blob: 88fcdc16040d9ffb596eaebe7475f43f99eeb4eb [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
62 "src/operators/fully-connected-nc.c",
63 "src/operators/global-average-pooling-ncw.c",
64 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070065 "src/operators/leaky-relu-nc.c",
66 "src/operators/max-pooling-nhwc.c",
67 "src/operators/prelu-nc.c",
68 "src/operators/resize-bilinear-nhwc.c",
69 "src/operators/sigmoid-nc.c",
70 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070071 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070072 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070073]
74
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070075SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070076 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077 "src/subgraph/add2.c",
78 "src/subgraph/argmax-pooling-2d.c",
79 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070080 "src/subgraph/bankers-rounding.c",
81 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070082 "src/subgraph/clamp.c",
83 "src/subgraph/convolution-2d.c",
84 "src/subgraph/deconvolution-2d.c",
85 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070087 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070089 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070090 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070091 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070092 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/maximum2.c",
94 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070096 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/prelu.c",
98 "src/subgraph/sigmoid.c",
99 "src/subgraph/softmax.c",
100 "src/subgraph/static-constant-pad.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700101 "src/subgraph/square.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700102 "src/subgraph/squared-difference.c",
103 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700104 "src/subgraph/unpooling-2d.c",
105]
106
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800107TABLE_SRCS = [
108 "src/tables/exp2-k-over-64.c",
109 "src/tables/exp2-k-over-2048.c",
110]
111
Marat Dukhan08c4a432019-10-03 09:29:21 -0700112SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800113 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800114 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800115 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700116 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
117 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700118 "src/f32-clamp/gen/scalar-x1.c",
119 "src/f32-clamp/gen/scalar-x2.c",
120 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -0800121 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800122 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700123 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
124 "src/f32-dwconv-chw/3x3p1-scalar.c",
125 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
126 "src/f32-dwconv-chw/5x5p2-scalar.c",
127 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700128 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
129 "src/f32-dwconv/gen/up1x4-scalar.c",
130 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
131 "src/f32-dwconv/gen/up1x9-scalar.c",
132 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
133 "src/f32-dwconv/gen/up1x25-scalar.c",
134 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
135 "src/f32-dwconv/gen/up2x4-scalar.c",
136 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
137 "src/f32-dwconv/gen/up2x9-scalar.c",
138 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
139 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700140 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
141 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
142 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
143 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700144 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
145 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700146 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
147 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
148 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
149 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700150 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
151 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700152 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700153 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
154 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700155 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
156 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
157 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700158 "src/f32-gemm/gen/1x4-scalar.c",
159 "src/f32-gemm/gen/2x4-scalar.c",
160 "src/f32-gemm/gen/4x2-scalar.c",
161 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700162 "src/f32-gemm/gen/1x4-relu-scalar.c",
163 "src/f32-gemm/gen/2x4-relu-scalar.c",
164 "src/f32-gemm/gen/4x2-relu-scalar.c",
165 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700166 "src/f32-gemm/gen/1x4-minmax-scalar.c",
167 "src/f32-gemm/gen/2x4-minmax-scalar.c",
168 "src/f32-gemm/gen/4x2-minmax-scalar.c",
169 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800170 "src/f32-hswish/gen/scalar-x1.c",
171 "src/f32-hswish/gen/scalar-x2.c",
172 "src/f32-hswish/gen/scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700173 "src/f32-ibilinear/gen/scalar-c1.c",
174 "src/f32-ibilinear/gen/scalar-c2.c",
175 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700176 "src/f32-igemm/gen/1x4-scalar.c",
177 "src/f32-igemm/gen/2x4-scalar.c",
178 "src/f32-igemm/gen/4x2-scalar.c",
179 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700180 "src/f32-igemm/gen/1x4-relu-scalar.c",
181 "src/f32-igemm/gen/2x4-relu-scalar.c",
182 "src/f32-igemm/gen/4x2-relu-scalar.c",
183 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700184 "src/f32-igemm/gen/1x4-minmax-scalar.c",
185 "src/f32-igemm/gen/2x4-minmax-scalar.c",
186 "src/f32-igemm/gen/4x2-minmax-scalar.c",
187 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700188 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
189 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
190 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700191 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
192 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
193 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
194 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800195 "src/f32-prelu/gen/scalar-2x1.c",
196 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800197 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
198 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
199 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
200 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
201 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
202 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
203 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700209 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800210 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
211 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
212 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
213 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
214 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
215 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
216 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
217 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
218 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700219 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
220 "src/f32-spmm/gen/1x1-minmax-scalar.c",
221 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
222 "src/f32-spmm/gen/2x1-minmax-scalar.c",
223 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
224 "src/f32-spmm/gen/4x1-minmax-scalar.c",
225 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
226 "src/f32-spmm/gen/8x1-minmax-scalar.c",
227 "src/f32-spmm/gen/8x2-minmax-scalar.c",
228 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700229 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
230 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
231 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
232 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
233 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
234 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
235 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
236 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
237 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
238 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
239 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
240 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800241 "src/f32-vbinary/gen/vmax-scalar-x1.c",
242 "src/f32-vbinary/gen/vmax-scalar-x2.c",
243 "src/f32-vbinary/gen/vmax-scalar-x4.c",
244 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
245 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
246 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
247 "src/f32-vbinary/gen/vmin-scalar-x1.c",
248 "src/f32-vbinary/gen/vmin-scalar-x2.c",
249 "src/f32-vbinary/gen/vmin-scalar-x4.c",
250 "src/f32-vbinary/gen/vminc-scalar-x1.c",
251 "src/f32-vbinary/gen/vminc-scalar-x2.c",
252 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700253 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
254 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
255 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
256 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
257 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
258 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
259 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
260 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
261 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
262 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
263 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
264 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700265 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
266 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
267 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
268 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
269 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
270 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700271 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
272 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
273 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
274 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
275 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
276 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700277 "src/f32-vlrelu/gen/scalar-x1.c",
278 "src/f32-vlrelu/gen/scalar-x2.c",
279 "src/f32-vlrelu/gen/scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700280 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
281 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
282 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700283 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
284 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
285 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
286 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
287 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
288 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
289 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
290 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
291 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
292 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
293 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
294 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700295 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
296 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
297 "src/f32-vsqrt/gen/scalar-sqrt-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 Dukhan7c1f8082020-06-25 13:26:20 -0700425 "src/f32-prelu/gen/wasm-2x1.c",
426 "src/f32-prelu/gen/wasm-2x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700427 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
428 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
429 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
430 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
431 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
432 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
433 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
434 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
435 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
436 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
437 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
438 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800439 "src/f32-vbinary/gen/vmax-wasm-x1.c",
440 "src/f32-vbinary/gen/vmax-wasm-x2.c",
441 "src/f32-vbinary/gen/vmax-wasm-x4.c",
442 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
443 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
444 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
445 "src/f32-vbinary/gen/vmin-wasm-x1.c",
446 "src/f32-vbinary/gen/vmin-wasm-x2.c",
447 "src/f32-vbinary/gen/vmin-wasm-x4.c",
448 "src/f32-vbinary/gen/vminc-wasm-x1.c",
449 "src/f32-vbinary/gen/vminc-wasm-x2.c",
450 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700451 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
452 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
453 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
454 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
455 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
456 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
457 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
458 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
459 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
460 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
461 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
462 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
463 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
464 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
465 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
466 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
467 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
468 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700469 "src/f32-vlrelu/gen/wasm-x1.c",
470 "src/f32-vlrelu/gen/wasm-x2.c",
471 "src/f32-vlrelu/gen/wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700472 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
473 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
474 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800475]
476
Marat Dukhan290055c2020-06-09 12:24:29 -0700477WASMSIMD_UKERNELS = [
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700478 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
479 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
480 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
481 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
482 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
483 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700484 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
485 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
486 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
487 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
488 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
489 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
490 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
491 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
492 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
493 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
494 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
495 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
496 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
497 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
498 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
499 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
500 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
501 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
502 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
503 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
504 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
505 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
506 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
507 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700508 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
509 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
510 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
511 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat-arm.c",
512 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat-arm.c",
513 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat-x86.c",
514 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700515 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
516 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
517 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
518 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
519 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
520 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
521 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
522 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
523 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
524 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
525 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
526 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
527 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
528 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
529 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
530 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
531 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
532 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
533 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
534 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
535 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
536 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
537 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
538 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
539 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
540 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
541 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
542 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
543 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
544 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
545 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
546 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
547 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
548 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
549 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
550 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
551 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
552 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
553 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
554 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
555 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
556 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
557 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
558 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
559 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
560 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
561 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
562 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
563 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
564 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
565 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
566 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
567 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
568 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
569 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
570 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
571 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
572 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
573 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
574 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700575 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
576 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
577 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
578 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat-arm.c",
579 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat-arm.c",
580 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat-x86.c",
581 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700582 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
583 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
584 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
585 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
586 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
587 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
588 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
589 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
590 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
591 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
592 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
593 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
594 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
595 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
596 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
597 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
598 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
599 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
600 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
601 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
602 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
603 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
604 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
605 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
606 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
607 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
608 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
609 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
610 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
611 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhan9baec802020-06-25 21:34:35 -0700612 "src/f32-hswish/gen/wasmsimd-arm-x4.c",
613 "src/f32-hswish/gen/wasmsimd-arm-x8.c",
Marat Dukhanc303fe62020-06-26 10:09:25 -0700614 "src/f32-hswish/gen/wasmsimd-arm-x16.c",
Marat Dukhan9baec802020-06-25 21:34:35 -0700615 "src/f32-hswish/gen/wasmsimd-x86-x4.c",
616 "src/f32-hswish/gen/wasmsimd-x86-x8.c",
Marat Dukhanc303fe62020-06-26 10:09:25 -0700617 "src/f32-hswish/gen/wasmsimd-x86-x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700618 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
619 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
620 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
621 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700622 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
623 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
624 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
625 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700626 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
627 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700628 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
629 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
630 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
631 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700632 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
633 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700634 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
635 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
636 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
637 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700638 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
639 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700640 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
641 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
642 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
643 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700644 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
645 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700646 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
647 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
648 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
649 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700650 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
651 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
652 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
653 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700654 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
655 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
656 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
657 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700658 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
659 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
660 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
661 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700662 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
663 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
664 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
665 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700666 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
667 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700668 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
669 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
670 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
671 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700672 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
673 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700674 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
675 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
676 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
677 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700678 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
679 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700680 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
681 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
682 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
683 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700684 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
685 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700686 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
687 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700688 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
689 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700690 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
691 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
692 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
693 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700694 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
695 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700696 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
697 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
698 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
699 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700700 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
701 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan72b399a2020-06-28 22:41:38 -0700702 "src/f32-vrnd/gen/vrndne-wasmsimd-x4.c",
703 "src/f32-vrnd/gen/vrndne-wasmsimd-x8.c",
704 "src/f32-vrnd/gen/vrndz-wasmsimd-x4.c",
705 "src/f32-vrnd/gen/vrndz-wasmsimd-x8.c",
706 "src/f32-vrnd/gen/vrndu-wasmsimd-x4.c",
707 "src/f32-vrnd/gen/vrndu-wasmsimd-x8.c",
708 "src/f32-vrnd/gen/vrndd-wasmsimd-x4.c",
709 "src/f32-vrnd/gen/vrndd-wasmsimd-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700710 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
711 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700712 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
713 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
714 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
715 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
716 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
717 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700718 "src/math/roundne-wasmsimd-addsub.c",
719 "src/math/roundd-wasmsimd-addsub.c",
720 "src/math/roundu-wasmsimd-addsub.c",
721 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700722 "src/requantization/fp32-wasmsimd.c",
723]
724
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800725PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800726 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800727 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800728 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700729 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
730 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700731 "src/f32-clamp/gen/psimd-x4.c",
732 "src/f32-clamp/gen/psimd-x8.c",
Erich Elsen0a1970e2020-06-10 09:24:59 -0700733 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700734 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700735 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700736 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700737 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700738 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
739 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
740 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
741 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
742 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
743 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
744 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
745 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
746 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
747 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
748 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
749 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Erich Elsen6f278b52020-06-10 16:13:11 -0700750 "src/f32-gavgpool-cw/psimd-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700751 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
752 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700753 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
754 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
755 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
756 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
757 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
758 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
759 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
760 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
761 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
762 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
763 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
764 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
765 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
766 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
767 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
768 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
769 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
770 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
771 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800772 "src/f32-hswish/gen/psimd-x4.c",
773 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700774 "src/f32-ibilinear/gen/psimd-c4.c",
775 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700776 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
777 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
778 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
779 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
780 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
781 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
782 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
783 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
784 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
785 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700786 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700787 "src/f32-spmm/gen/4x1-minmax-psimd.c",
788 "src/f32-spmm/gen/8x1-minmax-psimd.c",
789 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700790 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
791 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700792 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800793 "src/f32-prelu/gen/psimd-2x4.c",
794 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800795 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700796 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
797 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
798 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
799 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
800 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
801 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
802 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
803 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800804 "src/f32-vbinary/gen/vmax-psimd-x4.c",
805 "src/f32-vbinary/gen/vmax-psimd-x8.c",
806 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
807 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
808 "src/f32-vbinary/gen/vmin-psimd-x4.c",
809 "src/f32-vbinary/gen/vmin-psimd-x8.c",
810 "src/f32-vbinary/gen/vminc-psimd-x4.c",
811 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700812 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
813 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
814 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
815 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
816 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
