2899 ":wasm_ukernels",
2900 ],
2901 wasmsimd_deps = [
2902 ":wasm_ukernels",
2903 ],
2904 x86_deps = [
2905 ":sse2_ukernels",
2906 ":ssse3_ukernels",
2907 ":sse41_ukernels",
2908 ":avx_ukernels",
2909 ":fma3_ukernels",
2910 ":avx2_ukernels",
2911 ":avx512f_ukernels",
2912 ],
2913)
2914
Marat Dukhan08c4a432019-10-03 09:29:21 -07002915xnnpack_cc_library(
2916 name = "im2col",
2917 srcs = ["src/im2col.c"],
2918 hdrs = [
2919 "src/xnnpack/common.h",
2920 "src/xnnpack/im2col.h",
2921 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002922 gcc_copts = xnnpack_gcc_std_copts(),
2923 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002924)
2925
2926xnnpack_cc_library(
2927 name = "indirection",
2928 srcs = ["src/indirection.c"],
2929 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002930 gcc_copts = xnnpack_gcc_std_copts(),
2931 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002932 deps = [
2933 "@FP16",
2934 "@FXdiv",
2935 "@pthreadpool",
2936 ],
2937)
2938
2939xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002940 name = "indirection_test_mode",
2941 srcs = ["src/indirection.c"],
2942 hdrs = INTERNAL_HDRS,
2943 copts = [
2944 "-UNDEBUG",
2945 "-DXNN_TEST_MODE=1",
2946 ],
2947 gcc_copts = xnnpack_gcc_std_copts(),
2948 msvc_copts = xnnpack_msvc_std_copts(),
2949 deps = [
2950 "@FP16",
2951 "@FXdiv",
2952 "@pthreadpool",
2953 ],
2954)
2955
2956xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002957 name = "operator_run",
2958 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002959 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002960 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07002961 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2962 "//conditions:default": [],
2963 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07002964 gcc_copts = xnnpack_gcc_std_copts(),
2965 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002966 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07002967 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002968 "@FP16",
2969 "@FXdiv",
2970 "@clog",
2971 "@pthreadpool",
2972 ],
2973)
2974
Chao Mei6ddfc602020-05-13 22:29:36 -07002975xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002976 name = "operator_run_test_mode",
2977 srcs = ["src/operator-run.c"],
2978 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2979 copts = LOGGING_COPTS + [
2980 "-UNDEBUG",
2981 "-DXNN_TEST_MODE=1",
2982 ] + select({
2983 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
2984 "//conditions:default": [],
2985 }),
2986 gcc_copts = xnnpack_gcc_std_copts(),
2987 msvc_copts = xnnpack_msvc_std_copts(),
2988 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07002989 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07002990 "@FP16",
2991 "@FXdiv",
2992 "@clog",
2993 "@pthreadpool",
2994 ],
2995)
2996
2997xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07002998 name = "memory_planner",
2999 srcs = ["src/memory-planner.c"],
3000 hdrs = INTERNAL_HDRS,
3001 defines = select({
3002 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3003 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3004 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3005 }),
3006 gcc_copts = xnnpack_gcc_std_copts(),
3007 msvc_copts = xnnpack_msvc_std_copts(),
3008 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003009 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003010 "@pthreadpool",
3011 ],
3012)
3013
Marat Dukhan33fcf782020-05-24 14:27:15 -07003014xnnpack_cc_library(
3015 name = "memory_planner_test_mode",
3016 srcs = ["src/memory-planner.c"],
3017 hdrs = INTERNAL_HDRS,
3018 copts = [
3019 "-UNDEBUG",
3020 "-DXNN_TEST_MODE=1",
3021 ],
3022 defines = select({
3023 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3024 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3025 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3026 }),
3027 gcc_copts = xnnpack_gcc_std_copts(),
3028 msvc_copts = xnnpack_msvc_std_copts(),
3029 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003030 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003031 "@pthreadpool",
3032 ],
3033)
3034
Marat Dukhan08c4a432019-10-03 09:29:21 -07003035cc_library(
3036 name = "enable_assembly",
3037 defines = select({
3038 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3039 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003040 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003041 }),
3042)
3043
Marat Dukhancf056b22019-10-07 10:26:29 -07003044xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003045 name = "operators",
3046 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003047 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003048 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003049 ],
3050 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003051 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003052 "-Isrc",
3053 "-Iinclude",
3054 ] + select({
3055 ":debug_build": [],
3056 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003057 }) + select({
3058 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3059 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003060 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003061 gcc_copts = xnnpack_gcc_std_copts(),
3062 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003063 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003064 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003065 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003066 "@FP16",
3067 "@FXdiv",
3068 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003069 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003070 ],
3071)
3072
Marat Dukhan10a38082020-04-17 03:58:35 -07003073xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003074 name = "operators_test_mode",
3075 srcs = OPERATOR_SRCS + [
3076 "src/memory.c",
3077 "src/operator-delete.c",
3078 ],
3079 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3080 copts = LOGGING_COPTS + [
3081 "-Isrc",
3082 "-Iinclude",
3083 "-UNDEBUG",
3084 "-DXNN_TEST_MODE=1",
3085 ] + select({
3086 ":debug_build": [],
3087 "//conditions:default": xnnpack_min_size_copts(),
3088 }) + select({
3089 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3090 "//conditions:default": [],
3091 }),
3092 gcc_copts = xnnpack_gcc_std_copts(),
3093 msvc_copts = xnnpack_msvc_std_copts(),
3094 deps = [
3095 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003096 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003097 "@FP16",
3098 "@FXdiv",
3099 "@clog",
3100 "@pthreadpool",
3101 ],
3102)
3103
3104xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003105 name = "XNNPACK",
3106 srcs = [
3107 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003108 "src/runtime.c",
3109 "src/subgraph.c",
3110 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003111 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003112 hdrs = ["include/xnnpack.h"],
3113 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003114 "-Isrc",
3115 "-Iinclude",
3116 ] + select({
3117 ":debug_build": [],
3118 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003119 }) + select({
3120 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3121 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003122 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003123 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003124 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003125 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003126 visibility = xnnpack_visibility(),
3127 deps = [
3128 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003129 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003130 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003131 ":operator_run",
3132 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003133 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003134 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003135 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003136 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003137 ] + select({
3138 ":emscripten": [],
3139 "//conditions:default": ["@cpuinfo"],
3140 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003141)
3142
Marat Dukhan10a38082020-04-17 03:58:35 -07003143xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003144 name = "XNNPACK_test_mode",
3145 srcs = [
3146 "src/init.c",
3147 "src/runtime.c",
3148 "src/subgraph.c",
3149 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003150 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003151 hdrs = ["include/xnnpack.h"],
3152 copts = LOGGING_COPTS + [
3153 "-Isrc",
3154 "-Iinclude",
3155 "-UNDEBUG",
3156 "-DXNN_TEST_MODE=1",
3157 ] + select({
3158 ":debug_build": [],
3159 "//conditions:default": xnnpack_min_size_copts(),
3160 }) + select({
3161 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3162 "//conditions:default": [],
3163 }),
3164 gcc_copts = xnnpack_gcc_std_copts(),
3165 includes = ["include"],
3166 msvc_copts = xnnpack_msvc_std_copts(),
3167 visibility = xnnpack_visibility(),
3168 deps = [
3169 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003170 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003171 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003172 ":operator_run_test_mode",
3173 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003174 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003175 "@clog",
3176 "@FP16",
3177 "@pthreadpool",
3178 ] + select({
3179 ":emscripten": [],
3180 "//conditions:default": ["@cpuinfo"],
3181 }),
3182)
3183
3184xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003185 name = "xnnpack_f32",
3186 srcs = [
3187 "src/init.c",
3188 "src/runtime.c",
3189 "src/subgraph.c",
3190 "src/tensor.c",
3191 ] + SUBGRAPH_SRCS,
3192 hdrs = ["include/xnnpack.h"],