817 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
818 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
819 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700820 "src/f32-vbinary/gen/vsqrdiff-psimd-x4.c",
821 "src/f32-vbinary/gen/vsqrdiff-psimd-x8.c",
822 "src/f32-vbinary/gen/vsqrdiffc-psimd-x4.c",
823 "src/f32-vbinary/gen/vsqrdiffc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700824 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
825 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
826 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
827 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700828 "src/f32-vlrelu/gen/psimd-x4.c",
829 "src/f32-vlrelu/gen/psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700830 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
831 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700832 "src/f32-vunary/gen/vabs-psimd-x4.c",
833 "src/f32-vunary/gen/vabs-psimd-x8.c",
834 "src/f32-vunary/gen/vneg-psimd-x4.c",
835 "src/f32-vunary/gen/vneg-psimd-x8.c",
836 "src/f32-vunary/gen/vsqr-psimd-x4.c",
837 "src/f32-vunary/gen/vsqr-psimd-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700838 "src/x32-fill/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700839 "src/x32-packx/x4-psimd.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700840 "src/x32-pad/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700841 "src/x32-unpool/psimd.c",
842 "src/x32-zip/x2-psimd.c",
843 "src/x32-zip/x3-psimd.c",
844 "src/x32-zip/x4-psimd.c",
845 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700846 "src/requantization/precise-psimd.c",
847 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700848]
849
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800850PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800851 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
852 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
853 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
854 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
855 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
856 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
857 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
858 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
859 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
860 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
861 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
862 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800863 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
864 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
865 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
866 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
867 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
868 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700869 "src/f32-vrnd/gen/vrndne-psimd-x4.c",
870 "src/f32-vrnd/gen/vrndne-psimd-x8.c",
871 "src/f32-vrnd/gen/vrndz-psimd-x4.c",
872 "src/f32-vrnd/gen/vrndz-psimd-x8.c",
873 "src/f32-vrnd/gen/vrndu-psimd-x4.c",
874 "src/f32-vrnd/gen/vrndu-psimd-x8.c",
875 "src/f32-vrnd/gen/vrndd-psimd-x4.c",
876 "src/f32-vrnd/gen/vrndd-psimd-x8.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700877 "src/math/roundne-psimd-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700878 "src/math/roundd-psimd-addsub.c",
879 "src/math/roundu-psimd-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700880 "src/math/roundz-psimd-addsub.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800881 "src/math/sigmoid-psimd-p5-div.c",
882]
883
Marat Dukhan08c4a432019-10-03 09:29:21 -0700884# ISA-specific micro-kernels
885NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700886 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
887 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700888 "src/f32-clamp/gen/neon-x4.c",
889 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700890 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
891 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700892 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
893 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700894 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
895 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700896 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
897 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700898 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
899 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
900 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
901 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700902 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
903 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
904 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
905 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700906 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
907 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
908 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
909 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700910 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700911 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
912 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700913 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
914 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
915 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
916 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
917 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
918 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
919 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
920 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
921 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
922 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
923 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
924 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
925 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
926 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
927 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
928 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
929 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
930 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
931 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
932 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
933 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
934 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
935 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
936 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
937 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
938 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
939 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
940 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
941 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
942 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
943 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800944 "src/f32-hswish/gen/neon-x4.c",
945 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700946 "src/f32-ibilinear/gen/neon-c4.c",
947 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700948 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
949 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
950 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
951 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
952 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
953 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
954 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
955 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
956 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
957 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
958 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
959 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
960 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
961 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
962 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
963 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700964 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
965 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
966 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700967 "src/f32-ppmm/gen/4x8-minmax-neon.c",
968 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800969 "src/f32-prelu/gen/neon-2x4.c",
970 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800971 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
972 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
973 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
974 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
975 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
976 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
977 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
978 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
979 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
980 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
981 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
982 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
983 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
984 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
985 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
986 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
987 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
988 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
989 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
990 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
991 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
992 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
993 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
994 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700995 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800996 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -0800997 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
998 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
999 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1000 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1001 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1002 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1003 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1004 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1005 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1006 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1007 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1008 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1009 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1010 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1011 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1012 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1013 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1014 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001015 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1016 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1017 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1018 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001019 "src/f32-vbinary/gen/vmax-neon-x4.c",
1020 "src/f32-vbinary/gen/vmax-neon-x8.c",
1021 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1022 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1023 "src/f32-vbinary/gen/vmin-neon-x4.c",
1024 "src/f32-vbinary/gen/vmin-neon-x8.c",
1025 "src/f32-vbinary/gen/vminc-neon-x4.c",
1026 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001027 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1028 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1029 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1030 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1031 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1032 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001033 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1034 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1035 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1036 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001037 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1038 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1039 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1040 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001041 "src/f32-vlrelu/gen/neon-x4.c",
1042 "src/f32-vlrelu/gen/neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001043 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1044 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001045 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1046 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1047 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1048 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1049 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1050 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1051 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1052 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001053 "src/f32-vunary/gen/vabs-neon-x4.c",
1054 "src/f32-vunary/gen/vabs-neon-x8.c",
1055 "src/f32-vunary/gen/vneg-neon-x4.c",
1056 "src/f32-vunary/gen/vneg-neon-x8.c",
1057 "src/f32-vunary/gen/vsqr-neon-x4.c",
1058 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001059 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
1060 "src/q8-avgpool/9x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001061 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001062 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
1063 "src/q8-gavgpool/7x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001064 "src/q8-gemm/4x8-minmax-neon.c",
1065 "src/q8-gemm/8x8-minmax-neon.c",
1066 "src/q8-igemm/4x8-minmax-neon.c",
1067 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001068 "src/q8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001069 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001070 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001071 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001072 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001073 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001074 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001075 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001076 "src/x32-zip/x2-neon.c",
1077 "src/x32-zip/x3-neon.c",
1078 "src/x32-zip/x4-neon.c",
1079 "src/x32-zip/xm-neon.c",
1080 "src/x8-zip/x2-neon.c",
1081 "src/x8-zip/x3-neon.c",
1082 "src/x8-zip/x4-neon.c",
1083 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001084 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001085 "src/math/roundd-neon-addsub.c",
1086 "src/math/roundd-neon-cvt.c",
1087 "src/math/roundu-neon-addsub.c",
1088 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001089 "src/math/roundz-neon-addsub.c",
1090 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001091 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001092 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1093 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1094 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1095 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1096 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1097 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001098 "src/math/sqrt-neon-nr1rsqrts.c",
1099 "src/math/sqrt-neon-nr2rsqrts.c",
1100 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001101 "src/requantization/precise-neon.c",
1102 "src/requantization/fp32-neon.c",
1103 "src/requantization/q31-neon.c",
1104 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001105]
1106
1107NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001108 "src/f32-ibilinear/gen/neonfma-c4.c",
1109 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001110 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1111 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1112 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1113 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1114 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1115 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1116 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1117 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1118 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001119 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1120 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1121 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1122 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001123 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1124 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1125 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1126 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001127 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1128 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1129 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1130 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001131 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1132 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1133 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1134 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1135 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1136 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1137 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1138 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1139 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1140 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1141 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1142 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1143 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1144 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1145 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1146 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1147 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1148 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001149 "src/f32-hswish/gen/neonfma-x4.c",
1150 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001151 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1152 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001153 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1154 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1155 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1156 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1157 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1158 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1159 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1160 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1161 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1162 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1163 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1164 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1165 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1166 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1167 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1168 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1169 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1170 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1171 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1172 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1173 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1174 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1175 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1176 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001177 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1178 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1179 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1180 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1181 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1182 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1183 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1184 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1185 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1186 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1187 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1188 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1189 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1190 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1191 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1192 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1193 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1194 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1195 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1196 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1197 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1198 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1199 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1200 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1201 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1202 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1203 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1204 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1205 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1206 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1207 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1208 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1209 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1210 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1211 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1212 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1213 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1214 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1215 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1216 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1217 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1218 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1219 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1220 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1221 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1222 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1223 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1224 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1225 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1226 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1227 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1228 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1229 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1230 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001231 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1232 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001233 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1234 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1235 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1236 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1237 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1238 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1239 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1240 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1241 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1242 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1243 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1244 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1245 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1246 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1247 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1248 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1249 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1250 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1251 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1252 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001253 "src/math/exp-neonfma-lut64-p2.c",
1254 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001255 "src/math/expminus-neonfma-lut2048-p1.c",
1256 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001257 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001258 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1259 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1260 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1261 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1262 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1263 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1264 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1265 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1266 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1267 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1268 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1269 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1270 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1271 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1272 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1273 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1274 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1275 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001276 "src/math/sqrt-neonfma-nr1fma.c",
1277 "src/math/sqrt-neonfma-nr2fma.c",
1278 "src/math/sqrt-neonfma-nr3fma.c",
1279 "src/math/sqrt-neonfma-nr2fma1adj.c",
1280 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001281]
1282
1283AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001284 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1285 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1286 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1287 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1288 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1289 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001290 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1291 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1292 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1293 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1294 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1295 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1296 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1297 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1298 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1299 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1300 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1301 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1302 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1303 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1304 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1305 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1306 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1307 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1308 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1309 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001310 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1311 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001312 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1313 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001314 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1315 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001316 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1317 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001318 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1319 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1320 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1321 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1322 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001323 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1324 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1325 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1326 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1327 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1328 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1329 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1330 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1331 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1332 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1333 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1334 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1335 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1336 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1337 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1338 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1339 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1340 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001341 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1342 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1343 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1344 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1345 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1346 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1347 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1348 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1349 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1350 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1351 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1352 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1353 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1354 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1355 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1356 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1357 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1358 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001359 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1360 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001361 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1362 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1363 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1364 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1365 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1366 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001367]
1368
Marat Dukhan8853b822020-05-07 12:19:01 -07001369NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001370 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1371 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1372 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1373 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1374 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1375 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1376 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1377 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001378 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001379 "src/math/roundd-neonv8.c",
1380 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001381 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001382]
1383
Marat Dukhan08c4a432019-10-03 09:29:21 -07001384AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001385 "src/f16-clamp/gen/neonfp16arith-x16.c",
1386 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001387 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1388 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1389 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1390 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1391 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1392 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1393 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1394 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1395 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1396 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1397 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1398 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001399 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1400 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001401 "src/f16-hswish/gen/neonfp16arith-x16.c",
1402 "src/f16-hswish/gen/neonfp16arith-x8.c",
1403 "src/f16-prelu/gen/neonfp16arith-2x16.c",
1404 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001405 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1406 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1407 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1408 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1409 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1410 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1411 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1412 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1413 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1414 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1415 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1416 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1417 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1418 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1419 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1420 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1421 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1422 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1423 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1424 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1425 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1426 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1427 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1428 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1429 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1430 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1431 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1432 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001433 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1434 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001435 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1436 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1437 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1438 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1439 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1440 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1441 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1442 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001443 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1444 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1445 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1446 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001447 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1448 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1449 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1450 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1451 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1452 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1453 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1454 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1455 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1456 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1457 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1458 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001459 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1460 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1461 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1462 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1463 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1464 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1465 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1466 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001467]
1468
1469SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001470 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1471 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001472 "src/f32-clamp/gen/sse-x4.c",
1473 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001474 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1475 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001476 "src/f32-dwconv-chw/3x3p1-sse.c",
1477 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001478 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1479 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1480 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1481 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1482 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1483 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1484 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1485 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1486 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1487 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1488 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1489 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001490 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001491 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1492 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001493 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1494 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1495 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1496 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1497 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1498 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1499 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1500 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1501 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1502 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1503 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1504 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1505 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001506 "src/f32-hswish/gen/sse-x4.c",
1507 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001508 "src/f32-ibilinear/gen/sse-c4.c",
1509 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001510 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1511 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1512 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1513 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1514 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1515 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1516 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001517 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1518 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1519 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001520 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001521 "src/f32-prelu/gen/sse-2x4.c",
1522 "src/f32-prelu/gen/sse-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001523 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001524 "src/f32-spmm/gen/4x1-minmax-sse.c",
1525 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001526 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001527 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1528 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1529 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1530 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1531 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1532 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1533 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1534 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001535 "src/f32-vbinary/gen/vmax-sse-x4.c",
1536 "src/f32-vbinary/gen/vmax-sse-x8.c",
1537 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1538 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1539 "src/f32-vbinary/gen/vmin-sse-x4.c",
1540 "src/f32-vbinary/gen/vmin-sse-x8.c",
1541 "src/f32-vbinary/gen/vminc-sse-x4.c",
1542 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001543 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1544 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1545 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1546 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1547 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1548 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1549 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1550 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001551 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1552 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1553 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1554 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001555 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1556 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1557 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1558 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001559 "src/f32-vlrelu/gen/sse-x4.c",
1560 "src/f32-vlrelu/gen/sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001561 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1562 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001563 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1564 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001565 "src/f32-vunary/gen/vabs-sse-x4.c",
1566 "src/f32-vunary/gen/vabs-sse-x8.c",
1567 "src/f32-vunary/gen/vneg-sse-x4.c",
1568 "src/f32-vunary/gen/vneg-sse-x8.c",
1569 "src/f32-vunary/gen/vsqr-sse-x4.c",
1570 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001571 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001572 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001573 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001574 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001575 "src/math/roundd-sse-addsub.c",
1576 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001577 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001578 "src/math/sqrt-sse-hh1mac.c",
1579 "src/math/sqrt-sse-nr1mac.c",
1580 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001581]
1582
1583SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001584 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1585 "src/f32-argmaxpool/4x-sse2-c4.c",
1586 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001587 "src/f32-prelu/gen/sse2-2x4.c",
1588 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001589 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1590 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1591 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1592 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1593 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1594 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1595 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1596 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1597 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1598 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1599 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1600 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001601 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001602 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001603 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001604 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001605 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1606 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan0d3f4672020-06-25 16:42:58 -07001607 "src/f32-vlrelu/gen/sse2-x4.c",
1608 "src/f32-vlrelu/gen/sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001609 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1610 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1611 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1612 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1613 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1614 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1615 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1616 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001617 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1618 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001619 "src/q8-igemm/4x4c2-minmax-sse2.c",
1620 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001621 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1622 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001623 "src/q8-gemm/2x4c8-minmax-sse2.c",
1624 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001625 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001626 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001627 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001628 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001629 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001630 "src/x32-zip/x2-sse2.c",
1631 "src/x32-zip/x3-sse2.c",
1632 "src/x32-zip/x4-sse2.c",
1633 "src/x32-zip/xm-sse2.c",
1634 "src/x8-zip/x2-sse2.c",
1635 "src/x8-zip/x3-sse2.c",
1636 "src/x8-zip/x4-sse2.c",
1637 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001638 "src/math/exp-sse2-p5.c",
1639 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001640 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001641 "src/math/roundd-sse2-cvt.c",
1642 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001643 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001644 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001645 "src/requantization/precise-sse2.c",
1646 "src/requantization/fp32-sse2.c",
1647 "src/requantization/q31-sse2.c",
1648 "src/requantization/gemmlowp-sse2.c",
1649]
1650
1651SSSE3_UKERNELS = [
1652 "src/requantization/precise-ssse3.c",
1653 "src/requantization/q31-ssse3.c",
1654 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001655]
1656
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001657SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001658 "src/f32-prelu/gen/sse41-2x4.c",
1659 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001660 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001661 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001662 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001663 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001664 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1665 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan0d3f4672020-06-25 16:42:58 -07001666 "src/f32-vlrelu/gen/sse41-x4.c",
1667 "src/f32-vlrelu/gen/sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001668 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1669 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1670 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1671 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1672 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1673 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1674 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1675 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001676 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001677 "src/math/roundd-sse41.c",
1678 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001679 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001680 "src/requantization/precise-sse4.c",
1681 "src/requantization/q31-sse4.c",
1682 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001683]
1684
Marat Dukhan08c4a432019-10-03 09:29:21 -07001685AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001686 "src/f32-clamp/gen/avx-x8.c",
1687 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001688 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1689 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1690 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1691 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1692 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1693 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1694 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1695 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1696 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1697 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1698 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1699 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1700 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1701 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1702 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1703 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1704 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1705 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1706 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1707 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1708 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1709 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1710 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1711 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1712 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1713 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1714 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1715 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1716 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1717 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001718 "src/f32-hswish/gen/avx-x8.c",
1719 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001720 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1721 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1722 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1723 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1724 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1725 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1726 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1727 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1728 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001729 "src/f32-prelu/gen/avx-2x8.c",
1730 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001731 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001732 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1733 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1734 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1735 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1736 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1737 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1738 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1739 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001740 "src/f32-vbinary/gen/vmax-avx-x8.c",
1741 "src/f32-vbinary/gen/vmax-avx-x16.c",
1742 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1743 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1744 "src/f32-vbinary/gen/vmin-avx-x8.c",
1745 "src/f32-vbinary/gen/vmin-avx-x16.c",
1746 "src/f32-vbinary/gen/vminc-avx-x8.c",
1747 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001748 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1749 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1750 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1751 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1752 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1753 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1754 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1755 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001756 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1757 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1758 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1759 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001760 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1761 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1762 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1763 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001764 "src/f32-vlrelu/gen/avx-x8.c",
1765 "src/f32-vlrelu/gen/avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001766 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001767 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1768 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1769 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1770 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1771 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1772 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1773 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1774 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001775 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1776 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001777 "src/f32-vunary/gen/vabs-avx-x8.c",
1778 "src/f32-vunary/gen/vabs-avx-x16.c",
1779 "src/f32-vunary/gen/vneg-avx-x8.c",
1780 "src/f32-vunary/gen/vneg-avx-x16.c",
1781 "src/f32-vunary/gen/vsqr-avx-x8.c",
1782 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001783]
1784
Marat Dukhanfda12b82019-11-21 12:27:59 -08001785FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001786 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1787 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1788 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1789 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1790 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1791 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1792 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1793 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1794 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1795 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1796 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1797 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1798 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1799 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1800 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1801 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1802 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1803 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1804 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1805 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1806 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1807 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1808 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1809 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1810 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1811 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1812 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1813 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1814 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1815 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1816 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1817 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1818 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1819 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1820 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1821 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1822 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1823 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1824 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1825 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001826 "src/f32-hswish/gen/fma3-x8.c",
1827 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001828 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1829 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1830 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1831 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1832 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1833 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1834 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1835 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1836 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1837 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1838 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1839 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1840 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1841 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001842 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1843 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1844 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1845 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1846 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1847 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1848 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1849 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001850 "src/math/sqrt-fma3-nr1fma.c",
1851 "src/math/sqrt-fma3-nr2fma.c",
1852 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001853]
1854
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001855AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001856 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1857 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1858 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1859 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1860 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1861 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1862 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1863 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1864 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1865 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1866 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1867 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1868 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1869 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1870 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1871 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1872 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1873 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1874 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1875 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1876 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1877 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1878 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1879 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1880 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1881 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1882 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1883 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1884 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1885 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1886 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1887 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1888 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1889 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1890 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1891 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001892 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1893 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1894 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1895 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1896 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1897 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1898 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1899 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1900 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1901 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1902 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1903 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1904 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1905 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1906 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1907 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1908 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1909 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1910 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1911 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1912 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1913 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1914 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1915 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1916 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1917 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1918 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1919 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1920 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1921 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001922 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1923 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1924 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1925 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1926 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1927 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1928 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1929 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1930 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1931 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1932 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1933 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1934 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1935 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1936 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1937 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1938 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1939 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1940 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1941 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1942 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1943 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1944 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1945 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001946 "src/math/exp-avx2-p5.c",
1947 "src/math/exp-avx2-perm-p3.c",
1948 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001949 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001950 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001951 "src/math/sigmoid-avx2-rr2-p5-div.c",
1952 "src/math/sigmoid-avx2-rr1-p5-div.c",
1953 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1954 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1955 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1956 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001957]
1958
Marat Dukhan08c4a432019-10-03 09:29:21 -07001959AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001960 "src/f32-clamp/gen/avx512f-x16.c",
1961 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001962 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1963 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1964 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1965 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1966 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1967 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1968 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1969 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1970 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1971 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1972 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1973 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1974 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1975 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1976 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1977 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1978 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1979 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1980 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1981 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1982 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1983 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1984 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1985 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001986 "src/f32-hswish/gen/avx512f-x16.c",
1987 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001988 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
1989 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
1990 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
1991 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
1992 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
1993 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001994 "src/f32-prelu/gen/avx512f-2x16.c",
1995 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001996 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
1997 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
1998 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
1999 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2000 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2001 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2002 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2003 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2004 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2005 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2006 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2007 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2008 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2009 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2010 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2011 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2012 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2013 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2014 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2015 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2016 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2017 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2018 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2019 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2020 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2021 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2022 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2023 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2024 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2025 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2026 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2027 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2028 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2029 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2030 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2031 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002032 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002033 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2034 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2035 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2036 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2037 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2038 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2039 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2040 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002041 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2042 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2043 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2044 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2045 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2046 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2047 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2048 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002049 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2050 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2051 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2052 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2053 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2054 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2055 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2056 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002057 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2058 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2059 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2060 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002061 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2062 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2063 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2064 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07002065 "src/f32-vlrelu/gen/avx512f-x16.c",