3193 copts = LOGGING_COPTS + [
3194 "-Isrc",
3195 "-Iinclude",
3196 ] + select({
3197 ":debug_build": [],
3198 "//conditions:default": xnnpack_min_size_copts(),
3199 }) + select({
3200 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3201 "//conditions:default": [],
3202 }),
3203 defines = [
3204 "XNN_NO_Q8_OPERATORS",
3205 "XNN_NO_U8_OPERATORS",
3206 "XNN_NO_X8_OPERATORS",
3207 "XNN_NO_F16_OPERATORS",
3208 "XNN_NO_X16_OPERATORS",
3209 ],
3210 gcc_copts = xnnpack_gcc_std_copts(),
3211 includes = ["include"],
3212 msvc_copts = xnnpack_msvc_std_copts(),
3213 visibility = xnnpack_visibility(),
3214 deps = [
3215 ":enable_assembly",
3216 ":logging_utils",
3217 ":memory_planner",
3218 ":operator_run",
3219 ":operators",
3220 ":ukernels_no_f16",
3221 "@clog",
3222 "@FP16",
3223 "@pthreadpool",
3224 ] + select({
3225 ":emscripten": [],
3226 "//conditions:default": ["@cpuinfo"],
3227 }),
3228)
3229
3230xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003231 name = "xnnpack_operators_nhwc_f32",
3232 srcs = [
3233 "src/init.c",
3234 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003235 hdrs = ["include/xnnpack.h"],
3236 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003237 "-Isrc",
3238 "-Iinclude",
3239 ] + select({
3240 ":debug_build": [],
3241 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003242 }) + select({
3243 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3244 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003245 }),
3246 defines = [
3247 "XNN_NO_Q8_OPERATORS",
3248 "XNN_NO_U8_OPERATORS",
3249 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003250 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003251 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003252 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003253 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003254 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003255 visibility = xnnpack_visibility(),
3256 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003257 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003258 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003259 ":operator_run",
3260 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003261 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003262 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003263 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003264 ] + select({
3265 ":emscripten": [],
3266 "//conditions:default": ["@cpuinfo"],
3267 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003268)
3269
Marat Dukhancf056b22019-10-07 10:26:29 -07003270xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003271 name = "bench_utils",
3272 srcs = ["bench/utils.cc"],
3273 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003274 deps = [
3275 "@com_google_benchmark//:benchmark",
3276 "@cpuinfo",
3277 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003278)
3279
Frank Barchard7e955972019-10-11 10:34:25 -07003280######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003281
3282xnnpack_benchmark(
3283 name = "q8_gemm_bench",
3284 srcs = [
3285 "bench/gemm.h",
3286 "bench/q8-gemm.cc",
3287 "src/xnnpack/AlignedAllocator.h",
3288 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003289 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003290 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
3291)
3292
3293xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003294 name = "f16_igemm_bench",
3295 srcs = [
3296 "bench/f16-igemm.cc",
3297 "bench/conv.h",
3298 "bench/google/conv.h",
3299 "src/xnnpack/AlignedAllocator.h",
3300 ] + MICROKERNEL_BENCHMARK_HDRS,
3301 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3302)
3303
3304xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003305 name = "f16_gemm_bench",
3306 srcs = [
3307 "bench/f16-gemm.cc",
3308 "bench/gemm.h",
3309 "src/xnnpack/AlignedAllocator.h",
3310 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003311 deps = MICROKERNEL_BENCHMARK_DEPS,
3312)
3313
3314xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003315 name = "f16_spmm_bench",
3316 srcs = [
3317 "bench/f16-spmm.cc",
3318 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003319 "src/xnnpack/AlignedAllocator.h",
3320 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003321 deps = MICROKERNEL_BENCHMARK_DEPS,
3322)
3323
3324xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003325 name = "f32_igemm_bench",
3326 srcs = [
3327 "bench/f32-igemm.cc",
3328 "bench/conv.h",
3329 "src/xnnpack/AlignedAllocator.h",
3330 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07003331 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003332)
3333
3334xnnpack_benchmark(
3335 name = "f32_conv_hwc_bench",
3336 srcs = [
3337 "bench/f32-conv-hwc.cc",
3338 "bench/dconv.h",
3339 "src/xnnpack/AlignedAllocator.h",
3340 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003341 deps = MICROKERNEL_BENCHMARK_DEPS,
3342)
3343
3344xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003345 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003346 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003347 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003348 "bench/dconv.h",
3349 "src/xnnpack/AlignedAllocator.h",
3350 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Erich Elsen563df5f2019-10-23 08:02:21 -07003351 deps = MICROKERNEL_BENCHMARK_DEPS,
3352)
3353
3354xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003355 name = "f16_dwconv_bench",
3356 srcs = [
3357 "bench/f16-dwconv.cc",
3358 "bench/dwconv.h",
3359 "bench/google/dwconv.h",
3360 "src/xnnpack/AlignedAllocator.h",
3361 ] + MICROKERNEL_BENCHMARK_HDRS,
3362 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3363)
3364
3365xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003366 name = "f32_dwconv_bench",
3367 srcs = [
3368 "bench/f32-dwconv.cc",
3369 "bench/dwconv.h",
3370 "src/xnnpack/AlignedAllocator.h",
3371 ] + MICROKERNEL_BENCHMARK_HDRS,
3372 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3373)
3374
3375xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003376 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003377 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003378 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003379 "bench/dwconv.h",
3380 "src/xnnpack/AlignedAllocator.h",
3381 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3382 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3383)
3384
3385xnnpack_benchmark(
3386 name = "f32_gemm_bench",
3387 srcs = [
3388 "bench/f32-gemm.cc",
3389 "bench/gemm.h",
3390 "src/xnnpack/AlignedAllocator.h",
3391 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003392 copts = xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07003393 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003394)
3395
3396xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003397 name = "f32_raddexpminusmax_bench",
3398 srcs = [
3399 "bench/f32-raddexpminusmax.cc",
3400 "src/xnnpack/AlignedAllocator.h",
3401 ] + MICROKERNEL_BENCHMARK_HDRS,
3402 deps = MICROKERNEL_BENCHMARK_DEPS,
3403)
3404
3405xnnpack_benchmark(
3406 name = "f32_raddextexp_bench",
3407 srcs = [
3408 "bench/f32-raddextexp.cc",
3409 "src/xnnpack/AlignedAllocator.h",
3410 ] + MICROKERNEL_BENCHMARK_HDRS,
3411 deps = MICROKERNEL_BENCHMARK_DEPS,
3412)
3413
3414xnnpack_benchmark(
3415 name = "f32_raddstoreexpminusmax_bench",
3416 srcs = [
3417 "bench/f32-raddstoreexpminusmax.cc",
3418 "src/xnnpack/AlignedAllocator.h",
3419 ] + MICROKERNEL_BENCHMARK_HDRS,
3420 deps = MICROKERNEL_BENCHMARK_DEPS,
3421)
3422
3423xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003424 name = "f32_rmax_bench",
3425 srcs = [
3426 "bench/f32-rmax.cc",
3427 "src/xnnpack/AlignedAllocator.h",
3428 ] + MICROKERNEL_BENCHMARK_HDRS,
3429 deps = MICROKERNEL_BENCHMARK_DEPS,
3430)
3431
3432xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003433 name = "f32_sigmoid_bench",
3434 srcs = [
3435 "bench/f32-sigmoid.cc",
3436 "src/xnnpack/AlignedAllocator.h",
3437 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003438 deps = MICROKERNEL_BENCHMARK_DEPS,
3439)
3440
3441xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003442 name = "f32_spmm_bench",
3443 srcs = [
3444 "bench/f32-spmm.cc",
3445 "bench/gemm.h",
3446 "src/xnnpack/AlignedAllocator.h",
3447 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003448 deps = MICROKERNEL_BENCHMARK_DEPS,
3449)
3450
3451xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003452 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003453 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003454 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003455 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003456 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003457 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003458)
3459
3460xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003461 name = "f32_vscaleexpminusmax_bench",
3462 srcs = [
3463 "bench/f32-vscaleexpminusmax.cc",
3464 "src/xnnpack/AlignedAllocator.h",
3465 ] + MICROKERNEL_BENCHMARK_HDRS,
3466 deps = MICROKERNEL_BENCHMARK_DEPS,
3467)
3468
3469xnnpack_benchmark(
3470 name = "f32_vscaleextexp_bench",
3471 srcs = [
3472 "bench/f32-vscaleextexp.cc",
3473 "src/xnnpack/AlignedAllocator.h",
3474 ] + MICROKERNEL_BENCHMARK_HDRS,
3475 deps = MICROKERNEL_BENCHMARK_DEPS,
3476)
3477
3478xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003479 name = "f32_im2col_gemm_bench",
3480 srcs = [
3481 "bench/f32-im2col-gemm.cc",
3482 "bench/conv.h",
3483 "src/xnnpack/AlignedAllocator.h",
3484 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3485 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
3486)
3487
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003488xnnpack_benchmark(
3489 name = "requantization_bench",
3490 srcs = [
3491 "bench/requantization.cc",
3492 "src/xnnpack/requantization-stubs.h",
3493 "src/xnnpack/AlignedAllocator.h",
3494 ] + MICROKERNEL_BENCHMARK_HDRS,
3495 deps = MICROKERNEL_BENCHMARK_DEPS,
3496)
3497
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003498xnnpack_benchmark(
3499 name = "rounding_bench",
3500 srcs = [
3501 "bench/rounding.cc",
3502 "src/xnnpack/math-stubs.h",
3503 "src/xnnpack/AlignedAllocator.h",
3504 ] + MICROKERNEL_BENCHMARK_HDRS,
3505 deps = MICROKERNEL_BENCHMARK_DEPS,
3506)
3507
Marat Dukhan08c4a432019-10-03 09:29:21 -07003508########################### Benchmarks for operators ###########################
3509
3510xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003511 name = "average_pooling_bench",
3512 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003513 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003514 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003515 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003516)
3517
3518xnnpack_benchmark(
3519 name = "channel_shuffle_bench",
3520 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003521 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003522)
3523
3524xnnpack_benchmark(
3525 name = "convolution_bench",
3526 srcs = ["bench/convolution.cc"],
3527 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003528 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003529 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003530)
3531
3532xnnpack_benchmark(
3533 name = "deconvolution_bench",
3534 srcs = ["bench/deconvolution.cc"],
3535 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003536 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003537 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003538)
3539
3540xnnpack_benchmark(
3541 name = "global_average_pooling_bench",
3542 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003543 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003544)
3545
3546xnnpack_benchmark(
3547 name = "max_pooling_bench",
3548 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003549 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003550)
3551
3552xnnpack_benchmark(
3553 name = "sigmoid_bench",
3554 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003555 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003556 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003557 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003558)
3559
3560xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003561 name = "prelu_bench",
3562 srcs = ["bench/prelu.cc"],
3563 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003564 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003565 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003566)
3567
3568xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003569 name = "softmax_bench",
3570 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003571 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003572 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003573 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003574)
3575
Marat Dukhanc068bb62019-10-04 13:24:39 -07003576############################# End-to-end benchmarks ############################
3577
3578cc_library(
3579 name = "mobilenet_v1",
3580 srcs = ["models/mobilenet-v1.cc"],
3581 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003582 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003583 linkstatic = True,
3584 deps = [
3585 ":XNNPACK",
3586 "@pthreadpool",
3587 ],
3588)
3589
3590cc_library(
3591 name = "mobilenet_v2",
3592 srcs = ["models/mobilenet-v2.cc"],
3593 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003594 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003595 linkstatic = True,
3596 deps = [
3597 ":XNNPACK",
3598 "@pthreadpool",
3599 ],
3600)
3601
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003602cc_library(
3603 name = "mobilenet_v3_large",
3604 srcs = ["models/mobilenet-v3-large.cc"],
3605 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003606 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003607 linkstatic = True,
3608 deps = [
3609 ":XNNPACK",
3610 "@pthreadpool",
3611 ],
3612)
3613
3614cc_library(
3615 name = "mobilenet_v3_small",
3616 srcs = ["models/mobilenet-v3-small.cc"],
3617 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003618 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003619 linkstatic = True,
3620 deps = [
3621 ":XNNPACK",
3622 "@pthreadpool",
3623 ],
3624)
3625
Marat Dukhanc068bb62019-10-04 13:24:39 -07003626xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07003627 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003628 srcs = [
3629 "bench/f32-dwconv-e2e.cc",
3630 "bench/end2end.h",
3631 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07003632 deps = MICROKERNEL_BENCHMARK_DEPS + [
3633 ":XNNPACK",
3634 ":mobilenet_v1",
3635 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003636 ":mobilenet_v3_large",
3637 ":mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07003638 ],
3639)
3640
3641xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07003642 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003643 srcs = [
3644 "bench/f32-gemm-e2e.cc",
3645 "bench/end2end.h",
3646 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07003647 deps = MICROKERNEL_BENCHMARK_DEPS + [
3648 ":XNNPACK",
3649 ":mobilenet_v1",
3650 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003651 ":mobilenet_v3_large",
3652 ":mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07003653 ],
3654)
3655
3656xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07003657 name = "end2end_bench",
3658 srcs = ["bench/end2end.cc"],
3659 deps = [
3660 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07003661 ":bench_utils",
Marat Dukhanc068bb62019-10-04 13:24:39 -07003662 ":mobilenet_v1",
3663 ":mobilenet_v2",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003664 ":mobilenet_v3_large",
3665 ":mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07003666 "@pthreadpool",
3667 ],
3668)
3669
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003670#################### Accuracy evaluation for math functions ####################
3671
3672xnnpack_benchmark(
3673 name = "f32_exp_eval",
3674 srcs = [
3675 "eval/f32-exp.cc",
3676 "src/xnnpack/AlignedAllocator.h",
3677 ] + ACCURACY_EVAL_HDRS,
3678 deps = ACCURACY_EVAL_DEPS,
3679)
3680
Marat Dukhan515c9772019-10-17 18:07:57 -07003681xnnpack_benchmark(
3682 name = "f32_expminus_eval",
3683 srcs = [
3684 "eval/f32-expminus.cc",
3685 "src/xnnpack/AlignedAllocator.h",
3686 ] + ACCURACY_EVAL_HDRS,
3687 deps = ACCURACY_EVAL_DEPS,
3688)
3689
Marat Dukhan98ba4412019-10-23 02:14:28 -07003690xnnpack_benchmark(
3691 name = "f32_extexp_eval",
3692 srcs = [
3693 "eval/f32-extexp.cc",
3694 "src/xnnpack/AlignedAllocator.h",
3695 ] + ACCURACY_EVAL_HDRS,
3696 deps = ACCURACY_EVAL_DEPS,
3697)
3698
Marat Dukhan8853b822020-05-07 12:19:01 -07003699xnnpack_unit_test(
3700 name = "f32_roundne_eval",
3701 srcs = [
3702 "eval/f32-roundne.cc",
3703 "src/xnnpack/AlignedAllocator.h",
3704 "src/xnnpack/math-stubs.h",
3705 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07003706 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07003707 deps = MICROKERNEL_TEST_DEPS,
3708)
3709
Marat Dukhan2dbb9442020-05-12 20:43:43 -07003710xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07003711 name = "f32_roundd_eval",
3712 srcs = [
3713 "eval/f32-roundd.cc",
3714 "src/xnnpack/AlignedAllocator.h",
3715 "src/xnnpack/math-stubs.h",
3716 ] + MICROKERNEL_TEST_HDRS,
3717 automatic = False,
3718 deps = MICROKERNEL_TEST_DEPS,
3719)
3720
3721xnnpack_unit_test(
3722 name = "f32_roundu_eval",
3723 srcs = [
3724 "eval/f32-roundu.cc",
3725 "src/xnnpack/AlignedAllocator.h",
3726 "src/xnnpack/math-stubs.h",
3727 ] + MICROKERNEL_TEST_HDRS,
3728 automatic = False,
3729 deps = MICROKERNEL_TEST_DEPS,
3730)
3731
3732xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07003733 name = "f32_roundz_eval",
3734 srcs = [
3735 "eval/f32-roundz.cc",
3736 "src/xnnpack/AlignedAllocator.h",
3737 "src/xnnpack/math-stubs.h",
3738 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07003739 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07003740 deps = MICROKERNEL_TEST_DEPS,
3741)
3742
Marat Dukhan346a9e52019-11-15 09:06:30 -08003743xnnpack_benchmark(
3744 name = "f32_sigmoid_eval",
3745 srcs = [
3746 "eval/f32-sigmoid.cc",
3747 "src/xnnpack/AlignedAllocator.h",
3748 ] + ACCURACY_EVAL_HDRS,
3749 deps = ACCURACY_EVAL_DEPS,
3750)
3751
Marat Dukhan08c4a432019-10-03 09:29:21 -07003752######################### Unit tests for micro-kernels #########################
3753
3754xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07003755 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003756 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07003757 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003758 "test/gemm-microkernel-tester.h",
3759 "src/xnnpack/AlignedAllocator.h",
3760 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3761 deps = MICROKERNEL_TEST_DEPS,
3762)
3763
3764xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07003765 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003766 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07003767 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003768 "test/spmm-microkernel-tester.h",
3769 "src/xnnpack/AlignedAllocator.h",
3770 ] + MICROKERNEL_TEST_HDRS,
3771 deps = MICROKERNEL_TEST_DEPS,
3772)
3773
3774xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003775 name = "f32_argmaxpool_test",
3776 srcs = [
3777 "test/f32-argmaxpool.cc",
3778 "test/argmaxpool-microkernel-tester.h",
3779 "src/xnnpack/AlignedAllocator.h",