2066 "src/f32-vlrelu/gen/avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002067 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002068 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2069 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2070 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2071 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2072 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2073 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2074 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2075 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2076 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2077 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2078 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2079 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2080 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2081 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2082 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2083 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2084 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2085 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2086 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2087 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2088 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2089 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2090 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2091 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002092 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2093 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2094 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2095 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2096 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2097 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2098 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2099 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002100 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2101 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2102 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2103 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2104 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2105 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2106 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2107 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002108 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2109 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2110 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2111 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2112 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2113 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002114 "src/math/exp-avx512f-p5-scalef.c",
2115 "src/math/exp-avx512f-p5.c",
2116 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002117 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002118 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002119 "src/math/sqrt-avx512f-nr1fma.c",
2120 "src/math/sqrt-avx512f-nr2fma.c",
2121 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002122]
2123
2124AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002125 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002126 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002127 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2128 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
2129 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2130 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002131 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2132 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S",
2133 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2134 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002135 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2136 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
2137 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2138 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002139]
2140
2141AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002142 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002143 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2144 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002145 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002146 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2147 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002148 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2149 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2150 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002151 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002152 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2153 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2154 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002155 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002156 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2157 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002158 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002159 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2160 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2161 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2162 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2163 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2164 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2165 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2166 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2167 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2168 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2169 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2170 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2171 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2172 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2173 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2174 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2175 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2176 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002177 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2178 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002179 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002180 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2181 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2182 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2183 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2184 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2185 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2186 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2187 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2188 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2189 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2190 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2191 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2192 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2193 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2194 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2195 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2196 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2197 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002198 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2199 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2200 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2201 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2202 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2203 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2204 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2205 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2206 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2207 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2208 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2209 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2210 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2211 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2212 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2213 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2214 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2215 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002216]
2217
Marat Dukhan1b354632020-03-23 12:50:22 -07002218INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002219 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002220 "src/xnnpack/argmaxpool.h",
2221 "src/xnnpack/avgpool.h",
2222 "src/xnnpack/clamp.h",
2223 "src/xnnpack/common.h",
2224 "src/xnnpack/conv.h",
2225 "src/xnnpack/dwconv.h",
2226 "src/xnnpack/gavgpool.h",
2227 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002228 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002229 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002230 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002231 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002232 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002233 "src/xnnpack/lut.h",
2234 "src/xnnpack/math.h",
2235 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002236 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002237 "src/xnnpack/packx.h",
2238 "src/xnnpack/pad.h",
2239 "src/xnnpack/params.h",
2240 "src/xnnpack/pavgpool.h",
2241 "src/xnnpack/ppmm.h",
2242 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002243 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002244 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002245 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002246 "src/xnnpack/rmax.h",
2247 "src/xnnpack/scalar-utils.h",
2248 "src/xnnpack/spmm.h",
2249 "src/xnnpack/unpool.h",
2250 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002251 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002252 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002253 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002254 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002255 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002256 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002257 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002258]
2259
2260INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002261 "include/xnnpack.h",
2262 "src/xnnpack/allocator.h",
2263 "src/xnnpack/compute.h",
2264 "src/xnnpack/im2col.h",
2265 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002266 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002267 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002268 "src/xnnpack/operator.h",
2269 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002270 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002271 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002272 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002273 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002274]
2275
Marat Dukhan1b354632020-03-23 12:50:22 -07002276ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002277 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002278]
2279
Marat Dukhan1b354632020-03-23 12:50:22 -07002280MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002281 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002282 "include/xnnpack.h",
2283]
2284
Marat Dukhan1b354632020-03-23 12:50:22 -07002285MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002286 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002287 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002288 "src/xnnpack/requantization.h",
2289 "include/xnnpack.h",
2290]
2291
2292OPERATOR_TEST_PARAMS_HDRS = [
2293 "src/xnnpack/params.h",
2294 "src/xnnpack/common.h",
2295]
2296
2297WEIGHTS_PACK_HDRS = [
2298 "src/xnnpack/pack.h",
2299 "src/xnnpack/operator.h",
2300 "src/xnnpack/compute.h",
2301]
2302
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002303LOGGING_COPTS = select({
2304 # No logging in optimized mode
2305 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2306 # Full logging in debug mode
2307 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2308 # Error-only logging in default (fastbuild) mode
2309 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2310})
2311
Marat Dukhan3b59de22020-06-03 20:15:19 -07002312LOGGING_SRCS = select({
2313 # No logging in optimized mode
2314 ":optimized_build": [],
2315 "//conditions:default": [
2316 "src/operator-strings.c",
2317 "src/subgraph-strings.c",
2318 ],
2319})
2320
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002321LOGGING_HDRS = [
2322 "src/xnnpack/log.h",
2323]
2324
Marat Dukhan08c4a432019-10-03 09:29:21 -07002325xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002326 name = "tables",
2327 srcs = TABLE_SRCS,
2328 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002329 gcc_copts = xnnpack_gcc_std_copts(),
2330 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002331)
2332
2333xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002334 name = "scalar_ukernels",
2335 srcs = SCALAR_UKERNELS,
2336 hdrs = INTERNAL_HDRS,
2337 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002338 gcc_copts = xnnpack_gcc_std_copts(),
2339 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002340 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002341 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002342 "@FP16",
2343 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002344 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002345 ],
2346)
2347
2348xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002349 name = "scalar_ukernels_test_mode",
2350 srcs = SCALAR_UKERNELS,
2351 hdrs = INTERNAL_HDRS,
2352 aarch32_copts = ["-marm"],
2353 copts = [
2354 "-UNDEBUG",
2355 "-DXNN_TEST_MODE=1",
2356 ],
2357 gcc_copts = xnnpack_gcc_std_copts(),
2358 msvc_copts = xnnpack_msvc_std_copts(),
2359 deps = [
2360 ":tables",
2361 "@FP16",
2362 "@FXdiv",
2363 "@pthreadpool",
2364 ],
2365)
2366
2367xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002368 name = "wasm_ukernels",
2369 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002370 gcc_copts = xnnpack_gcc_std_copts(),
2371 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002372 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002373 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002374 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002375 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002376 "@FP16",
2377 "@FXdiv",
2378 "@pthreadpool",
2379 ],
2380)
2381
2382xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002383 name = "wasm_ukernels_test_mode",
2384 hdrs = INTERNAL_HDRS,
2385 copts = [
2386 "-UNDEBUG",
2387 "-DXNN_TEST_MODE=1",
2388 ],
2389 gcc_copts = xnnpack_gcc_std_copts(),
2390 msvc_copts = xnnpack_msvc_std_copts(),
2391 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002392 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002393 deps = [
2394 ":tables",
2395 "@FP16",
2396 "@FXdiv",
2397 "@pthreadpool",
2398 ],
2399)
2400
2401xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002402 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002403 hdrs = INTERNAL_HDRS,
2404 aarch32_copts = [
2405 "-marm",
2406 "-mfpu=neon",
2407 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002408 gcc_copts = xnnpack_gcc_std_copts(),
2409 gcc_x86_copts = ["-msse2"],
2410 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002411 optimized_copts = [
2412 "-O3",
2413 "-ffast-math",
2414 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002415 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002416 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002417 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002418 "@FP16",
2419 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002420 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002421 ],
2422)
2423
2424xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002425 name = "psimd_fastmath_ukernels_test_mode",
2426 hdrs = INTERNAL_HDRS,
2427 aarch32_copts = [
2428 "-marm",
2429 "-mfpu=neon",
2430 ],
2431 copts = [
2432 "-UNDEBUG",
2433 "-DXNN_TEST_MODE=1",
2434 ],
2435 gcc_copts = xnnpack_gcc_std_copts(),
2436 gcc_x86_copts = ["-msse2"],
2437 msvc_copts = xnnpack_msvc_std_copts(),
2438 optimized_copts = [
2439 "-O3",
2440 "-ffast-math",
2441 ],
2442 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2443 deps = [
2444 ":tables",
2445 "@FP16",
2446 "@psimd",
2447 "@pthreadpool",
2448 ],
2449)
2450
2451xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002452 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002453 hdrs = INTERNAL_HDRS,
2454 aarch32_copts = [
2455 "-marm",
2456 "-mfpu=neon",
2457 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002458 gcc_copts = xnnpack_gcc_std_copts(),
2459 gcc_x86_copts = ["-msse2"],
2460 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002461 optimized_copts = [
2462 "-O3",
2463 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002464 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002465 deps = [
2466 ":tables",
2467 "@FP16",
2468 "@psimd",
2469 "@pthreadpool",
2470 ],
2471)
2472
2473xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002474 name = "psimd_accmath_ukernels_test_mode",
2475 hdrs = INTERNAL_HDRS,
2476 aarch32_copts = [
2477 "-marm",
2478 "-mfpu=neon",
2479 ],
2480 copts = [
2481 "-UNDEBUG",
2482 "-DXNN_TEST_MODE=1",
2483 ],
2484 gcc_copts = xnnpack_gcc_std_copts(),
2485 gcc_x86_copts = ["-msse2"],
2486 msvc_copts = xnnpack_msvc_std_copts(),
2487 optimized_copts = [
2488 "-O3",
2489 ],
2490 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2491 deps = [
2492 ":tables",
2493 "@FP16",
2494 "@psimd",
2495 "@pthreadpool",
2496 ],
2497)
2498
2499xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002500 name = "neon_ukernels",
2501 hdrs = INTERNAL_HDRS,
2502 aarch32_copts = [
2503 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002504 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002505 "-mfpu=neon",
2506 ],
2507 aarch32_srcs = NEON_UKERNELS,
2508 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002509 gcc_copts = xnnpack_gcc_std_copts(),
2510 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002511 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002512 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002513 "@FP16",
2514 "@pthreadpool",
2515 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002516)
2517
2518xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002519 name = "neon_ukernels_test_mode",
2520 hdrs = INTERNAL_HDRS,
2521 aarch32_copts = [
2522 "-marm",
2523 "-march=armv7-a",
2524 "-mfpu=neon",
2525 ],
2526 aarch32_srcs = NEON_UKERNELS,
2527 aarch64_srcs = NEON_UKERNELS,
2528 copts = [
2529 "-UNDEBUG",
2530 "-DXNN_TEST_MODE=1",
2531 ],
2532 gcc_copts = xnnpack_gcc_std_copts(),
2533 msvc_copts = xnnpack_msvc_std_copts(),
2534 deps = [
2535 ":tables",
2536 "@FP16",
2537 "@pthreadpool",
2538 ],
2539)
2540
2541xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002542 name = "neonfma_ukernels",
2543 hdrs = INTERNAL_HDRS,
2544 aarch32_copts = [
2545 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002546 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002547 "-mfpu=neon-vfpv4",
2548 ],
2549 aarch32_srcs = NEONFMA_UKERNELS,
2550 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002551 apple_aarch32_copts = [
2552 "-mcpu=swift",
2553 "-mtune=generic",
2554 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002555 gcc_copts = xnnpack_gcc_std_copts(),
2556 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002557 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002558 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002559 "@FP16",
2560 "@pthreadpool",
2561 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002562)
2563
2564xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002565 name = "neonfma_ukernels_test_mode",
2566 hdrs = INTERNAL_HDRS,
2567 aarch32_copts = [
2568 "-marm",
2569 "-march=armv7-a",
2570 "-mfpu=neon-vfpv4",
2571 ],
2572 aarch32_srcs = NEONFMA_UKERNELS,
2573 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002574 apple_aarch32_copts = [
2575 "-mcpu=swift",
2576 "-mtune=generic",
2577 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002578 copts = [
2579 "-UNDEBUG",
2580 "-DXNN_TEST_MODE=1",
2581 ],
2582 gcc_copts = xnnpack_gcc_std_copts(),
2583 msvc_copts = xnnpack_msvc_std_copts(),
2584 deps = [
2585 ":tables",
2586 "@FP16",
2587 "@pthreadpool",
2588 ],
2589)
2590
2591xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002592 name = "neonv8_ukernels",
2593 hdrs = INTERNAL_HDRS,
2594 aarch32_copts = [
2595 "-marm",
2596 "-march=armv8-a",
2597 "-mfpu=neon-fp-armv8",
2598 ],
2599 aarch32_srcs = NEONV8_UKERNELS,
2600 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002601 apple_aarch32_copts = [
2602 "-mcpu=cyclone",
2603 "-mtune=generic",
2604 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002605 gcc_copts = xnnpack_gcc_std_copts(),
2606 msvc_copts = xnnpack_msvc_std_copts(),
2607 deps = [
2608 ":tables",
2609 "@FP16",
2610 "@pthreadpool",
2611 ],
2612)
2613
2614xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002615 name = "neonv8_ukernels_test_mode",
2616 hdrs = INTERNAL_HDRS,
2617 aarch32_copts = [
2618 "-marm",
2619 "-march=armv8-a",
2620 "-mfpu=neon-fp-armv8",
2621 ],
2622 aarch32_srcs = NEONV8_UKERNELS,
2623 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002624 apple_aarch32_copts = [
2625 "-mcpu=cyclone",
2626 "-mtune=generic",
2627 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002628 copts = [
2629 "-UNDEBUG",
2630 "-DXNN_TEST_MODE=1",
2631 ],
2632 gcc_copts = xnnpack_gcc_std_copts(),
2633 msvc_copts = xnnpack_msvc_std_copts(),
2634 deps = [
2635 ":tables",
2636 "@FP16",
2637 "@pthreadpool",
2638 ],
2639)
2640
2641xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002642 name = "neonfp16arith_ukernels",
2643 hdrs = INTERNAL_HDRS,
2644 aarch64_copts = ["-march=armv8.2-a+fp16"],
2645 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002646 gcc_copts = xnnpack_gcc_std_copts(),
2647 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002648 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002649 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002650 "@FP16",
2651 "@pthreadpool",
2652 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002653)
2654
2655xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002656 name = "neonfp16arith_ukernels_test_mode",
2657 hdrs = INTERNAL_HDRS,
2658 aarch64_copts = ["-march=armv8.2-a+fp16"],
2659 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2660 copts = [
2661 "-UNDEBUG",
2662 "-DXNN_TEST_MODE=1",
2663 ],
2664 gcc_copts = xnnpack_gcc_std_copts(),
2665 msvc_copts = xnnpack_msvc_std_copts(),
2666 deps = [
2667 ":tables",
2668 "@FP16",
2669 "@pthreadpool",
2670 ],
2671)
2672
2673xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002674 name = "sse2_ukernels",
2675 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002676 gcc_copts = xnnpack_gcc_std_copts(),
2677 gcc_x86_copts = ["-msse2"],
2678 msvc_copts = xnnpack_msvc_std_copts(),
2679 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002680 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002681 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002682 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002683 "@FP16",
2684 "@pthreadpool",
2685 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002686)
2687
2688xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002689 name = "sse2_ukernels_test_mode",
2690 hdrs = INTERNAL_HDRS,
2691 copts = [
2692 "-UNDEBUG",
2693 "-DXNN_TEST_MODE=1",
2694 ],
2695 gcc_copts = xnnpack_gcc_std_copts(),
2696 gcc_x86_copts = ["-msse2"],
2697 msvc_copts = xnnpack_msvc_std_copts(),
2698 msvc_x86_32_copts = ["/arch:SSE2"],
2699 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2700 deps = [
2701 ":tables",
2702 "@FP16",
2703 "@pthreadpool",
2704 ],
2705)
2706
2707xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002708 name = "ssse3_ukernels",
2709 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002710 gcc_copts = xnnpack_gcc_std_copts(),
2711 gcc_x86_copts = ["-mssse3"],
2712 msvc_copts = xnnpack_msvc_std_copts(),
2713 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002714 x86_srcs = SSSE3_UKERNELS,
2715 deps = [
2716 ":tables",
2717 "@FP16",
2718 "@pthreadpool",
2719 ],
2720)
2721
2722xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002723 name = "ssse3_ukernels_test_mode",
2724 hdrs = INTERNAL_HDRS,
2725 copts = [
2726 "-UNDEBUG",
2727 "-DXNN_TEST_MODE=1",
2728 ],
2729 gcc_copts = xnnpack_gcc_std_copts(),
2730 gcc_x86_copts = ["-mssse3"],
2731 msvc_copts = xnnpack_msvc_std_copts(),
2732 msvc_x86_32_copts = ["/arch:SSE2"],
2733 x86_srcs = SSSE3_UKERNELS,
2734 deps = [
2735 ":tables",
2736 "@FP16",
2737 "@pthreadpool",
2738 ],
2739)
2740
2741xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002742 name = "sse41_ukernels",
2743 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002744 gcc_copts = xnnpack_gcc_std_copts(),
2745 gcc_x86_copts = ["-msse4.1"],
2746 msvc_copts = xnnpack_msvc_std_copts(),
2747 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002748 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002749 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002750 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002751 "@FP16",
2752 "@pthreadpool",
2753 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002754)
2755
2756xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002757 name = "sse41_ukernels_test_mode",
2758 hdrs = INTERNAL_HDRS,
2759 copts = [
2760 "-UNDEBUG",
2761 "-DXNN_TEST_MODE=1",
2762 ],
2763 gcc_copts = xnnpack_gcc_std_copts(),
2764 gcc_x86_copts = ["-msse4.1"],
2765 msvc_copts = xnnpack_msvc_std_copts(),
2766 msvc_x86_32_copts = ["/arch:SSE2"],
2767 x86_srcs = SSE41_UKERNELS,
2768 deps = [
2769 ":tables",
2770 "@FP16",
2771 "@pthreadpool",
2772 ],
2773)
2774
2775xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002776 name = "avx_ukernels",
2777 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002778 gcc_copts = xnnpack_gcc_std_copts(),
2779 gcc_x86_copts = ["-mavx"],
2780 msvc_copts = xnnpack_msvc_std_copts(),
2781 msvc_x86_32_copts = ["/arch:AVX"],
2782 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002783 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002784 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002785 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002786 "@FP16",
2787 "@pthreadpool",
2788 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002789)
2790
2791xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002792 name = "avx_ukernels_test_mode",
2793 hdrs = INTERNAL_HDRS,
2794 copts = [
2795 "-UNDEBUG",
2796 "-DXNN_TEST_MODE=1",
2797 ],
2798 gcc_copts = xnnpack_gcc_std_copts(),
2799 gcc_x86_copts = ["-mavx"],
2800 msvc_copts = xnnpack_msvc_std_copts(),
2801 msvc_x86_32_copts = ["/arch:AVX"],
2802 msvc_x86_64_copts = ["/arch:AVX"],
2803 x86_srcs = AVX_UKERNELS,
2804 deps = [
2805 ":tables",
2806 "@FP16",
2807 "@pthreadpool",
2808 ],
2809)
2810
2811xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002812 name = "fma3_ukernels",
2813 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002814 gcc_copts = xnnpack_gcc_std_copts(),
2815 gcc_x86_copts = ["-mfma"],
2816 msvc_copts = xnnpack_msvc_std_copts(),
2817 msvc_x86_32_copts = ["/arch:AVX"],
2818 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002819 x86_srcs = FMA3_UKERNELS,
2820 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002821 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002822 "@FP16",
2823 "@pthreadpool",
2824 ],
2825)
2826
2827xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002828 name = "fma3_ukernels_test_mode",
2829 hdrs = INTERNAL_HDRS,
2830 copts = [
2831 "-UNDEBUG",
2832 "-DXNN_TEST_MODE=1",
2833 ],
2834 gcc_copts = xnnpack_gcc_std_copts(),
2835 gcc_x86_copts = ["-mfma"],
2836 msvc_copts = xnnpack_msvc_std_copts(),
2837 msvc_x86_32_copts = ["/arch:AVX"],
2838 msvc_x86_64_copts = ["/arch:AVX"],
2839 x86_srcs = FMA3_UKERNELS,
2840 deps = [
2841 ":tables",
2842 "@FP16",
2843 "@pthreadpool",
2844 ],
2845)
2846
2847xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002848 name = "avx2_ukernels",
2849 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002850 gcc_copts = xnnpack_gcc_std_copts(),
2851 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002852 "-mfma",
2853 "-mavx2",
2854 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002855 msvc_copts = xnnpack_msvc_std_copts(),
2856 msvc_x86_32_copts = ["/arch:AVX2"],
2857 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002858 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002859 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002860 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002861 "@FP16",
2862 "@pthreadpool",
2863 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002864)
2865
2866xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002867 name = "avx2_ukernels_test_mode",
2868 hdrs = INTERNAL_HDRS,
2869 copts = [
2870 "-UNDEBUG",
2871 "-DXNN_TEST_MODE=1",
2872 ],
2873 gcc_copts = xnnpack_gcc_std_copts(),
2874 gcc_x86_copts = [
2875 "-mfma",
2876 "-mavx2",
2877 ],
2878 msvc_copts = xnnpack_msvc_std_copts(),
2879 msvc_x86_32_copts = ["/arch:AVX2"],
2880 msvc_x86_64_copts = ["/arch:AVX2"],
2881 x86_srcs = AVX2_UKERNELS,
2882 deps = [
2883 ":tables",
2884 "@FP16",
2885 "@pthreadpool",
2886 ],
2887)
2888
2889xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002890 name = "avx512f_ukernels",
2891 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002892 gcc_copts = xnnpack_gcc_std_copts(),
2893 gcc_x86_copts = ["-mavx512f"],
2894 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2895 msvc_copts = xnnpack_msvc_std_copts(),
2896 msvc_x86_32_copts = ["/arch:AVX512"],
2897 msvc_x86_64_copts = ["/arch:AVX512"],
2898 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002899 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002900 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002901 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002902 "@FP16",
2903 "@pthreadpool",
2904 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002905)
2906
2907xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002908 name = "avx512f_ukernels_test_mode",
2909 hdrs = INTERNAL_HDRS,
2910 copts = [
2911 "-UNDEBUG",
2912 "-DXNN_TEST_MODE=1",
2913 ],
2914 gcc_copts = xnnpack_gcc_std_copts(),
2915 gcc_x86_copts = ["-mavx512f"],
2916 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2917 msvc_copts = xnnpack_msvc_std_copts(),
2918 msvc_x86_32_copts = ["/arch:AVX512"],
2919 msvc_x86_64_copts = ["/arch:AVX512"],
2920 msys_copts = ["-fno-asynchronous-unwind-tables"],
2921 x86_srcs = AVX512F_UKERNELS,
2922 deps = [
2923 ":tables",
2924 "@FP16",
2925 "@pthreadpool",
2926 ],
2927)
2928
2929xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002930 name = "asm_ukernels",
2931 hdrs = ["src/xnnpack/assembly.h"],
2932 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002933 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002934 aarch64_srcs = AARCH64_ASM_UKERNELS,
2935)
2936
Marat Dukhan3b59de22020-06-03 20:15:19 -07002937xnnpack_cc_library(
2938 name = "logging_utils",
2939 srcs = LOGGING_SRCS,
2940 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2941 copts = LOGGING_COPTS + [
2942 "-Isrc",
2943 "-Iinclude",
2944 ] + select({
2945 ":debug_build": [],
2946 "//conditions:default": xnnpack_min_size_copts(),
2947 }),
2948 gcc_copts = xnnpack_gcc_std_copts(),
2949 msvc_copts = xnnpack_msvc_std_copts(),
2950 visibility = xnnpack_visibility(),
2951 deps = [
2952 "@FP16",
2953 "@clog",
2954 "@pthreadpool",
2955 ],
2956)
2957
Marat Dukhan08c4a432019-10-03 09:29:21 -07002958xnnpack_aggregate_library(
2959 name = "ukernels",
2960 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002961 ":neon_ukernels",
2962 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002963 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002964 ":asm_ukernels",
2965 ],
2966 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002967 ":neon_ukernels",
2968 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002969 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002970 ":neonfp16arith_ukernels",
2971 ":asm_ukernels",
2972 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002973 generic_deps = [
2974 ":scalar_ukernels",
2975 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002976 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002977 ":psimd_fastmath_ukernels",
2978 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002979 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002980 wasm_deps = [
2981 ":wasm_ukernels",
2982 ],
2983 wasmsimd_deps = [
2984 ":wasm_ukernels",
2985 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002986 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002987 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002988 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002989 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002990 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002991 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002992 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002993 ":avx512f_ukernels",
2994 ],
2995)
2996
Marat Dukhan33fcf782020-05-24 14:27:15 -07002997xnnpack_aggregate_library(
2998 name = "ukernels_test_mode",
2999 aarch32_deps = [
3000 ":neon_ukernels_test_mode",
3001 ":neonfma_ukernels_test_mode",
3002 ":neonv8_ukernels_test_mode",
3003 ":asm_ukernels",
3004 ],
3005 aarch64_deps = [
3006 ":neon_ukernels_test_mode",
3007 ":neonfma_ukernels_test_mode",
3008 ":neonv8_ukernels_test_mode",
3009 ":neonfp16arith_ukernels_test_mode",
3010 ":asm_ukernels",
3011 ],
3012 generic_deps = [
3013 ":scalar_ukernels_test_mode",
3014 ],
3015 psimd_deps = [
3016 ":psimd_fastmath_ukernels_test_mode",
3017 ":psimd_accmath_ukernels_test_mode",
3018 ],
3019 wasm_deps = [
3020 ":wasm_ukernels_test_mode",
3021 ],
3022 wasmsimd_deps = [
3023 ":wasm_ukernels_test_mode",
3024 ],
3025 x86_deps = [
3026 ":sse2_ukernels_test_mode",
3027 ":ssse3_ukernels_test_mode",
3028 ":sse41_ukernels_test_mode",
3029 ":avx_ukernels_test_mode",
3030 ":fma3_ukernels_test_mode",
3031 ":avx2_ukernels_test_mode",
3032 ":avx512f_ukernels_test_mode",
3033 ],
3034)
3035
Marat Dukhan86df3842020-06-17 14:03:31 -07003036# This target must be compatible with Android NDK r17 and compile with Clang 6.