3780 ] + MICROKERNEL_TEST_HDRS,
3781 deps = MICROKERNEL_TEST_DEPS,
3782)
3783
3784xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003785 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003786 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003787 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003788 "test/avgpool-microkernel-tester.h",
3789 "src/xnnpack/AlignedAllocator.h",
3790 ] + MICROKERNEL_TEST_HDRS,
3791 deps = MICROKERNEL_TEST_DEPS,
3792)
3793
3794xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07003795 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08003796 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07003797 "test/f32-ibilinear.cc",
3798 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08003799 "src/xnnpack/AlignedAllocator.h",
3800 ] + MICROKERNEL_TEST_HDRS,
3801 deps = MICROKERNEL_TEST_DEPS,
3802)
3803
3804xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07003805 name = "f16_clamp_test",
3806 srcs = [
3807 "test/f16-clamp.cc",
3808 "test/clamp-microkernel-tester.h",
3809 ] + MICROKERNEL_TEST_HDRS,
3810 deps = MICROKERNEL_TEST_DEPS,
3811)
3812
3813xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003814 name = "f32_clamp_test",
3815 srcs = [
3816 "test/f32-clamp.cc",
3817 "test/clamp-microkernel-tester.h",
3818 ] + MICROKERNEL_TEST_HDRS,
3819 deps = MICROKERNEL_TEST_DEPS,
3820)
3821
3822xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07003823 name = "f32_igemm_test",
3824 srcs = [
3825 "test/f32-igemm.cc",
3826 "test/gemm-microkernel-tester.h",
3827 "src/xnnpack/AlignedAllocator.h",
3828 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3829 deps = MICROKERNEL_TEST_DEPS,
3830)
3831
3832xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07003833 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003834 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07003835 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003836 "test/gemm-microkernel-tester.h",
3837 "src/xnnpack/AlignedAllocator.h",
3838 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3839 deps = MICROKERNEL_TEST_DEPS,
3840)
3841
3842xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07003843 name = "f32_igemm_minmax_test",
3844 srcs = [
3845 "test/f32-igemm-minmax.cc",
3846 "test/gemm-microkernel-tester.h",
3847 "src/xnnpack/AlignedAllocator.h",
3848 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3849 deps = MICROKERNEL_TEST_DEPS,
3850)
3851
3852xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07003853 name = "f16_igemm_minmax_test",
3854 srcs = [
3855 "test/f16-igemm-minmax.cc",
3856 "test/gemm-microkernel-tester.h",
3857 "src/xnnpack/AlignedAllocator.h",
3858 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3859 deps = MICROKERNEL_TEST_DEPS,
3860)
3861
3862xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003863 name = "f32_conv_hwc_test",
3864 srcs = [
3865 "test/f32-conv-hwc.cc",
3866 "test/conv-hwc-microkernel-tester.h",
3867 "src/xnnpack/AlignedAllocator.h",
3868 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3869 deps = MICROKERNEL_TEST_DEPS,
3870)
3871
3872xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003873 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003874 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003875 "test/f32-conv-hwc2chw.cc",
3876 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003877 "src/xnnpack/AlignedAllocator.h",
3878 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3879 deps = MICROKERNEL_TEST_DEPS,
3880)
3881
3882xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07003883 name = "f32_dwconv_test",
3884 srcs = [
3885 "test/f32-dwconv.cc",
3886 "test/dwconv-microkernel-tester.h",
3887 "src/xnnpack/AlignedAllocator.h",
3888 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3889 deps = MICROKERNEL_TEST_DEPS,
3890)
3891
3892xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07003893 name = "f16_dwconv_minmax_test",
3894 srcs = [
3895 "test/f16-dwconv-minmax.cc",
3896 "test/dwconv-microkernel-tester.h",
3897 "src/xnnpack/AlignedAllocator.h",
3898 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3899 deps = MICROKERNEL_TEST_DEPS,
3900)
3901
3902xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07003903 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003904 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003905 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003906 "test/dwconv-microkernel-tester.h",
3907 "src/xnnpack/AlignedAllocator.h",
3908 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3909 deps = MICROKERNEL_TEST_DEPS,
3910)
3911
3912xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003913 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003914 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003915 "test/f32-dwconv-chw.cc",
3916 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003917 "src/xnnpack/AlignedAllocator.h",
3918 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3919 deps = MICROKERNEL_TEST_DEPS,
3920)
3921
3922xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07003923 name = "f16_gavgpool_minmax_test",
3924 srcs = [
3925 "test/f16-gavgpool-minmax.cc",
3926 "test/gavgpool-microkernel-tester.h",
3927 "src/xnnpack/AlignedAllocator.h",
3928 ] + MICROKERNEL_TEST_HDRS,
3929 deps = MICROKERNEL_TEST_DEPS,
3930)
3931
3932xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07003933 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003934 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07003935 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003936 "test/gavgpool-microkernel-tester.h",
3937 "src/xnnpack/AlignedAllocator.h",
3938 ] + MICROKERNEL_TEST_HDRS,
3939 deps = MICROKERNEL_TEST_DEPS,
3940)
3941
3942xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003943 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003944 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003945 "test/f32-gavgpool-cw.cc",
3946 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003947 "src/xnnpack/AlignedAllocator.h",
3948 ] + MICROKERNEL_TEST_HDRS,
3949 deps = MICROKERNEL_TEST_DEPS,
3950)
3951
3952xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07003953 name = "f32_gemm_test",
3954 srcs = [
3955 "test/f32-gemm.cc",
3956 "test/gemm-microkernel-tester.h",
3957 "src/xnnpack/AlignedAllocator.h",
3958 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3959 deps = MICROKERNEL_TEST_DEPS,
3960)
3961
3962xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07003963 name = "f32_gemm_relu_test",
3964 srcs = [
3965 "test/f32-gemm-relu.cc",
3966 "test/gemm-microkernel-tester.h",
3967 "src/xnnpack/AlignedAllocator.h",
3968 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3969 deps = MICROKERNEL_TEST_DEPS,
3970)
3971
3972xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07003973 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003974 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003975 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003976 "test/gemm-microkernel-tester.h",
3977 "src/xnnpack/AlignedAllocator.h",
3978 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3979 deps = MICROKERNEL_TEST_DEPS,
3980)
3981
3982xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07003983 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003984 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07003985 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003986 "test/gemm-microkernel-tester.h",
3987 "src/xnnpack/AlignedAllocator.h",
3988 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
3989 deps = MICROKERNEL_TEST_DEPS,
3990)
3991
3992xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07003993 name = "f16_hswish_test",
3994 srcs = [
3995 "test/f16-hswish.cc",
3996 "test/hswish-microkernel-tester.h",
3997 ] + MICROKERNEL_TEST_HDRS,
3998 deps = MICROKERNEL_TEST_DEPS,
3999)
4000
4001xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004002 name = "f32_hswish_test",
4003 srcs = [
4004 "test/f32-hswish.cc",
4005 "test/hswish-microkernel-tester.h",
4006 ] + MICROKERNEL_TEST_HDRS,
4007 deps = MICROKERNEL_TEST_DEPS,
4008)
4009
4010xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004011 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004012 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004013 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004014 "test/maxpool-microkernel-tester.h",
4015 ] + MICROKERNEL_TEST_HDRS,
4016 deps = MICROKERNEL_TEST_DEPS,
4017)
4018
4019xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004020 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004021 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004022 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004023 "test/avgpool-microkernel-tester.h",
4024 "src/xnnpack/AlignedAllocator.h",
4025 ] + MICROKERNEL_TEST_HDRS,
4026 deps = MICROKERNEL_TEST_DEPS,
4027)
4028
4029xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004030 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004031 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004032 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004033 "test/gemm-microkernel-tester.h",
4034 "src/xnnpack/AlignedAllocator.h",
4035 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4036 deps = MICROKERNEL_TEST_DEPS,
4037)
4038
4039xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004040 name = "f16_prelu_test",
4041 srcs = [
4042 "test/f16-prelu.cc",
4043 "test/prelu-microkernel-tester.h",
4044 "src/xnnpack/AlignedAllocator.h",
4045 ] + MICROKERNEL_TEST_HDRS,
4046 deps = MICROKERNEL_TEST_DEPS,
4047)
4048
4049xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004050 name = "f32_prelu_test",
4051 srcs = [
4052 "test/f32-prelu.cc",
4053 "test/prelu-microkernel-tester.h",
4054 "src/xnnpack/AlignedAllocator.h",
4055 ] + MICROKERNEL_TEST_HDRS,
4056 deps = MICROKERNEL_TEST_DEPS,
4057)
4058
4059xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004060 name = "f32_raddexpminusmax_test",
4061 srcs = [
4062 "test/f32-raddexpminusmax.cc",
4063 "test/raddexpminusmax-microkernel-tester.h",
4064 ] + MICROKERNEL_TEST_HDRS,
4065 deps = MICROKERNEL_TEST_DEPS,
4066)
4067
4068xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004069 name = "f32_raddextexp_test",
4070 srcs = [
4071 "test/f32-raddextexp.cc",
4072 "test/raddextexp-microkernel-tester.h",
4073 ] + MICROKERNEL_TEST_HDRS,
4074 deps = MICROKERNEL_TEST_DEPS,
4075)
4076
4077xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004078 name = "f32_raddstoreexpminusmax_test",
4079 srcs = [
4080 "test/f32-raddstoreexpminusmax.cc",
4081 "test/raddstoreexpminusmax-microkernel-tester.h",
4082 ] + MICROKERNEL_TEST_HDRS,
4083 deps = MICROKERNEL_TEST_DEPS,
4084)
4085
4086xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004087 name = "f32_rmax_test",
4088 srcs = [
4089 "test/f32-rmax.cc",
4090 "test/rmax-microkernel-tester.h",
4091 ] + MICROKERNEL_TEST_HDRS,
4092 deps = MICROKERNEL_TEST_DEPS,
4093)
4094
4095xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004096 name = "f32_sigmoid_test",
4097 srcs = [
4098 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004099 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004100 ] + MICROKERNEL_TEST_HDRS,
4101 deps = MICROKERNEL_TEST_DEPS,
4102)
4103
4104xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004105 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004106 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004107 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004108 "test/spmm-microkernel-tester.h",
4109 "src/xnnpack/AlignedAllocator.h",
4110 ] + MICROKERNEL_TEST_HDRS,
4111 deps = MICROKERNEL_TEST_DEPS,
4112)
4113
4114xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004115 name = "f16_vadd_minmax_test",
4116 srcs = [
4117 "test/f16-vadd-minmax.cc",
4118 "test/vbinary-microkernel-tester.h",
4119 ] + MICROKERNEL_TEST_HDRS,
4120 deps = MICROKERNEL_TEST_DEPS,
4121)
4122
4123xnnpack_unit_test(
4124 name = "f16_vaddc_minmax_test",
4125 srcs = [
4126 "test/f16-vaddc-minmax.cc",
4127 "test/vbinaryc-microkernel-tester.h",
4128 ] + MICROKERNEL_TEST_HDRS,
4129 deps = MICROKERNEL_TEST_DEPS,
4130)
4131
4132xnnpack_unit_test(
4133 name = "f16_vdiv_minmax_test",
4134 srcs = [
4135 "test/f16-vdiv-minmax.cc",
4136 "test/vbinary-microkernel-tester.h",
4137 ] + MICROKERNEL_TEST_HDRS,
4138 deps = MICROKERNEL_TEST_DEPS,
4139)
4140
4141xnnpack_unit_test(
4142 name = "f16_vdivc_minmax_test",
4143 srcs = [
4144 "test/f16-vdivc-minmax.cc",
4145 "test/vbinaryc-microkernel-tester.h",
4146 ] + MICROKERNEL_TEST_HDRS,
4147 deps = MICROKERNEL_TEST_DEPS,
4148)
4149
4150xnnpack_unit_test(
4151 name = "f16_vrdivc_minmax_test",
4152 srcs = [
4153 "test/f16-vrdivc-minmax.cc",
4154 "test/vbinaryc-microkernel-tester.h",
4155 ] + MICROKERNEL_TEST_HDRS,
4156 deps = MICROKERNEL_TEST_DEPS,
4157)
4158
4159xnnpack_unit_test(
4160 name = "f16_vmax_test",
4161 srcs = [
4162 "test/f16-vmax.cc",
4163 "test/vbinary-microkernel-tester.h",
4164 ] + MICROKERNEL_TEST_HDRS,
4165 deps = MICROKERNEL_TEST_DEPS,
4166)
4167
4168xnnpack_unit_test(
4169 name = "f16_vmaxc_test",
4170 srcs = [
4171 "test/f16-vmaxc.cc",
4172 "test/vbinaryc-microkernel-tester.h",
4173 ] + MICROKERNEL_TEST_HDRS,
4174 deps = MICROKERNEL_TEST_DEPS,
4175)
4176
4177xnnpack_unit_test(
4178 name = "f16_vmin_test",
4179 srcs = [
4180 "test/f16-vmin.cc",
4181 "test/vbinary-microkernel-tester.h",
4182 ] + MICROKERNEL_TEST_HDRS,
4183 deps = MICROKERNEL_TEST_DEPS,
4184)
4185
4186xnnpack_unit_test(
4187 name = "f16_vminc_test",
4188 srcs = [
4189 "test/f16-vminc.cc",
4190 "test/vbinaryc-microkernel-tester.h",
4191 ] + MICROKERNEL_TEST_HDRS,
4192 deps = MICROKERNEL_TEST_DEPS,
4193)
4194
4195xnnpack_unit_test(
4196 name = "f16_vmul_minmax_test",
4197 srcs = [
4198 "test/f16-vmul-minmax.cc",
4199 "test/vbinary-microkernel-tester.h",
4200 ] + MICROKERNEL_TEST_HDRS,
4201 deps = MICROKERNEL_TEST_DEPS,
4202)
4203
4204xnnpack_unit_test(
4205 name = "f16_vmulc_minmax_test",
4206 srcs = [
4207 "test/f16-vmulc-minmax.cc",
4208 "test/vbinaryc-microkernel-tester.h",
4209 ] + MICROKERNEL_TEST_HDRS,
4210 deps = MICROKERNEL_TEST_DEPS,
4211)
4212
4213xnnpack_unit_test(
4214 name = "f16_vsub_minmax_test",
4215 srcs = [
4216 "test/f16-vsub-minmax.cc",
4217 "test/vbinary-microkernel-tester.h",
4218 ] + MICROKERNEL_TEST_HDRS,
4219 deps = MICROKERNEL_TEST_DEPS,
4220)
4221
4222xnnpack_unit_test(
4223 name = "f16_vsubc_minmax_test",
4224 srcs = [
4225 "test/f16-vsubc-minmax.cc",
4226 "test/vbinaryc-microkernel-tester.h",
4227 ] + MICROKERNEL_TEST_HDRS,
4228 deps = MICROKERNEL_TEST_DEPS,
4229)
4230
4231xnnpack_unit_test(
4232 name = "f16_vrsubc_minmax_test",
4233 srcs = [
4234 "test/f16-vrsubc-minmax.cc",
4235 "test/vbinaryc-microkernel-tester.h",
4236 ] + MICROKERNEL_TEST_HDRS,
4237 deps = MICROKERNEL_TEST_DEPS,
4238)
4239
4240xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004241 name = "f32_vabs_test",
4242 srcs = [
4243 "test/f32-vabs.cc",
4244 "test/vunary-microkernel-tester.h",
4245 ] + MICROKERNEL_TEST_HDRS,
4246 deps = MICROKERNEL_TEST_DEPS,
4247)
4248
4249xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004250 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004251 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004252 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004253 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004254 ] + MICROKERNEL_TEST_HDRS,
4255 deps = MICROKERNEL_TEST_DEPS,
4256)
4257
4258xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004259 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004260 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004261 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004262 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004263 ] + MICROKERNEL_TEST_HDRS,
4264 deps = MICROKERNEL_TEST_DEPS,
4265)
4266
4267xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004268 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004269 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004270 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004271 "test/vbinary-microkernel-tester.h",
4272 ] + MICROKERNEL_TEST_HDRS,
4273 deps = MICROKERNEL_TEST_DEPS,
4274)
4275
4276xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004277 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004278 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004279 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004280 "test/vbinaryc-microkernel-tester.h",
4281 ] + MICROKERNEL_TEST_HDRS,
4282 deps = MICROKERNEL_TEST_DEPS,
4283)
4284
4285xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004286 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004287 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004288 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004289 "test/vbinaryc-microkernel-tester.h",
4290 ] + MICROKERNEL_TEST_HDRS,
4291 deps = MICROKERNEL_TEST_DEPS,
4292)
4293
4294xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004295 name = "f32_vmax_test",
4296 srcs = [
4297 "test/f32-vmax.cc",
4298 "test/vbinary-microkernel-tester.h",
4299 ] + MICROKERNEL_TEST_HDRS,
4300 deps = MICROKERNEL_TEST_DEPS,
4301)
4302
4303xnnpack_unit_test(
4304 name = "f32_vmaxc_test",
4305 srcs = [
4306 "test/f32-vmaxc.cc",
4307 "test/vbinaryc-microkernel-tester.h",
4308 ] + MICROKERNEL_TEST_HDRS,
4309 deps = MICROKERNEL_TEST_DEPS,
4310)
4311
4312xnnpack_unit_test(
4313 name = "f32_vmin_test",
4314 srcs = [
4315 "test/f32-vmin.cc",
4316 "test/vbinary-microkernel-tester.h",
4317 ] + MICROKERNEL_TEST_HDRS,
4318 deps = MICROKERNEL_TEST_DEPS,
4319)
4320
4321xnnpack_unit_test(
4322 name = "f32_vminc_test",
4323 srcs = [
4324 "test/f32-vminc.cc",
4325 "test/vbinaryc-microkernel-tester.h",
4326 ] + MICROKERNEL_TEST_HDRS,
4327 deps = MICROKERNEL_TEST_DEPS,
4328)
4329
4330xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004331 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004332 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004333 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004334 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004335 ] + MICROKERNEL_TEST_HDRS,
4336 deps = MICROKERNEL_TEST_DEPS,
4337)
4338
4339xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004340 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004341 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004342 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004343 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004344 ] + MICROKERNEL_TEST_HDRS,
4345 deps = MICROKERNEL_TEST_DEPS,
4346)
4347
4348xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004349 name = "f16_vmulcaddc_minmax_test",
4350 srcs = [
4351 "test/f16-vmulcaddc-minmax.cc",
4352 "test/vmulcaddc-microkernel-tester.h",
4353 "src/xnnpack/AlignedAllocator.h",
4354 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4355 deps = MICROKERNEL_TEST_DEPS,
4356)
4357
4358xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004359 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004360 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004361 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004362 "test/vmulcaddc-microkernel-tester.h",
4363 "src/xnnpack/AlignedAllocator.h",
4364 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4365 deps = MICROKERNEL_TEST_DEPS,
4366)
4367
4368xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004369 name = "f32_vlrelu_test",
4370 srcs = [
4371 "test/f32-vlrelu.cc",
4372 "test/vunary-microkernel-tester.h",
4373 ] + MICROKERNEL_TEST_HDRS,
4374 deps = MICROKERNEL_TEST_DEPS,
4375)
4376
4377xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004378 name = "f32_vneg_test",