3037# It is similar to :ukernels target, but excludes microkernels which require
3038# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003039xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003040 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003041 aarch32_deps = [
3042 ":neon_ukernels",
3043 ":neonfma_ukernels",
3044 ":neonv8_ukernels",
3045 ":asm_ukernels",
3046 ],
3047 aarch64_deps = [
3048 ":neon_ukernels",
3049 ":neonfma_ukernels",
3050 ":neonv8_ukernels",
3051 ":asm_ukernels",
3052 ],
3053 generic_deps = [
3054 ":scalar_ukernels",
3055 ],
3056 psimd_deps = [
3057 ":psimd_fastmath_ukernels",
3058 ":psimd_accmath_ukernels",
3059 ],
3060 wasm_deps = [
3061 ":wasm_ukernels",
3062 ],
3063 wasmsimd_deps = [
3064 ":wasm_ukernels",
3065 ],
3066 x86_deps = [
3067 ":sse2_ukernels",
3068 ":ssse3_ukernels",
3069 ":sse41_ukernels",
3070 ":avx_ukernels",
3071 ":fma3_ukernels",
3072 ":avx2_ukernels",
3073 ":avx512f_ukernels",
3074 ],
3075)
3076
Marat Dukhan08c4a432019-10-03 09:29:21 -07003077xnnpack_cc_library(
3078 name = "im2col",
3079 srcs = ["src/im2col.c"],
3080 hdrs = [
3081 "src/xnnpack/common.h",
3082 "src/xnnpack/im2col.h",
3083 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003084 gcc_copts = xnnpack_gcc_std_copts(),
3085 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003086)
3087
3088xnnpack_cc_library(
3089 name = "indirection",
3090 srcs = ["src/indirection.c"],
3091 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003092 gcc_copts = xnnpack_gcc_std_copts(),
3093 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003094 deps = [
3095 "@FP16",
3096 "@FXdiv",
3097 "@pthreadpool",
3098 ],
3099)
3100
3101xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003102 name = "indirection_test_mode",
3103 srcs = ["src/indirection.c"],
3104 hdrs = INTERNAL_HDRS,
3105 copts = [
3106 "-UNDEBUG",
3107 "-DXNN_TEST_MODE=1",
3108 ],
3109 gcc_copts = xnnpack_gcc_std_copts(),
3110 msvc_copts = xnnpack_msvc_std_copts(),
3111 deps = [
3112 "@FP16",
3113 "@FXdiv",
3114 "@pthreadpool",
3115 ],
3116)
3117
3118xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003119 name = "operator_run",
3120 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003121 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003122 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003123 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3124 "//conditions:default": [],
3125 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003126 gcc_copts = xnnpack_gcc_std_copts(),
3127 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003128 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003129 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003130 "@FP16",
3131 "@FXdiv",
3132 "@clog",
3133 "@pthreadpool",
3134 ],
3135)
3136
Chao Mei6ddfc602020-05-13 22:29:36 -07003137xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003138 name = "operator_run_test_mode",
3139 srcs = ["src/operator-run.c"],
3140 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3141 copts = LOGGING_COPTS + [
3142 "-UNDEBUG",
3143 "-DXNN_TEST_MODE=1",
3144 ] + select({
3145 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3146 "//conditions:default": [],
3147 }),
3148 gcc_copts = xnnpack_gcc_std_copts(),
3149 msvc_copts = xnnpack_msvc_std_copts(),
3150 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003151 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003152 "@FP16",
3153 "@FXdiv",
3154 "@clog",
3155 "@pthreadpool",
3156 ],
3157)
3158
3159xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003160 name = "memory_planner",
3161 srcs = ["src/memory-planner.c"],
3162 hdrs = INTERNAL_HDRS,
3163 defines = select({
3164 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3165 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3166 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3167 }),
3168 gcc_copts = xnnpack_gcc_std_copts(),
3169 msvc_copts = xnnpack_msvc_std_copts(),
3170 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003171 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003172 "@pthreadpool",
3173 ],
3174)
3175
Marat Dukhan33fcf782020-05-24 14:27:15 -07003176xnnpack_cc_library(
3177 name = "memory_planner_test_mode",
3178 srcs = ["src/memory-planner.c"],
3179 hdrs = INTERNAL_HDRS,
3180 copts = [
3181 "-UNDEBUG",
3182 "-DXNN_TEST_MODE=1",
3183 ],
3184 defines = select({
3185 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3186 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3187 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3188 }),
3189 gcc_copts = xnnpack_gcc_std_copts(),
3190 msvc_copts = xnnpack_msvc_std_copts(),
3191 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003192 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003193 "@pthreadpool",
3194 ],
3195)
3196
Marat Dukhan08c4a432019-10-03 09:29:21 -07003197cc_library(
3198 name = "enable_assembly",
3199 defines = select({
3200 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3201 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003202 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003203 }),
3204)
3205
Marat Dukhan9de90e02020-06-18 16:04:12 -07003206cc_library(
3207 name = "enable_sparse",
3208 defines = select({
3209 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3210 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003211 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003212 }),
3213)
3214
Marat Dukhancf056b22019-10-07 10:26:29 -07003215xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003216 name = "operators",
3217 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003218 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003219 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003220 ],
3221 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003222 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003223 "-Isrc",
3224 "-Iinclude",
3225 ] + select({
3226 ":debug_build": [],
3227 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003228 }) + select({
3229 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3230 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003231 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003232 gcc_copts = xnnpack_gcc_std_copts(),
3233 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003234 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003235 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003236 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003237 "@FP16",
3238 "@FXdiv",
3239 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003240 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003241 ],
3242)
3243
Marat Dukhan10a38082020-04-17 03:58:35 -07003244xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003245 name = "operators_test_mode",
3246 srcs = OPERATOR_SRCS + [
3247 "src/memory.c",
3248 "src/operator-delete.c",
3249 ],
3250 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3251 copts = LOGGING_COPTS + [
3252 "-Isrc",
3253 "-Iinclude",
3254 "-UNDEBUG",
3255 "-DXNN_TEST_MODE=1",
3256 ] + select({
3257 ":debug_build": [],
3258 "//conditions:default": xnnpack_min_size_copts(),
3259 }) + select({
3260 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3261 "//conditions:default": [],
3262 }),
3263 gcc_copts = xnnpack_gcc_std_copts(),
3264 msvc_copts = xnnpack_msvc_std_copts(),
3265 deps = [
3266 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003267 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003268 "@FP16",
3269 "@FXdiv",
3270 "@clog",
3271 "@pthreadpool",
3272 ],
3273)
3274
3275xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003276 name = "XNNPACK",
3277 srcs = [
3278 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003279 "src/runtime.c",
3280 "src/subgraph.c",
3281 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003282 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003283 hdrs = ["include/xnnpack.h"],
3284 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003285 "-Isrc",
3286 "-Iinclude",
3287 ] + select({
3288 ":debug_build": [],
3289 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003290 }) + select({
3291 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3292 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003293 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003294 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003295 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003296 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003297 visibility = xnnpack_visibility(),
3298 deps = [
3299 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003300 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003301 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003302 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003303 ":operator_run",
3304 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003305 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003306 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003307 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003308 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003309 ] + select({
3310 ":emscripten": [],
3311 "//conditions:default": ["@cpuinfo"],
3312 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003313)
3314
Marat Dukhan10a38082020-04-17 03:58:35 -07003315xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003316 name = "XNNPACK_test_mode",
3317 srcs = [
3318 "src/init.c",
3319 "src/runtime.c",
3320 "src/subgraph.c",
3321 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003322 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003323 hdrs = ["include/xnnpack.h"],
3324 copts = LOGGING_COPTS + [
3325 "-Isrc",
3326 "-Iinclude",
3327 "-UNDEBUG",
3328 "-DXNN_TEST_MODE=1",
3329 ] + select({
3330 ":debug_build": [],
3331 "//conditions:default": xnnpack_min_size_copts(),
3332 }) + select({
3333 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3334 "//conditions:default": [],
3335 }),
3336 gcc_copts = xnnpack_gcc_std_copts(),
3337 includes = ["include"],
3338 msvc_copts = xnnpack_msvc_std_copts(),
3339 visibility = xnnpack_visibility(),
3340 deps = [
3341 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003342 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003343 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003344 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003345 ":operator_run_test_mode",
3346 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003347 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003348 "@clog",
3349 "@FP16",
3350 "@pthreadpool",
3351 ] + select({
3352 ":emscripten": [],
3353 "//conditions:default": ["@cpuinfo"],
3354 }),
3355)
3356
3357xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003358 name = "xnnpack_f32",
3359 srcs = [
3360 "src/init.c",
3361 "src/runtime.c",
3362 "src/subgraph.c",
3363 "src/tensor.c",
3364 ] + SUBGRAPH_SRCS,
3365 hdrs = ["include/xnnpack.h"],
3366 copts = LOGGING_COPTS + [
3367 "-Isrc",
3368 "-Iinclude",
3369 ] + select({
3370 ":debug_build": [],
3371 "//conditions:default": xnnpack_min_size_copts(),
3372 }) + select({
3373 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3374 "//conditions:default": [],
3375 }),
3376 defines = [
3377 "XNN_NO_Q8_OPERATORS",
3378 "XNN_NO_U8_OPERATORS",
3379 "XNN_NO_X8_OPERATORS",
3380 "XNN_NO_F16_OPERATORS",
3381 "XNN_NO_X16_OPERATORS",
3382 ],
3383 gcc_copts = xnnpack_gcc_std_copts(),
3384 includes = ["include"],
3385 msvc_copts = xnnpack_msvc_std_copts(),
3386 visibility = xnnpack_visibility(),
3387 deps = [
3388 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003389 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003390 ":logging_utils",
3391 ":memory_planner",
3392 ":operator_run",
3393 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003394 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003395 "@clog",
3396 "@FP16",
3397 "@pthreadpool",
3398 ] + select({
3399 ":emscripten": [],
3400 "//conditions:default": ["@cpuinfo"],
3401 }),
3402)
3403
3404xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003405 name = "xnnpack_operators_nhwc_f32",
3406 srcs = [
3407 "src/init.c",
3408 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003409 hdrs = ["include/xnnpack.h"],
3410 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003411 "-Isrc",
3412 "-Iinclude",
3413 ] + select({
3414 ":debug_build": [],
3415 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003416 }) + select({
3417 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3418 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003419 }),
3420 defines = [
3421 "XNN_NO_Q8_OPERATORS",
3422 "XNN_NO_U8_OPERATORS",
3423 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003424 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003425 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003426 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003427 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003428 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003429 visibility = xnnpack_visibility(),
3430 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003431 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003432 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003433 ":operator_run",
3434 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003435 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003436 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003437 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003438 ] + select({
3439 ":emscripten": [],
3440 "//conditions:default": ["@cpuinfo"],
3441 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003442)
3443
Marat Dukhancf056b22019-10-07 10:26:29 -07003444xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003445 name = "bench_utils",
3446 srcs = ["bench/utils.cc"],
3447 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003448 deps = [
3449 "@com_google_benchmark//:benchmark",
3450 "@cpuinfo",
3451 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003452)
3453
Frank Barchard7e955972019-10-11 10:34:25 -07003454######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003455
3456xnnpack_benchmark(
3457 name = "q8_gemm_bench",
3458 srcs = [
3459 "bench/gemm.h",
3460 "bench/q8-gemm.cc",
3461 "src/xnnpack/AlignedAllocator.h",
3462 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003463 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003464 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
3465)
3466
3467xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003468 name = "f16_igemm_bench",
3469 srcs = [
3470 "bench/f16-igemm.cc",
3471 "bench/conv.h",
3472 "bench/google/conv.h",
3473 "src/xnnpack/AlignedAllocator.h",
3474 ] + MICROKERNEL_BENCHMARK_HDRS,
3475 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3476)
3477
3478xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003479 name = "f16_gemm_bench",
3480 srcs = [
3481 "bench/f16-gemm.cc",
3482 "bench/gemm.h",
3483 "src/xnnpack/AlignedAllocator.h",
3484 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003485 deps = MICROKERNEL_BENCHMARK_DEPS,
3486)
3487
3488xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003489 name = "f16_spmm_bench",
3490 srcs = [
3491 "bench/f16-spmm.cc",
3492 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003493 "src/xnnpack/AlignedAllocator.h",
3494 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003495 deps = MICROKERNEL_BENCHMARK_DEPS,
3496)
3497
3498xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003499 name = "f32_igemm_bench",
3500 srcs = [
3501 "bench/f32-igemm.cc",
3502 "bench/conv.h",
3503 "src/xnnpack/AlignedAllocator.h",
3504 ] + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard7e955972019-10-11 10:34:25 -07003505 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003506)
3507
3508xnnpack_benchmark(
3509 name = "f32_conv_hwc_bench",
3510 srcs = [
3511 "bench/f32-conv-hwc.cc",
3512 "bench/dconv.h",
3513 "src/xnnpack/AlignedAllocator.h",
3514 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003515 deps = MICROKERNEL_BENCHMARK_DEPS,
3516)
3517
3518xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003519 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003520 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003521 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003522 "bench/dconv.h",
3523 "src/xnnpack/AlignedAllocator.h",
3524 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Erich Elsen563df5f2019-10-23 08:02:21 -07003525 deps = MICROKERNEL_BENCHMARK_DEPS,
3526)
3527
3528xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003529 name = "f16_dwconv_bench",
3530 srcs = [
3531 "bench/f16-dwconv.cc",
3532 "bench/dwconv.h",
3533 "bench/google/dwconv.h",
3534 "src/xnnpack/AlignedAllocator.h",
3535 ] + MICROKERNEL_BENCHMARK_HDRS,
3536 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3537)
3538
3539xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003540 name = "f32_dwconv_bench",
3541 srcs = [
3542 "bench/f32-dwconv.cc",
3543 "bench/dwconv.h",
3544 "src/xnnpack/AlignedAllocator.h",
3545 ] + MICROKERNEL_BENCHMARK_HDRS,
3546 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3547)
3548
3549xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003550 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003551 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003552 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003553 "bench/dwconv.h",
3554 "src/xnnpack/AlignedAllocator.h",
3555 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3556 deps = MICROKERNEL_BENCHMARK_DEPS + [":indirection"],
3557)
3558
3559xnnpack_benchmark(
3560 name = "f32_gemm_bench",
3561 srcs = [
3562 "bench/f32-gemm.cc",
3563 "bench/gemm.h",
3564 "src/xnnpack/AlignedAllocator.h",
3565 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003566 copts = xnnpack_optional_ruy_copts(),
Frank Barchard7e955972019-10-11 10:34:25 -07003567 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003568)
3569
3570xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003571 name = "f32_hswish_bench",
3572 srcs = [
3573 "bench/f32-hswish.cc",
3574 "src/xnnpack/AlignedAllocator.h",
3575 ] + MICROKERNEL_BENCHMARK_HDRS,
3576 deps = MICROKERNEL_BENCHMARK_DEPS,
3577)
3578
3579xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003580 name = "f32_raddexpminusmax_bench",
3581 srcs = [
3582 "bench/f32-raddexpminusmax.cc",
3583 "src/xnnpack/AlignedAllocator.h",
3584 ] + MICROKERNEL_BENCHMARK_HDRS,
3585 deps = MICROKERNEL_BENCHMARK_DEPS,
3586)
3587
3588xnnpack_benchmark(
3589 name = "f32_raddextexp_bench",
3590 srcs = [
3591 "bench/f32-raddextexp.cc",
3592 "src/xnnpack/AlignedAllocator.h",
3593 ] + MICROKERNEL_BENCHMARK_HDRS,
3594 deps = MICROKERNEL_BENCHMARK_DEPS,
3595)
3596
3597xnnpack_benchmark(
3598 name = "f32_raddstoreexpminusmax_bench",
3599 srcs = [
3600 "bench/f32-raddstoreexpminusmax.cc",
3601 "src/xnnpack/AlignedAllocator.h",
3602 ] + MICROKERNEL_BENCHMARK_HDRS,
3603 deps = MICROKERNEL_BENCHMARK_DEPS,
3604)
3605
3606xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003607 name = "f32_rmax_bench",
3608 srcs = [
3609 "bench/f32-rmax.cc",
3610 "src/xnnpack/AlignedAllocator.h",
3611 ] + MICROKERNEL_BENCHMARK_HDRS,
3612 deps = MICROKERNEL_BENCHMARK_DEPS,
3613)
3614
3615xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003616 name = "f32_sigmoid_bench",
3617 srcs = [
3618 "bench/f32-sigmoid.cc",
3619 "src/xnnpack/AlignedAllocator.h",
3620 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003621 deps = MICROKERNEL_BENCHMARK_DEPS,
3622)
3623
3624xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003625 name = "f32_spmm_bench",
3626 srcs = [
3627 "bench/f32-spmm.cc",
3628 "bench/gemm.h",
3629 "src/xnnpack/AlignedAllocator.h",
3630 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003631 deps = MICROKERNEL_BENCHMARK_DEPS,
3632)
3633
3634xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003635 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003636 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003637 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003638 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003639 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003640 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003641)
3642
3643xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003644 name = "f32_vscaleexpminusmax_bench",
3645 srcs = [
3646 "bench/f32-vscaleexpminusmax.cc",
3647 "src/xnnpack/AlignedAllocator.h",
3648 ] + MICROKERNEL_BENCHMARK_HDRS,
3649 deps = MICROKERNEL_BENCHMARK_DEPS,
3650)
3651
3652xnnpack_benchmark(
3653 name = "f32_vscaleextexp_bench",
3654 srcs = [
3655 "bench/f32-vscaleextexp.cc",
3656 "src/xnnpack/AlignedAllocator.h",
3657 ] + MICROKERNEL_BENCHMARK_HDRS,
3658 deps = MICROKERNEL_BENCHMARK_DEPS,
3659)
3660
3661xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003662 name = "f32_vsqrt_bench",
3663 srcs = [
3664 "bench/f32-vsqrt.cc",
3665 "src/xnnpack/AlignedAllocator.h",
3666 ] + MICROKERNEL_BENCHMARK_HDRS,
3667 deps = MICROKERNEL_BENCHMARK_DEPS,
3668)
3669
3670xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003671 name = "f32_im2col_gemm_bench",
3672 srcs = [
3673 "bench/f32-im2col-gemm.cc",
3674 "bench/conv.h",
3675 "src/xnnpack/AlignedAllocator.h",
3676 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3677 deps = MICROKERNEL_BENCHMARK_DEPS + [":im2col"],
3678)
3679
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003680xnnpack_benchmark(
3681 name = "requantization_bench",
3682 srcs = [
3683 "bench/requantization.cc",
3684 "src/xnnpack/requantization-stubs.h",
3685 "src/xnnpack/AlignedAllocator.h",
3686 ] + MICROKERNEL_BENCHMARK_HDRS,
3687 deps = MICROKERNEL_BENCHMARK_DEPS,
3688)
3689
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003690xnnpack_benchmark(
3691 name = "rounding_bench",
3692 srcs = [
3693 "bench/rounding.cc",
3694 "src/xnnpack/math-stubs.h",
3695 "src/xnnpack/AlignedAllocator.h",
3696 ] + MICROKERNEL_BENCHMARK_HDRS,
3697 deps = MICROKERNEL_BENCHMARK_DEPS,
3698)
3699
Marat Dukhan08c4a432019-10-03 09:29:21 -07003700########################### Benchmarks for operators ###########################
3701
3702xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003703 name = "average_pooling_bench",
3704 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003705 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003706 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003707 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003708)
3709
3710xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003711 name = "bankers_rounding_bench",
3712 srcs = ["bench/bankers-rounding.cc"],
3713 copts = xnnpack_optional_tflite_copts(),
3714 tags = ["nowin32"],
3715 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3716)
3717
3718xnnpack_benchmark(
3719 name = "ceiling_bench",
3720 srcs = ["bench/ceiling.cc"],
3721 copts = xnnpack_optional_tflite_copts(),
3722 tags = ["nowin32"],
3723 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3724)
3725
3726xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003727 name = "channel_shuffle_bench",
3728 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003729 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003730)
3731
3732xnnpack_benchmark(
3733 name = "convolution_bench",
3734 srcs = ["bench/convolution.cc"],
3735 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003736 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003737 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003738)
3739
3740xnnpack_benchmark(
3741 name = "deconvolution_bench",
3742 srcs = ["bench/deconvolution.cc"],
3743 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003744 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003745 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003746)
3747
3748xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003749 name = "floor_bench",
3750 srcs = ["bench/floor.cc"],
3751 copts = xnnpack_optional_tflite_copts(),
3752 tags = ["nowin32"],
3753 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3754)
3755
3756xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003757 name = "global_average_pooling_bench",
3758 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003759 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003760)
3761
3762xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07003763 name = "hardswish_bench",
3764 srcs = ["bench/hardswish.cc"],
3765 copts = xnnpack_optional_tflite_copts(),
3766 tags = ["nowin32"],
3767 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3768)
3769
3770xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003771 name = "max_pooling_bench",
3772 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003773 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003774)
3775
3776xnnpack_benchmark(
3777 name = "sigmoid_bench",
3778 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003779 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003780 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003781 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003782)
3783
3784xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003785 name = "prelu_bench",
3786 srcs = ["bench/prelu.cc"],
3787 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003788 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003789 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003790)
3791
3792xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003793 name = "softmax_bench",
3794 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003795 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003796 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003797 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003798)
3799
Marat Dukhan87727142020-06-24 15:24:10 -07003800xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07003801 name = "square_root_bench",
3802 srcs = ["bench/square-root.cc"],
3803 copts = xnnpack_optional_tflite_copts(),
3804 tags = ["nowin32"],
3805 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3806)
3807
3808xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003809 name = "truncation_bench",
3810 srcs = ["bench/truncation.cc"],
3811 deps = OPERATOR_BENCHMARK_DEPS,
3812)
3813
Marat Dukhanc068bb62019-10-04 13:24:39 -07003814############################# End-to-end benchmarks ############################
3815
3816cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003817 name = "fp32_mobilenet_v1",
3818 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003819 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003820 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003821 linkstatic = True,
3822 deps = [
3823 ":XNNPACK",
3824 "@pthreadpool",
3825 ],
3826)
3827
3828cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003829 name = "fp16_mobilenet_v1",
3830 srcs = ["models/fp16-mobilenet-v1.cc"],
3831 hdrs = ["models/models.h"],
3832 copts = xnnpack_std_cxxopts(),
3833 linkstatic = True,
3834 deps = [
3835 ":XNNPACK",
3836 "@FP16",
3837 "@pthreadpool",
3838 ],
3839)
3840
3841cc_library(
3842 name = "fp32_mobilenet_v2",
3843 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003844 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003845 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003846 linkstatic = True,
3847 deps = [
3848 ":XNNPACK",
3849 "@pthreadpool",
3850 ],
3851)
3852
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003853cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003854 name = "fp16_mobilenet_v2",
3855 srcs = ["models/fp16-mobilenet-v2.cc"],
3856 hdrs = ["models/models.h"],
3857 copts = xnnpack_std_cxxopts(),
3858 linkstatic = True,
3859 deps = [
3860 ":XNNPACK",
3861 "@FP16",
3862 "@pthreadpool",
3863 ],
3864)
3865
3866cc_library(
3867 name = "fp32_mobilenet_v3_large",
3868 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003869 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003870 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003871 linkstatic = True,
3872 deps = [
3873 ":XNNPACK",
3874 "@pthreadpool",
3875 ],
3876)
3877
3878cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003879 name = "fp32_mobilenet_v3_small",
3880 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003881 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003882 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003883 linkstatic = True,
3884 deps = [
3885 ":XNNPACK",
3886 "@pthreadpool",
3887 ],
3888)
3889
Marat Dukhanc068bb62019-10-04 13:24:39 -07003890xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07003891 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003892 srcs = [
3893 "bench/f32-dwconv-e2e.cc",
3894 "bench/end2end.h",
3895 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07003896 deps = MICROKERNEL_BENCHMARK_DEPS + [
3897 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003898 ":fp32_mobilenet_v1",
3899 ":fp32_mobilenet_v2",
3900 ":fp32_mobilenet_v3_large",
3901 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07003902 ],
3903)
3904
3905xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07003906 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003907 srcs = [
3908 "bench/f32-gemm-e2e.cc",
3909 "bench/end2end.h",
3910 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07003911 deps = MICROKERNEL_BENCHMARK_DEPS + [
3912 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003913 ":fp32_mobilenet_v1",
3914 ":fp32_mobilenet_v2",
3915 ":fp32_mobilenet_v3_large",
3916 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07003917 ],
3918)
3919
3920xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07003921 name = "end2end_bench",
3922 srcs = ["bench/end2end.cc"],
3923 deps = [
3924 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07003925 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003926 ":fp16_mobilenet_v1",
3927 ":fp16_mobilenet_v2",
3928 ":fp32_mobilenet_v1",
3929 ":fp32_mobilenet_v2",
3930 ":fp32_mobilenet_v3_large",
3931 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07003932 "@pthreadpool",
3933 ],
3934)
3935
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003936#################### Accuracy evaluation for math functions ####################
3937
3938xnnpack_benchmark(
3939 name = "f32_exp_eval",
3940 srcs = [
3941 "eval/f32-exp.cc",
3942 "src/xnnpack/AlignedAllocator.h",
3943 ] + ACCURACY_EVAL_HDRS,
3944 deps = ACCURACY_EVAL_DEPS,
3945)
3946
Marat Dukhan515c9772019-10-17 18:07:57 -07003947xnnpack_benchmark(
3948 name = "f32_expminus_eval",
3949 srcs = [
3950 "eval/f32-expminus.cc",
3951 "src/xnnpack/AlignedAllocator.h",
3952 ] + ACCURACY_EVAL_HDRS,
3953 deps = ACCURACY_EVAL_DEPS,
3954)
3955
Marat Dukhan98ba4412019-10-23 02:14:28 -07003956xnnpack_benchmark(
3957 name = "f32_extexp_eval",
3958 srcs = [
3959 "eval/f32-extexp.cc",
3960 "src/xnnpack/AlignedAllocator.h",
3961 ] + ACCURACY_EVAL_HDRS,
3962 deps = ACCURACY_EVAL_DEPS,
3963)
3964
Marat Dukhan8853b822020-05-07 12:19:01 -07003965xnnpack_unit_test(
3966 name = "f32_roundne_eval",
3967 srcs = [
3968 "eval/f32-roundne.cc",
3969 "src/xnnpack/AlignedAllocator.h",
3970 "src/xnnpack/math-stubs.h",
3971 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07003972 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07003973 deps = MICROKERNEL_TEST_DEPS,
3974)
3975
Marat Dukhan2dbb9442020-05-12 20:43:43 -07003976xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07003977 name = "f32_roundd_eval",
3978 srcs = [
3979 "eval/f32-roundd.cc",
3980 "src/xnnpack/AlignedAllocator.h",
3981 "src/xnnpack/math-stubs.h",
3982 ] + MICROKERNEL_TEST_HDRS,
3983 automatic = False,
3984 deps = MICROKERNEL_TEST_DEPS,
3985)
3986
3987xnnpack_unit_test(
3988 name = "f32_roundu_eval",
3989 srcs = [
3990 "eval/f32-roundu.cc",
3991 "src/xnnpack/AlignedAllocator.h",
3992 "src/xnnpack/math-stubs.h",
3993 ] + MICROKERNEL_TEST_HDRS,
3994 automatic = False,
3995 deps = MICROKERNEL_TEST_DEPS,
3996)
3997
3998xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07003999 name = "f32_roundz_eval",
4000 srcs = [
4001 "eval/f32-roundz.cc",
4002 "src/xnnpack/AlignedAllocator.h",
4003 "src/xnnpack/math-stubs.h",
4004 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004005 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004006 deps = MICROKERNEL_TEST_DEPS,
4007)
4008
Marat Dukhan346a9e52019-11-15 09:06:30 -08004009xnnpack_benchmark(
4010 name = "f32_sigmoid_eval",
4011 srcs = [
4012 "eval/f32-sigmoid.cc",
4013 "src/xnnpack/AlignedAllocator.h",
4014 ] + ACCURACY_EVAL_HDRS,
4015 deps = ACCURACY_EVAL_DEPS,
4016)
4017
Marat Dukhan84000762020-06-29 18:38:43 -07004018xnnpack_benchmark(
4019 name = "f32_sqrt_eval",
4020 srcs = [
4021 "eval/f32-sqrt.cc",
4022 "src/xnnpack/AlignedAllocator.h",
4023 ] + ACCURACY_EVAL_HDRS,
4024 deps = ACCURACY_EVAL_DEPS + [
4025 ":bench_utils",
4026 ],
4027)
4028
Marat Dukhan08c4a432019-10-03 09:29:21 -07004029######################### Unit tests for micro-kernels #########################
4030
4031xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004032 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004033 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004034 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004035 "test/gemm-microkernel-tester.h",
4036 "src/xnnpack/AlignedAllocator.h",
4037 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4038 deps = MICROKERNEL_TEST_DEPS,
4039)
4040
4041xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004042 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004043 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004044 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004045 "test/spmm-microkernel-tester.h",
4046 "src/xnnpack/AlignedAllocator.h",
4047 ] + MICROKERNEL_TEST_HDRS,
4048 deps = MICROKERNEL_TEST_DEPS,
4049)
4050
4051xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004052 name = "f32_argmaxpool_test",
4053 srcs = [
4054 "test/f32-argmaxpool.cc",
4055 "test/argmaxpool-microkernel-tester.h",
4056 "src/xnnpack/AlignedAllocator.h",
4057 ] + MICROKERNEL_TEST_HDRS,
4058 deps = MICROKERNEL_TEST_DEPS,
4059)
4060
4061xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004062 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004063 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004064 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004065 "test/avgpool-microkernel-tester.h",
4066 "src/xnnpack/AlignedAllocator.h",
4067 ] + MICROKERNEL_TEST_HDRS,
4068 deps = MICROKERNEL_TEST_DEPS,
4069)
4070
4071xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004072 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004073 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004074 "test/f32-ibilinear.cc",
4075 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004076 "src/xnnpack/AlignedAllocator.h",
4077 ] + MICROKERNEL_TEST_HDRS,
4078 deps = MICROKERNEL_TEST_DEPS,
4079)
4080
4081xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004082 name = "f16_clamp_test",
4083 srcs = [
4084 "test/f16-clamp.cc",
4085 "test/clamp-microkernel-tester.h",
4086 ] + MICROKERNEL_TEST_HDRS,
4087 deps = MICROKERNEL_TEST_DEPS,
4088)
4089
4090xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004091 name = "f32_clamp_test",
4092 srcs = [
4093 "test/f32-clamp.cc",
4094 "test/clamp-microkernel-tester.h",
4095 ] + MICROKERNEL_TEST_HDRS,
4096 deps = MICROKERNEL_TEST_DEPS,
4097)
4098
4099xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004100 name = "f32_igemm_test",
4101 srcs = [
4102 "test/f32-igemm.cc",
4103 "test/gemm-microkernel-tester.h",
4104 "src/xnnpack/AlignedAllocator.h",
4105 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4106 deps = MICROKERNEL_TEST_DEPS,
4107)
4108
4109xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004110 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004111 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004112 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004113 "test/gemm-microkernel-tester.h",