4379 srcs = [
4380 "test/f32-vneg.cc",
4381 "test/vunary-microkernel-tester.h",
4382 ] + MICROKERNEL_TEST_HDRS,
4383 deps = MICROKERNEL_TEST_DEPS,
4384)
4385
4386xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004387 name = "f32_vrndne_test",
4388 srcs = [
4389 "test/f32-vrndne.cc",
4390 "test/vunary-microkernel-tester.h",
4391 ] + MICROKERNEL_TEST_HDRS,
4392 deps = MICROKERNEL_TEST_DEPS,
4393)
4394
4395xnnpack_unit_test(
4396 name = "f32_vrndz_test",
4397 srcs = [
4398 "test/f32-vrndz.cc",
4399 "test/vunary-microkernel-tester.h",
4400 ] + MICROKERNEL_TEST_HDRS,
4401 deps = MICROKERNEL_TEST_DEPS,
4402)
4403
4404xnnpack_unit_test(
4405 name = "f32_vrndu_test",
4406 srcs = [
4407 "test/f32-vrndu.cc",
4408 "test/vunary-microkernel-tester.h",
4409 ] + MICROKERNEL_TEST_HDRS,
4410 deps = MICROKERNEL_TEST_DEPS,
4411)
4412
4413xnnpack_unit_test(
4414 name = "f32_vrndd_test",
4415 srcs = [
4416 "test/f32-vrndd.cc",
4417 "test/vunary-microkernel-tester.h",
4418 ] + MICROKERNEL_TEST_HDRS,
4419 deps = MICROKERNEL_TEST_DEPS,
4420)
4421
4422xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004423 name = "f32_vscale_test",
4424 srcs = [
4425 "test/f32-vscale.cc",
4426 "test/vscale-microkernel-tester.h",
4427 ] + MICROKERNEL_TEST_HDRS,
4428 deps = MICROKERNEL_TEST_DEPS,
4429)
4430
4431xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004432 name = "f32_vscaleexpminusmax_test",
4433 srcs = [
4434 "test/f32-vscaleexpminusmax.cc",
4435 "test/vscaleexpminusmax-microkernel-tester.h",
4436 ] + MICROKERNEL_TEST_HDRS,
4437 deps = MICROKERNEL_TEST_DEPS,
4438)
4439
4440xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004441 name = "f32_vscaleextexp_test",
4442 srcs = [
4443 "test/f32-vscaleextexp.cc",
4444 "test/vscaleextexp-microkernel-tester.h",
4445 ] + MICROKERNEL_TEST_HDRS,
4446 deps = MICROKERNEL_TEST_DEPS,
4447)
4448
4449xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004450 name = "f32_vsqr_test",
4451 srcs = [
4452 "test/f32-vsqr.cc",
4453 "test/vunary-microkernel-tester.h",
4454 ] + MICROKERNEL_TEST_HDRS,
4455 deps = MICROKERNEL_TEST_DEPS,
4456)
4457
4458xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004459 name = "f32_vsqrdiff_test",
4460 srcs = [
4461 "test/f32-vsqrdiff.cc",
4462 "test/vbinary-microkernel-tester.h",
4463 ] + MICROKERNEL_TEST_HDRS,
4464 deps = MICROKERNEL_TEST_DEPS,
4465)
4466
4467xnnpack_unit_test(
4468 name = "f32_vsqrdiffc_test",
4469 srcs = [
4470 "test/f32-vsqrdiffc.cc",
4471 "test/vbinaryc-microkernel-tester.h",
4472 ] + MICROKERNEL_TEST_HDRS,
4473 deps = MICROKERNEL_TEST_DEPS,
4474)
4475
4476xnnpack_unit_test(
4477 name = "f32_vrsqrdiffc_test",
4478 srcs = [
4479 "test/f32-vrsqrdiffc.cc",
4480 "test/vbinaryc-microkernel-tester.h",
4481 ] + MICROKERNEL_TEST_HDRS,
4482 deps = MICROKERNEL_TEST_DEPS,
4483)
4484
4485xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004486 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07004487 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004488 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004489 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004490 ] + MICROKERNEL_TEST_HDRS,
4491 deps = MICROKERNEL_TEST_DEPS,
4492)
4493
4494xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004495 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004496 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004497 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004498 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004499 ] + MICROKERNEL_TEST_HDRS,
4500 deps = MICROKERNEL_TEST_DEPS,
4501)
4502
4503xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004504 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004505 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004506 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004507 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07004508 ] + MICROKERNEL_TEST_HDRS,
4509 deps = MICROKERNEL_TEST_DEPS,
4510)
4511
4512xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004513 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004514 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004515 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004516 "test/avgpool-microkernel-tester.h",
4517 "src/xnnpack/AlignedAllocator.h",
4518 ] + MICROKERNEL_TEST_HDRS,
4519 deps = MICROKERNEL_TEST_DEPS,
4520)
4521
4522xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004523 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004524 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004525 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004526 "test/gemm-microkernel-tester.h",
4527 "src/xnnpack/AlignedAllocator.h",
4528 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4529 deps = MICROKERNEL_TEST_DEPS,
4530)
4531
4532xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004533 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004534 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004535 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004536 "test/dwconv-microkernel-tester.h",
4537 "src/xnnpack/AlignedAllocator.h",
4538 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4539 deps = MICROKERNEL_TEST_DEPS,
4540)
4541
4542xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004543 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004544 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004545 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004546 "test/gavgpool-microkernel-tester.h",
4547 "src/xnnpack/AlignedAllocator.h",
4548 ] + MICROKERNEL_TEST_HDRS,
4549 deps = MICROKERNEL_TEST_DEPS,
4550)
4551
4552xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004553 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004554 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004555 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004556 "test/gemm-microkernel-tester.h",
4557 "src/xnnpack/AlignedAllocator.h",
4558 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4559 deps = MICROKERNEL_TEST_DEPS,
4560)
4561
4562xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004563 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004564 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004565 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004566 "test/vadd-microkernel-tester.h",
4567 ] + MICROKERNEL_TEST_HDRS,
4568 deps = MICROKERNEL_TEST_DEPS,
4569)
4570
4571xnnpack_unit_test(
4572 name = "u8_clamp_test",
4573 srcs = [
4574 "test/u8-clamp.cc",
4575 "test/clamp-microkernel-tester.h",
4576 ] + MICROKERNEL_TEST_HDRS,
4577 deps = MICROKERNEL_TEST_DEPS,
4578)
4579
4580xnnpack_unit_test(
4581 name = "u8_lut32norm_test",
4582 srcs = [
4583 "test/u8-lut32norm.cc",
4584 "test/lut-norm-microkernel-tester.h",
4585 ] + MICROKERNEL_TEST_HDRS,
4586 deps = MICROKERNEL_TEST_DEPS,
4587)
4588
4589xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004590 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004591 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004592 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004593 "test/maxpool-microkernel-tester.h",
4594 ] + MICROKERNEL_TEST_HDRS,
4595 deps = MICROKERNEL_TEST_DEPS,
4596)
4597
4598xnnpack_unit_test(
4599 name = "u8_rmax_test",
4600 srcs = [
4601 "test/u8-rmax.cc",
4602 "test/rmax-microkernel-tester.h",
4603 ] + MICROKERNEL_TEST_HDRS,
4604 deps = MICROKERNEL_TEST_DEPS,
4605)
4606
4607xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07004608 name = "x32_fill_test",
4609 srcs = [
4610 "test/x32-fill.cc",
4611 "test/fill-microkernel-tester.h",
4612 ] + MICROKERNEL_TEST_HDRS,
4613 deps = MICROKERNEL_TEST_DEPS,
4614)
4615
4616xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004617 name = "x32_packx_test",
4618 srcs = [
4619 "test/x32-packx.cc",
4620 "test/pack-microkernel-tester.h",
4621 "src/xnnpack/AlignedAllocator.h",
4622 ] + MICROKERNEL_TEST_HDRS,
4623 deps = MICROKERNEL_TEST_DEPS,
4624)
4625
4626xnnpack_unit_test(
4627 name = "x32_pad_test",
4628 srcs = [
4629 "test/x32-pad.cc",
4630 "test/pad-microkernel-tester.h",
4631 ] + MICROKERNEL_TEST_HDRS,
4632 deps = MICROKERNEL_TEST_DEPS,
4633)
4634
4635xnnpack_unit_test(
4636 name = "x32_unpool_test",
4637 srcs = [
4638 "test/x32-unpool.cc",
4639 "test/unpool-microkernel-tester.h",
4640 ] + MICROKERNEL_TEST_HDRS,
4641 deps = MICROKERNEL_TEST_DEPS,
4642)
4643
4644xnnpack_unit_test(
4645 name = "x32_zip_test",
4646 srcs = [
4647 "test/x32-zip.cc",
4648 "test/zip-microkernel-tester.h",
4649 ] + MICROKERNEL_TEST_HDRS,
4650 deps = MICROKERNEL_TEST_DEPS,
4651)
4652
4653xnnpack_unit_test(
4654 name = "x8_lut_test",
4655 srcs = [
4656 "test/x8-lut.cc",
4657 "test/lut-microkernel-tester.h",
4658 ] + MICROKERNEL_TEST_HDRS,
4659 deps = MICROKERNEL_TEST_DEPS,
4660)
4661
4662xnnpack_unit_test(
4663 name = "x8_zip_test",
4664 srcs = [
4665 "test/x8-zip.cc",
4666 "test/zip-microkernel-tester.h",
4667 ] + MICROKERNEL_TEST_HDRS,
4668 deps = MICROKERNEL_TEST_DEPS,
4669)
4670
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004671xnnpack_unit_test(
4672 name = "requantization_test",
4673 srcs = [
4674 "src/xnnpack/requantization-stubs.h",
4675 "test/requantization.cc",
4676 "test/requantization-tester.h",
4677 ] + MICROKERNEL_TEST_HDRS,
4678 deps = MICROKERNEL_TEST_DEPS,
4679)
4680
Marat Dukhan20c3b922020-03-10 03:45:06 -07004681########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004682
4683xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07004684 name = "operator_size_test",
4685 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004686 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004687)
4688
Marat Dukhan20c3b922020-03-10 03:45:06 -07004689xnnpack_binary(
4690 name = "subgraph_size_test",
4691 srcs = ["test/subgraph-size.c"],
4692 deps = [":XNNPACK"],
4693)
4694