4114 "src/xnnpack/AlignedAllocator.h",
4115 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4116 deps = MICROKERNEL_TEST_DEPS,
4117)
4118
4119xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004120 name = "f32_igemm_minmax_test",
4121 srcs = [
4122 "test/f32-igemm-minmax.cc",
4123 "test/gemm-microkernel-tester.h",
4124 "src/xnnpack/AlignedAllocator.h",
4125 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4126 deps = MICROKERNEL_TEST_DEPS,
4127)
4128
4129xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004130 name = "f16_igemm_minmax_test",
4131 srcs = [
4132 "test/f16-igemm-minmax.cc",
4133 "test/gemm-microkernel-tester.h",
4134 "src/xnnpack/AlignedAllocator.h",
4135 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4136 deps = MICROKERNEL_TEST_DEPS,
4137)
4138
4139xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004140 name = "f32_conv_hwc_test",
4141 srcs = [
4142 "test/f32-conv-hwc.cc",
4143 "test/conv-hwc-microkernel-tester.h",
4144 "src/xnnpack/AlignedAllocator.h",
4145 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4146 deps = MICROKERNEL_TEST_DEPS,
4147)
4148
4149xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004150 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004151 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004152 "test/f32-conv-hwc2chw.cc",
4153 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004154 "src/xnnpack/AlignedAllocator.h",
4155 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4156 deps = MICROKERNEL_TEST_DEPS,
4157)
4158
4159xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004160 name = "f32_dwconv_test",
4161 srcs = [
4162 "test/f32-dwconv.cc",
4163 "test/dwconv-microkernel-tester.h",
4164 "src/xnnpack/AlignedAllocator.h",
4165 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4166 deps = MICROKERNEL_TEST_DEPS,
4167)
4168
4169xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004170 name = "f16_dwconv_minmax_test",
4171 srcs = [
4172 "test/f16-dwconv-minmax.cc",
4173 "test/dwconv-microkernel-tester.h",
4174 "src/xnnpack/AlignedAllocator.h",
4175 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4176 deps = MICROKERNEL_TEST_DEPS,
4177)
4178
4179xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004180 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004181 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004182 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004183 "test/dwconv-microkernel-tester.h",
4184 "src/xnnpack/AlignedAllocator.h",
4185 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4186 deps = MICROKERNEL_TEST_DEPS,
4187)
4188
4189xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004190 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004191 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004192 "test/f32-dwconv-chw.cc",
4193 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004194 "src/xnnpack/AlignedAllocator.h",
4195 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4196 deps = MICROKERNEL_TEST_DEPS,
4197)
4198
4199xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004200 name = "f16_gavgpool_minmax_test",
4201 srcs = [
4202 "test/f16-gavgpool-minmax.cc",
4203 "test/gavgpool-microkernel-tester.h",
4204 "src/xnnpack/AlignedAllocator.h",
4205 ] + MICROKERNEL_TEST_HDRS,
4206 deps = MICROKERNEL_TEST_DEPS,
4207)
4208
4209xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004210 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004211 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004212 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004213 "test/gavgpool-microkernel-tester.h",
4214 "src/xnnpack/AlignedAllocator.h",
4215 ] + MICROKERNEL_TEST_HDRS,
4216 deps = MICROKERNEL_TEST_DEPS,
4217)
4218
4219xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004220 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004221 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004222 "test/f32-gavgpool-cw.cc",
4223 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004224 "src/xnnpack/AlignedAllocator.h",
4225 ] + MICROKERNEL_TEST_HDRS,
4226 deps = MICROKERNEL_TEST_DEPS,
4227)
4228
4229xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004230 name = "f32_gemm_test",
4231 srcs = [
4232 "test/f32-gemm.cc",
4233 "test/gemm-microkernel-tester.h",
4234 "src/xnnpack/AlignedAllocator.h",
4235 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4236 deps = MICROKERNEL_TEST_DEPS,
4237)
4238
4239xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004240 name = "f32_gemm_relu_test",
4241 srcs = [
4242 "test/f32-gemm-relu.cc",
4243 "test/gemm-microkernel-tester.h",
4244 "src/xnnpack/AlignedAllocator.h",
4245 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4246 deps = MICROKERNEL_TEST_DEPS,
4247)
4248
4249xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004250 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004251 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004252 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004253 "test/gemm-microkernel-tester.h",
4254 "src/xnnpack/AlignedAllocator.h",
4255 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4256 deps = MICROKERNEL_TEST_DEPS,
4257)
4258
4259xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004260 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004261 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004262 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004263 "test/gemm-microkernel-tester.h",
4264 "src/xnnpack/AlignedAllocator.h",
4265 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4266 deps = MICROKERNEL_TEST_DEPS,
4267)
4268
4269xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004270 name = "f16_hswish_test",
4271 srcs = [
4272 "test/f16-hswish.cc",
4273 "test/hswish-microkernel-tester.h",
4274 ] + MICROKERNEL_TEST_HDRS,
4275 deps = MICROKERNEL_TEST_DEPS,
4276)
4277
4278xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004279 name = "f32_hswish_test",
4280 srcs = [
4281 "test/f32-hswish.cc",
4282 "test/hswish-microkernel-tester.h",
4283 ] + MICROKERNEL_TEST_HDRS,
4284 deps = MICROKERNEL_TEST_DEPS,
4285)
4286
4287xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004288 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004289 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004290 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004291 "test/maxpool-microkernel-tester.h",
4292 ] + MICROKERNEL_TEST_HDRS,
4293 deps = MICROKERNEL_TEST_DEPS,
4294)
4295
4296xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004297 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004298 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004299 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004300 "test/avgpool-microkernel-tester.h",
4301 "src/xnnpack/AlignedAllocator.h",
4302 ] + MICROKERNEL_TEST_HDRS,
4303 deps = MICROKERNEL_TEST_DEPS,
4304)
4305
4306xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004307 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004308 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004309 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004310 "test/gemm-microkernel-tester.h",
4311 "src/xnnpack/AlignedAllocator.h",
4312 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4313 deps = MICROKERNEL_TEST_DEPS,
4314)
4315
4316xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004317 name = "f16_prelu_test",
4318 srcs = [
4319 "test/f16-prelu.cc",
4320 "test/prelu-microkernel-tester.h",
4321 "src/xnnpack/AlignedAllocator.h",
4322 ] + MICROKERNEL_TEST_HDRS,
4323 deps = MICROKERNEL_TEST_DEPS,
4324)
4325
4326xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004327 name = "f32_prelu_test",
4328 srcs = [
4329 "test/f32-prelu.cc",
4330 "test/prelu-microkernel-tester.h",
4331 "src/xnnpack/AlignedAllocator.h",
4332 ] + MICROKERNEL_TEST_HDRS,
4333 deps = MICROKERNEL_TEST_DEPS,
4334)
4335
4336xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004337 name = "f32_raddexpminusmax_test",
4338 srcs = [
4339 "test/f32-raddexpminusmax.cc",
4340 "test/raddexpminusmax-microkernel-tester.h",
4341 ] + MICROKERNEL_TEST_HDRS,
4342 deps = MICROKERNEL_TEST_DEPS,
4343)
4344
4345xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004346 name = "f32_raddextexp_test",
4347 srcs = [
4348 "test/f32-raddextexp.cc",
4349 "test/raddextexp-microkernel-tester.h",
4350 ] + MICROKERNEL_TEST_HDRS,
4351 deps = MICROKERNEL_TEST_DEPS,
4352)
4353
4354xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004355 name = "f32_raddstoreexpminusmax_test",
4356 srcs = [
4357 "test/f32-raddstoreexpminusmax.cc",
4358 "test/raddstoreexpminusmax-microkernel-tester.h",
4359 ] + MICROKERNEL_TEST_HDRS,
4360 deps = MICROKERNEL_TEST_DEPS,
4361)
4362
4363xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004364 name = "f32_rmax_test",
4365 srcs = [
4366 "test/f32-rmax.cc",
4367 "test/rmax-microkernel-tester.h",
4368 ] + MICROKERNEL_TEST_HDRS,
4369 deps = MICROKERNEL_TEST_DEPS,
4370)
4371
4372xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004373 name = "f32_sigmoid_test",
4374 srcs = [
4375 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004376 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004377 ] + MICROKERNEL_TEST_HDRS,
4378 deps = MICROKERNEL_TEST_DEPS,
4379)
4380
4381xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004382 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004383 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004384 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004385 "test/spmm-microkernel-tester.h",
4386 "src/xnnpack/AlignedAllocator.h",
4387 ] + MICROKERNEL_TEST_HDRS,
4388 deps = MICROKERNEL_TEST_DEPS,
4389)
4390
4391xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004392 name = "f16_vadd_minmax_test",
4393 srcs = [
4394 "test/f16-vadd-minmax.cc",
4395 "test/vbinary-microkernel-tester.h",
4396 ] + MICROKERNEL_TEST_HDRS,
4397 deps = MICROKERNEL_TEST_DEPS,
4398)
4399
4400xnnpack_unit_test(
4401 name = "f16_vaddc_minmax_test",
4402 srcs = [
4403 "test/f16-vaddc-minmax.cc",
4404 "test/vbinaryc-microkernel-tester.h",
4405 ] + MICROKERNEL_TEST_HDRS,
4406 deps = MICROKERNEL_TEST_DEPS,
4407)
4408
4409xnnpack_unit_test(
4410 name = "f16_vdiv_minmax_test",
4411 srcs = [
4412 "test/f16-vdiv-minmax.cc",
4413 "test/vbinary-microkernel-tester.h",
4414 ] + MICROKERNEL_TEST_HDRS,
4415 deps = MICROKERNEL_TEST_DEPS,
4416)
4417
4418xnnpack_unit_test(
4419 name = "f16_vdivc_minmax_test",
4420 srcs = [
4421 "test/f16-vdivc-minmax.cc",
4422 "test/vbinaryc-microkernel-tester.h",
4423 ] + MICROKERNEL_TEST_HDRS,
4424 deps = MICROKERNEL_TEST_DEPS,
4425)
4426
4427xnnpack_unit_test(
4428 name = "f16_vrdivc_minmax_test",
4429 srcs = [
4430 "test/f16-vrdivc-minmax.cc",
4431 "test/vbinaryc-microkernel-tester.h",
4432 ] + MICROKERNEL_TEST_HDRS,
4433 deps = MICROKERNEL_TEST_DEPS,
4434)
4435
4436xnnpack_unit_test(
4437 name = "f16_vmax_test",
4438 srcs = [
4439 "test/f16-vmax.cc",
4440 "test/vbinary-microkernel-tester.h",
4441 ] + MICROKERNEL_TEST_HDRS,
4442 deps = MICROKERNEL_TEST_DEPS,
4443)
4444
4445xnnpack_unit_test(
4446 name = "f16_vmaxc_test",
4447 srcs = [
4448 "test/f16-vmaxc.cc",
4449 "test/vbinaryc-microkernel-tester.h",
4450 ] + MICROKERNEL_TEST_HDRS,
4451 deps = MICROKERNEL_TEST_DEPS,
4452)
4453
4454xnnpack_unit_test(
4455 name = "f16_vmin_test",
4456 srcs = [
4457 "test/f16-vmin.cc",
4458 "test/vbinary-microkernel-tester.h",
4459 ] + MICROKERNEL_TEST_HDRS,
4460 deps = MICROKERNEL_TEST_DEPS,
4461)
4462
4463xnnpack_unit_test(
4464 name = "f16_vminc_test",
4465 srcs = [
4466 "test/f16-vminc.cc",
4467 "test/vbinaryc-microkernel-tester.h",
4468 ] + MICROKERNEL_TEST_HDRS,
4469 deps = MICROKERNEL_TEST_DEPS,
4470)
4471
4472xnnpack_unit_test(
4473 name = "f16_vmul_minmax_test",
4474 srcs = [
4475 "test/f16-vmul-minmax.cc",
4476 "test/vbinary-microkernel-tester.h",
4477 ] + MICROKERNEL_TEST_HDRS,
4478 deps = MICROKERNEL_TEST_DEPS,
4479)
4480
4481xnnpack_unit_test(
4482 name = "f16_vmulc_minmax_test",
4483 srcs = [
4484 "test/f16-vmulc-minmax.cc",
4485 "test/vbinaryc-microkernel-tester.h",
4486 ] + MICROKERNEL_TEST_HDRS,
4487 deps = MICROKERNEL_TEST_DEPS,
4488)
4489
4490xnnpack_unit_test(
4491 name = "f16_vsub_minmax_test",
4492 srcs = [
4493 "test/f16-vsub-minmax.cc",
4494 "test/vbinary-microkernel-tester.h",
4495 ] + MICROKERNEL_TEST_HDRS,
4496 deps = MICROKERNEL_TEST_DEPS,
4497)
4498
4499xnnpack_unit_test(
4500 name = "f16_vsubc_minmax_test",
4501 srcs = [
4502 "test/f16-vsubc-minmax.cc",
4503 "test/vbinaryc-microkernel-tester.h",
4504 ] + MICROKERNEL_TEST_HDRS,
4505 deps = MICROKERNEL_TEST_DEPS,
4506)
4507
4508xnnpack_unit_test(
4509 name = "f16_vrsubc_minmax_test",
4510 srcs = [
4511 "test/f16-vrsubc-minmax.cc",
4512 "test/vbinaryc-microkernel-tester.h",
4513 ] + MICROKERNEL_TEST_HDRS,
4514 deps = MICROKERNEL_TEST_DEPS,
4515)
4516
4517xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004518 name = "f32_vabs_test",
4519 srcs = [
4520 "test/f32-vabs.cc",
4521 "test/vunary-microkernel-tester.h",
4522 ] + MICROKERNEL_TEST_HDRS,
4523 deps = MICROKERNEL_TEST_DEPS,
4524)
4525
4526xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004527 name = "f32_vadd_test",
4528 srcs = [
4529 "test/f32-vadd.cc",
4530 "test/vbinary-microkernel-tester.h",
4531 ] + MICROKERNEL_TEST_HDRS,
4532 deps = MICROKERNEL_TEST_DEPS,
4533)
4534
4535xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004536 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004537 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004538 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004539 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004540 ] + MICROKERNEL_TEST_HDRS,
4541 deps = MICROKERNEL_TEST_DEPS,
4542)
4543
4544xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004545 name = "f32_vaddc_test",
4546 srcs = [
4547 "test/f32-vaddc.cc",
4548 "test/vbinaryc-microkernel-tester.h",
4549 ] + MICROKERNEL_TEST_HDRS,
4550 deps = MICROKERNEL_TEST_DEPS,
4551)
4552
4553xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004554 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004555 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004556 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004557 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004558 ] + MICROKERNEL_TEST_HDRS,
4559 deps = MICROKERNEL_TEST_DEPS,
4560)
4561
4562xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004563 name = "f32_vdiv_test",
4564 srcs = [
4565 "test/f32-vdiv.cc",
4566 "test/vbinary-microkernel-tester.h",
4567 ] + MICROKERNEL_TEST_HDRS,
4568 deps = MICROKERNEL_TEST_DEPS,
4569)
4570
4571xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004572 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004573 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004574 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004575 "test/vbinary-microkernel-tester.h",
4576 ] + MICROKERNEL_TEST_HDRS,
4577 deps = MICROKERNEL_TEST_DEPS,
4578)
4579
4580xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004581 name = "f32_vdivc_test",
4582 srcs = [
4583 "test/f32-vdivc.cc",
4584 "test/vbinaryc-microkernel-tester.h",
4585 ] + MICROKERNEL_TEST_HDRS,
4586 deps = MICROKERNEL_TEST_DEPS,
4587)
4588
4589xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004590 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004591 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004592 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004593 "test/vbinaryc-microkernel-tester.h",
4594 ] + MICROKERNEL_TEST_HDRS,
4595 deps = MICROKERNEL_TEST_DEPS,
4596)
4597
4598xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004599 name = "f32_vrdivc_test",
4600 srcs = [
4601 "test/f32-vrdivc.cc",
4602 "test/vbinaryc-microkernel-tester.h",
4603 ] + MICROKERNEL_TEST_HDRS,
4604 deps = MICROKERNEL_TEST_DEPS,
4605)
4606
4607xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004608 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004609 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004610 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004611 "test/vbinaryc-microkernel-tester.h",
4612 ] + MICROKERNEL_TEST_HDRS,
4613 deps = MICROKERNEL_TEST_DEPS,
4614)
4615
4616xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004617 name = "f32_vmax_test",
4618 srcs = [
4619 "test/f32-vmax.cc",
4620 "test/vbinary-microkernel-tester.h",
4621 ] + MICROKERNEL_TEST_HDRS,
4622 deps = MICROKERNEL_TEST_DEPS,
4623)
4624
4625xnnpack_unit_test(
4626 name = "f32_vmaxc_test",
4627 srcs = [
4628 "test/f32-vmaxc.cc",
4629 "test/vbinaryc-microkernel-tester.h",
4630 ] + MICROKERNEL_TEST_HDRS,
4631 deps = MICROKERNEL_TEST_DEPS,
4632)
4633
4634xnnpack_unit_test(
4635 name = "f32_vmin_test",
4636 srcs = [
4637 "test/f32-vmin.cc",
4638 "test/vbinary-microkernel-tester.h",
4639 ] + MICROKERNEL_TEST_HDRS,
4640 deps = MICROKERNEL_TEST_DEPS,
4641)
4642
4643xnnpack_unit_test(
4644 name = "f32_vminc_test",
4645 srcs = [
4646 "test/f32-vminc.cc",
4647 "test/vbinaryc-microkernel-tester.h",
4648 ] + MICROKERNEL_TEST_HDRS,
4649 deps = MICROKERNEL_TEST_DEPS,
4650)
4651
4652xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004653 name = "f32_vmul_test",
4654 srcs = [
4655 "test/f32-vmul.cc",
4656 "test/vbinary-microkernel-tester.h",
4657 ] + MICROKERNEL_TEST_HDRS,
4658 deps = MICROKERNEL_TEST_DEPS,
4659)
4660
4661xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004662 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004663 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004664 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004665 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004666 ] + MICROKERNEL_TEST_HDRS,
4667 deps = MICROKERNEL_TEST_DEPS,
4668)
4669
4670xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004671 name = "f32_vmulc_test",
4672 srcs = [
4673 "test/f32-vmulc.cc",
4674 "test/vbinaryc-microkernel-tester.h",
4675 ] + MICROKERNEL_TEST_HDRS,
4676 deps = MICROKERNEL_TEST_DEPS,
4677)
4678