4695########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004696
4697xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004698 name = "abs_nc_test",
4699 srcs = [
4700 "test/abs-nc.cc",
4701 "test/abs-operator-tester.h",
4702 ],
4703 deps = OPERATOR_TEST_DEPS,
4704)
4705
4706xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08004707 name = "add_nd_test",
4708 srcs = [
4709 "test/add-nd.cc",
4710 "test/binary-elementwise-operator-tester.h",
4711 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004712 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08004713)
4714
4715xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004716 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004717 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004718 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004719 "test/argmax-pooling-operator-tester.h",
4720 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004721 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004722)
4723
4724xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004725 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004726 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004727 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004728 "test/average-pooling-operator-tester.h",
4729 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004730 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004731)
4732
4733xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07004734 name = "bankers_rounding_nc_test",
4735 srcs = [
4736 "test/bankers-rounding-nc.cc",
4737 "test/bankers-rounding-operator-tester.h",
4738 ],
4739 deps = OPERATOR_TEST_DEPS,
4740)
4741
4742xnnpack_unit_test(
4743 name = "ceiling_nc_test",
4744 srcs = [
4745 "test/ceiling-nc.cc",
4746 "test/ceiling-operator-tester.h",
4747 ],
4748 deps = OPERATOR_TEST_DEPS,
4749)
4750
4751xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004752 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004753 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004754 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004755 "test/channel-shuffle-operator-tester.h",
4756 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004757 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004758)
4759
4760xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004761 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004762 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004763 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004764 "test/clamp-operator-tester.h",
4765 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004766 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004767)
4768
4769xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07004770 name = "constant_pad_nd_test",
4771 srcs = [
4772 "test/constant-pad-nd.cc",
4773 "test/constant-pad-operator-tester.h",
4774 ],
4775 deps = OPERATOR_TEST_DEPS,
4776)
4777
4778xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004779 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004780 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004781 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004782 "test/convolution-operator-tester.h",
4783 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004784 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004785)
4786
4787xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004788 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004789 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004790 "test/convolution-nchw.cc",
4791 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004792 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004793 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004794)
4795
4796xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07004797 name = "copy_nc_test",
4798 srcs = [
4799 "test/copy-nc.cc",
4800 "test/copy-operator-tester.h",
4801 ],
4802 deps = OPERATOR_TEST_DEPS,
4803)
4804
4805xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004806 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004807 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004808 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004809 "test/deconvolution-operator-tester.h",
4810 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004811 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004812)
4813
4814xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08004815 name = "divide_nd_test",
4816 srcs = [
4817 "test/binary-elementwise-operator-tester.h",
4818 "test/divide-nd.cc",
4819 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004820 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08004821)
4822
4823xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004824 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004825 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004826 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004827 "test/fully-connected-operator-tester.h",
4828 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004829 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004830)
4831
4832xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07004833 name = "floor_nc_test",
4834 srcs = [
4835 "test/floor-nc.cc",
4836 "test/floor-operator-tester.h",
4837 ],
4838 deps = OPERATOR_TEST_DEPS,
4839)
4840
4841xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004842 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004843 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004844 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004845 "test/global-average-pooling-operator-tester.h",
4846 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004847 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004848)
4849
4850xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004851 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004852 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004853 "test/global-average-pooling-ncw.cc",
4854 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004855 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004856 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004857)
4858
4859xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004860 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004861 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004862 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004863 "test/hardswish-operator-tester.h",
4864 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004865 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004866)
4867
4868xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004869 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004870 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004871 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004872 "test/leaky-relu-operator-tester.h",
4873 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004874 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004875)
4876
4877xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004878 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004879 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004880 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004881 "test/max-pooling-operator-tester.h",
4882 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004883 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004884)
4885
4886xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08004887 name = "maximum_nd_test",
4888 srcs = [
4889 "test/binary-elementwise-operator-tester.h",
4890 "test/maximum-nd.cc",
4891 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004892 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08004893)
4894
4895xnnpack_unit_test(
4896 name = "minimum_nd_test",
4897 srcs = [
4898 "test/binary-elementwise-operator-tester.h",
4899 "test/minimum-nd.cc",
4900 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004901 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08004902)
4903
4904xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004905 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08004906 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08004907 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004908 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08004909 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004910 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08004911)
4912
4913xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004914 name = "negate_nc_test",
4915 srcs = [
4916 "test/negate-nc.cc",
4917 "test/negate-operator-tester.h",
4918 ],
4919 deps = OPERATOR_TEST_DEPS,
4920)
4921
4922xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004923 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004924 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004925 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004926 "test/prelu-operator-tester.h",
4927 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004928 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004929)
4930
4931xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004932 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08004933 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004934 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08004935 "test/resize-bilinear-operator-tester.h",
4936 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07004937 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08004938)
4939
4940xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004941 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004942 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004943 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004944 "test/sigmoid-operator-tester.h",
4945 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004946 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004947)
4948
4949xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004950 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004951 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004952 "test/softmax-nc.cc",
4953 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004954 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004955 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004956)
4957
4958xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004959 name = "square_nc_test",
4960 srcs = [
4961 "test/square-nc.cc",
4962 "test/square-operator-tester.h",
4963 ],
4964 deps = OPERATOR_TEST_DEPS,
4965)
4966
4967xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07004968 name = "squared_difference_nd_test",
4969 srcs = [
4970 "test/binary-elementwise-operator-tester.h",
4971 "test/squared-difference-nd.cc",
4972 ],
4973 deps = OPERATOR_TEST_DEPS,
4974)
4975
4976xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08004977 name = "subtract_nd_test",
4978 srcs = [
4979 "test/binary-elementwise-operator-tester.h",
4980 "test/subtract-nd.cc",
4981 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07004982 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08004983)
4984
4985xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07004986 name = "truncation_nc_test",
4987 srcs = [
4988 "test/truncation-nc.cc",
4989 "test/truncation-operator-tester.h",
4990 ],
4991 deps = OPERATOR_TEST_DEPS,
4992)
4993
4994xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08004995 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004996 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08004997 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004998 "test/unpooling-operator-tester.h",
4999 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005000 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005001)
5002
Chao Mei6ddfc602020-05-13 22:29:36 -07005003############################### Misc unit tests ###############################
5004
5005xnnpack_unit_test(
5006 name = "memory_planner_test",
5007 srcs = [
5008 "test/memory-planner-test.cc",
5009 ],
5010 deps = [
5011 ":XNNPACK",
5012 ":memory_planner",
5013 ],
5014)
5015
Marat Dukhan08c4a432019-10-03 09:29:21 -07005016############################# Build configurations #############################
5017
Marat Dukhanb8642352019-10-30 15:43:02 -07005018# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005019config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005020 name = "xnn_enable_assembly_explicit_true",
5021 define_values = {"xnn_enable_assembly": "true"},
5022)
5023
5024# Disables usage of assembly kernels.
5025config_setting(
5026 name = "xnn_enable_assembly_explicit_false",
5027 define_values = {"xnn_enable_assembly": "false"},
5028)
5029
Marat Dukhan05702cf2020-03-26 15:41:33 -07005030# Disables usage of HMP-aware optimizations.
5031config_setting(
5032 name = "xnn_enable_hmp_explicit_false",
5033 define_values = {"xnn_enable_hmp": "false"},
5034)
5035
Chao Mei6ddfc602020-05-13 22:29:36 -07005036# Enable usage of optimized memory allocation
5037config_setting(
5038 name = "xnn_enable_memopt_explicit_true",
5039 define_values = {"xnnpack_enable_memopt": "true"},
5040)
5041
5042# Disable usage of optimized memory allocation
5043config_setting(
5044 name = "xnn_enable_memopt_explicit_false",
5045 define_values = {"xnnpack_enable_memopt": "false"},
5046)
5047
Marat Dukhanb8642352019-10-30 15:43:02 -07005048# Builds with -c dbg
5049config_setting(
5050 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005051 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005052 "compilation_mode": "dbg",
5053 },
5054)
5055
5056# Builds with -c opt
5057config_setting(
5058 name = "optimized_build",
5059 values = {
5060 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005061 },
5062)
5063
5064config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005065 name = "linux_k8",
5066 values = {"cpu": "k8"},
5067)
5068
5069config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005070 name = "linux_arm",
5071 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005072)
5073
5074config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005075 name = "linux_armeabi",
5076 values = {"cpu": "armeabi"},
5077)
5078
5079config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005080 name = "linux_armhf",
5081 values = {"cpu": "armhf"},
5082)
5083
5084config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005085 name = "linux_armv7a",
5086 values = {"cpu": "armv7a"},
5087)
5088
5089config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005090 name = "linux_aarch64",
5091 values = {"cpu": "aarch64"},
5092)
5093
5094config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005095 name = "android",
5096 values = {"crosstool_top": "//external:android/crosstool"},
5097)
5098
5099config_setting(
5100 name = "android_armv7",
5101 values = {
5102 "crosstool_top": "//external:android/crosstool",
5103 "cpu": "armeabi-v7a",
5104 },
5105)
5106
5107config_setting(
5108 name = "android_arm64",
5109 values = {
5110 "crosstool_top": "//external:android/crosstool",
5111 "cpu": "arm64-v8a",
5112 },
5113)
5114
5115config_setting(
5116 name = "android_x86",
5117 values = {
5118 "crosstool_top": "//external:android/crosstool",
5119 "cpu": "x86",
5120 },
5121)
5122
5123config_setting(
5124 name = "android_x86_64",
5125 values = {
5126 "crosstool_top": "//external:android/crosstool",
5127 "cpu": "x86_64",
5128 },
5129)
5130
5131config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005132 name = "windows_x86_64",
5133 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005134)
5135
5136config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005137 name = "windows_x86_64_clang",
5138 values = {
5139 "compiler": "clang-cl",
5140 "cpu": "x64_windows",
5141 },
5142)
5143
5144config_setting(
5145 name = "windows_x86_64_mingw",
5146 values = {
5147 "compiler": "mingw-gcc",
5148 "cpu": "x64_windows",
5149 },
5150)
5151
5152config_setting(
5153 name = "windows_x86_64_msys",
5154 values = {
5155 "compiler": "msys-gcc",
5156 "cpu": "x64_windows",
5157 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005158)
5159
5160config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005161 name = "macos_x86_64",
5162 values = {
5163 "apple_platform_type": "macos",
5164 "cpu": "darwin",
5165 },
5166)
5167
5168config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005169 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005170 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005171)
5172
5173config_setting(
5174 name = "emscripten_wasm",
5175 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005176 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005177 "cpu": "wasm",
5178 },
5179)
5180
5181config_setting(
5182 name = "emscripten_wasmsimd",
5183 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005184 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005185 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005186 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005187 },
5188)
5189
5190config_setting(
5191 name = "emscripten_asmjs",
5192 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005193 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005194 "cpu": "asmjs",
5195 },
5196)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005197
5198config_setting(
5199 name = "ios_armv7",
5200 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005201 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005202 "cpu": "ios_armv7",
5203 },
5204)
5205
5206config_setting(
5207 name = "ios_arm64",
5208 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005209 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005210 "cpu": "ios_arm64",
5211 },
5212)
5213
5214config_setting(
5215 name = "ios_arm64e",
5216 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005217 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005218 "cpu": "ios_arm64e",
5219 },
5220)
5221
5222config_setting(
5223 name = "ios_x86",
5224 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005225 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005226 "cpu": "ios_i386",
5227 },
5228)
5229
5230config_setting(
5231 name = "ios_x86_64",
5232 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005233 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005234 "cpu": "ios_x86_64",
5235 },
5236)
5237
5238config_setting(
5239 name = "watchos_armv7k",
5240 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005241 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005242 "cpu": "watchos_armv7k",
5243 },
5244)
5245
5246config_setting(
5247 name = "watchos_arm64_32",
5248 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005249 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005250 "cpu": "watchos_arm64_32",
5251 },
5252)
5253
5254config_setting(
5255 name = "watchos_x86",
5256 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005257 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005258 "cpu": "watchos_i386",
5259 },
5260)
5261
5262config_setting(
5263 name = "watchos_x86_64",
5264 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005265 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005266 "cpu": "watchos_x86_64",
5267 },
5268)
5269
5270config_setting(
5271 name = "tvos_arm64",
5272 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005273 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005274 "cpu": "tvos_arm64",
5275 },
5276)
5277
5278config_setting(
5279 name = "tvos_x86_64",
5280 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005281 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005282 "cpu": "tvos_x86_64",
5283 },
5284)