4679xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004680 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004681 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004682 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004683 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004684 ] + MICROKERNEL_TEST_HDRS,
4685 deps = MICROKERNEL_TEST_DEPS,
4686)
4687
4688xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004689 name = "f16_vmulcaddc_minmax_test",
4690 srcs = [
4691 "test/f16-vmulcaddc-minmax.cc",
4692 "test/vmulcaddc-microkernel-tester.h",
4693 "src/xnnpack/AlignedAllocator.h",
4694 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4695 deps = MICROKERNEL_TEST_DEPS,
4696)
4697
4698xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004699 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004700 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004701 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004702 "test/vmulcaddc-microkernel-tester.h",
4703 "src/xnnpack/AlignedAllocator.h",
4704 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4705 deps = MICROKERNEL_TEST_DEPS,
4706)
4707
4708xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004709 name = "f32_vlrelu_test",
4710 srcs = [
4711 "test/f32-vlrelu.cc",
4712 "test/vunary-microkernel-tester.h",
4713 ] + MICROKERNEL_TEST_HDRS,
4714 deps = MICROKERNEL_TEST_DEPS,
4715)
4716
4717xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004718 name = "f32_vneg_test",
4719 srcs = [
4720 "test/f32-vneg.cc",
4721 "test/vunary-microkernel-tester.h",
4722 ] + MICROKERNEL_TEST_HDRS,
4723 deps = MICROKERNEL_TEST_DEPS,
4724)
4725
4726xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004727 name = "f32_vrndne_test",
4728 srcs = [
4729 "test/f32-vrndne.cc",
4730 "test/vunary-microkernel-tester.h",
4731 ] + MICROKERNEL_TEST_HDRS,
4732 deps = MICROKERNEL_TEST_DEPS,
4733)
4734
4735xnnpack_unit_test(
4736 name = "f32_vrndz_test",
4737 srcs = [
4738 "test/f32-vrndz.cc",
4739 "test/vunary-microkernel-tester.h",
4740 ] + MICROKERNEL_TEST_HDRS,
4741 deps = MICROKERNEL_TEST_DEPS,
4742)
4743
4744xnnpack_unit_test(
4745 name = "f32_vrndu_test",
4746 srcs = [
4747 "test/f32-vrndu.cc",
4748 "test/vunary-microkernel-tester.h",
4749 ] + MICROKERNEL_TEST_HDRS,
4750 deps = MICROKERNEL_TEST_DEPS,
4751)
4752
4753xnnpack_unit_test(
4754 name = "f32_vrndd_test",
4755 srcs = [
4756 "test/f32-vrndd.cc",
4757 "test/vunary-microkernel-tester.h",
4758 ] + MICROKERNEL_TEST_HDRS,
4759 deps = MICROKERNEL_TEST_DEPS,
4760)
4761
4762xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004763 name = "f32_vscale_test",
4764 srcs = [
4765 "test/f32-vscale.cc",
4766 "test/vscale-microkernel-tester.h",
4767 ] + MICROKERNEL_TEST_HDRS,
4768 deps = MICROKERNEL_TEST_DEPS,
4769)
4770
4771xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004772 name = "f32_vscaleexpminusmax_test",
4773 srcs = [
4774 "test/f32-vscaleexpminusmax.cc",
4775 "test/vscaleexpminusmax-microkernel-tester.h",
4776 ] + MICROKERNEL_TEST_HDRS,
4777 deps = MICROKERNEL_TEST_DEPS,
4778)
4779
4780xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004781 name = "f32_vscaleextexp_test",
4782 srcs = [
4783 "test/f32-vscaleextexp.cc",
4784 "test/vscaleextexp-microkernel-tester.h",
4785 ] + MICROKERNEL_TEST_HDRS,
4786 deps = MICROKERNEL_TEST_DEPS,
4787)
4788
4789xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004790 name = "f32_vsqr_test",
4791 srcs = [
4792 "test/f32-vsqr.cc",
4793 "test/vunary-microkernel-tester.h",
4794 ] + MICROKERNEL_TEST_HDRS,
4795 deps = MICROKERNEL_TEST_DEPS,
4796)
4797
4798xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004799 name = "f32_vsqrdiff_test",
4800 srcs = [
4801 "test/f32-vsqrdiff.cc",
4802 "test/vbinary-microkernel-tester.h",
4803 ] + MICROKERNEL_TEST_HDRS,
4804 deps = MICROKERNEL_TEST_DEPS,
4805)
4806
4807xnnpack_unit_test(
4808 name = "f32_vsqrdiffc_test",
4809 srcs = [
4810 "test/f32-vsqrdiffc.cc",
4811 "test/vbinaryc-microkernel-tester.h",
4812 ] + MICROKERNEL_TEST_HDRS,
4813 deps = MICROKERNEL_TEST_DEPS,
4814)
4815
4816xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004817 name = "f32_vsqrt_test",
4818 srcs = [
4819 "test/f32-vsqrt.cc",
4820 "test/vunary-microkernel-tester.h",
4821 ] + MICROKERNEL_TEST_HDRS,
4822 deps = MICROKERNEL_TEST_DEPS,
4823)
4824
4825xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004826 name = "f32_vsub_test",
4827 srcs = [
4828 "test/f32-vsub.cc",
4829 "test/vbinary-microkernel-tester.h",
4830 ] + MICROKERNEL_TEST_HDRS,
4831 deps = MICROKERNEL_TEST_DEPS,
4832)
4833
4834xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004835 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07004836 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004837 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004838 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004839 ] + MICROKERNEL_TEST_HDRS,
4840 deps = MICROKERNEL_TEST_DEPS,
4841)
4842
4843xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004844 name = "f32_vsubc_test",
4845 srcs = [
4846 "test/f32-vsubc.cc",
4847 "test/vbinaryc-microkernel-tester.h",
4848 ] + MICROKERNEL_TEST_HDRS,
4849 deps = MICROKERNEL_TEST_DEPS,
4850)
4851
4852xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004853 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004854 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004855 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004856 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004857 ] + MICROKERNEL_TEST_HDRS,
4858 deps = MICROKERNEL_TEST_DEPS,
4859)
4860
4861xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004862 name = "f32_vrsubc_test",
4863 srcs = [
4864 "test/f32-vrsubc.cc",
4865 "test/vbinaryc-microkernel-tester.h",
4866 ] + MICROKERNEL_TEST_HDRS,
4867 deps = MICROKERNEL_TEST_DEPS,
4868)
4869
4870xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004871 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004872 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004873 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004874 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07004875 ] + MICROKERNEL_TEST_HDRS,
4876 deps = MICROKERNEL_TEST_DEPS,
4877)
4878
4879xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004880 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004881 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004882 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004883 "test/avgpool-microkernel-tester.h",
4884 "src/xnnpack/AlignedAllocator.h",
4885 ] + MICROKERNEL_TEST_HDRS,
4886 deps = MICROKERNEL_TEST_DEPS,
4887)
4888
4889xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004890 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004891 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004892 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004893 "test/gemm-microkernel-tester.h",
4894 "src/xnnpack/AlignedAllocator.h",
4895 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4896 deps = MICROKERNEL_TEST_DEPS,
4897)
4898
4899xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004900 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004901 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004902 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004903 "test/dwconv-microkernel-tester.h",
4904 "src/xnnpack/AlignedAllocator.h",
4905 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4906 deps = MICROKERNEL_TEST_DEPS,
4907)
4908
4909xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004910 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004911 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004912 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004913 "test/gavgpool-microkernel-tester.h",
4914 "src/xnnpack/AlignedAllocator.h",
4915 ] + MICROKERNEL_TEST_HDRS,
4916 deps = MICROKERNEL_TEST_DEPS,
4917)
4918
4919xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004920 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004921 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004922 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004923 "test/gemm-microkernel-tester.h",
4924 "src/xnnpack/AlignedAllocator.h",
4925 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
4926 deps = MICROKERNEL_TEST_DEPS,
4927)
4928
4929xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004930 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004931 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004932 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004933 "test/vadd-microkernel-tester.h",
4934 ] + MICROKERNEL_TEST_HDRS,
4935 deps = MICROKERNEL_TEST_DEPS,
4936)
4937
4938xnnpack_unit_test(
4939 name = "u8_clamp_test",
4940 srcs = [
4941 "test/u8-clamp.cc",
4942 "test/clamp-microkernel-tester.h",
4943 ] + MICROKERNEL_TEST_HDRS,
4944 deps = MICROKERNEL_TEST_DEPS,
4945)
4946
4947xnnpack_unit_test(
4948 name = "u8_lut32norm_test",
4949 srcs = [
4950 "test/u8-lut32norm.cc",
4951 "test/lut-norm-microkernel-tester.h",
4952 ] + MICROKERNEL_TEST_HDRS,
4953 deps = MICROKERNEL_TEST_DEPS,
4954)
4955
4956xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004957 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004958 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004959 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004960 "test/maxpool-microkernel-tester.h",
4961 ] + MICROKERNEL_TEST_HDRS,
4962 deps = MICROKERNEL_TEST_DEPS,
4963)
4964
4965xnnpack_unit_test(
4966 name = "u8_rmax_test",
4967 srcs = [
4968 "test/u8-rmax.cc",
4969 "test/rmax-microkernel-tester.h",
4970 ] + MICROKERNEL_TEST_HDRS,
4971 deps = MICROKERNEL_TEST_DEPS,
4972)
4973
4974xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07004975 name = "x32_fill_test",
4976 srcs = [
4977 "test/x32-fill.cc",
4978 "test/fill-microkernel-tester.h",
4979 ] + MICROKERNEL_TEST_HDRS,
4980 deps = MICROKERNEL_TEST_DEPS,
4981)
4982
4983xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004984 name = "x32_packx_test",
4985 srcs = [
4986 "test/x32-packx.cc",
4987 "test/pack-microkernel-tester.h",
4988 "src/xnnpack/AlignedAllocator.h",
4989 ] + MICROKERNEL_TEST_HDRS,
4990 deps = MICROKERNEL_TEST_DEPS,
4991)
4992
4993xnnpack_unit_test(
4994 name = "x32_pad_test",
4995 srcs = [
4996 "test/x32-pad.cc",
4997 "test/pad-microkernel-tester.h",
4998 ] + MICROKERNEL_TEST_HDRS,
4999 deps = MICROKERNEL_TEST_DEPS,
5000)
5001
5002xnnpack_unit_test(
5003 name = "x32_unpool_test",
5004 srcs = [
5005 "test/x32-unpool.cc",
5006 "test/unpool-microkernel-tester.h",
5007 ] + MICROKERNEL_TEST_HDRS,
5008 deps = MICROKERNEL_TEST_DEPS,
5009)
5010
5011xnnpack_unit_test(
5012 name = "x32_zip_test",
5013 srcs = [
5014 "test/x32-zip.cc",
5015 "test/zip-microkernel-tester.h",
5016 ] + MICROKERNEL_TEST_HDRS,
5017 deps = MICROKERNEL_TEST_DEPS,
5018)
5019
5020xnnpack_unit_test(
5021 name = "x8_lut_test",
5022 srcs = [
5023 "test/x8-lut.cc",
5024 "test/lut-microkernel-tester.h",
5025 ] + MICROKERNEL_TEST_HDRS,
5026 deps = MICROKERNEL_TEST_DEPS,
5027)
5028
5029xnnpack_unit_test(
5030 name = "x8_zip_test",
5031 srcs = [
5032 "test/x8-zip.cc",
5033 "test/zip-microkernel-tester.h",
5034 ] + MICROKERNEL_TEST_HDRS,
5035 deps = MICROKERNEL_TEST_DEPS,
5036)
5037
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005038xnnpack_unit_test(
5039 name = "requantization_test",
5040 srcs = [
5041 "src/xnnpack/requantization-stubs.h",
5042 "test/requantization.cc",
5043 "test/requantization-tester.h",
5044 ] + MICROKERNEL_TEST_HDRS,
5045 deps = MICROKERNEL_TEST_DEPS,
5046)
5047
Marat Dukhan20c3b922020-03-10 03:45:06 -07005048########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005049
5050xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005051 name = "operator_size_test",
5052 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005053 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005054)
5055
Marat Dukhan20c3b922020-03-10 03:45:06 -07005056xnnpack_binary(
5057 name = "subgraph_size_test",
5058 srcs = ["test/subgraph-size.c"],
5059 deps = [":XNNPACK"],
5060)
5061
5062########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005063
5064xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005065 name = "abs_nc_test",
5066 srcs = [
5067 "test/abs-nc.cc",
5068 "test/abs-operator-tester.h",
5069 ],
5070 deps = OPERATOR_TEST_DEPS,
5071)
5072
5073xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005074 name = "add_nd_test",
5075 srcs = [
5076 "test/add-nd.cc",
5077 "test/binary-elementwise-operator-tester.h",
5078 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005079 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005080)
5081
5082xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005083 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005084 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005085 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005086 "test/argmax-pooling-operator-tester.h",
5087 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005088 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005089)
5090
5091xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005092 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005093 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005094 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005095 "test/average-pooling-operator-tester.h",
5096 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005097 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005098)
5099
5100xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005101 name = "bankers_rounding_nc_test",
5102 srcs = [
5103 "test/bankers-rounding-nc.cc",
5104 "test/bankers-rounding-operator-tester.h",
5105 ],
5106 deps = OPERATOR_TEST_DEPS,
5107)
5108
5109xnnpack_unit_test(
5110 name = "ceiling_nc_test",
5111 srcs = [
5112 "test/ceiling-nc.cc",
5113 "test/ceiling-operator-tester.h",
5114 ],
5115 deps = OPERATOR_TEST_DEPS,
5116)
5117
5118xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005119 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005120 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005121 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005122 "test/channel-shuffle-operator-tester.h",
5123 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005124 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005125)
5126
5127xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005128 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005129 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005130 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005131 "test/clamp-operator-tester.h",
5132 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005133 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005134)
5135
5136xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005137 name = "constant_pad_nd_test",
5138 srcs = [
5139 "test/constant-pad-nd.cc",
5140 "test/constant-pad-operator-tester.h",
5141 ],
5142 deps = OPERATOR_TEST_DEPS,
5143)
5144
5145xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005146 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005147 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005148 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005149 "test/convolution-operator-tester.h",
5150 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005151 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005152)
5153
5154xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005155 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005156 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005157 "test/convolution-nchw.cc",
5158 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005159 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005160 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005161)
5162
5163xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005164 name = "copy_nc_test",
5165 srcs = [
5166 "test/copy-nc.cc",
5167 "test/copy-operator-tester.h",
5168 ],
5169 deps = OPERATOR_TEST_DEPS,
5170)
5171
5172xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005173 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005174 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005175 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005176 "test/deconvolution-operator-tester.h",
5177 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005178 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005179)
5180
5181xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005182 name = "divide_nd_test",
5183 srcs = [
5184 "test/binary-elementwise-operator-tester.h",
5185 "test/divide-nd.cc",
5186 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005187 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005188)
5189
5190xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005191 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005192 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005193 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005194 "test/fully-connected-operator-tester.h",
5195 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005196 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005197)
5198
5199xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005200 name = "floor_nc_test",
5201 srcs = [
5202 "test/floor-nc.cc",
5203 "test/floor-operator-tester.h",
5204 ],
5205 deps = OPERATOR_TEST_DEPS,
5206)
5207
5208xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005209 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005210 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005211 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005212 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005213 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005214 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005215)
5216
5217xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005218 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005219 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005220 "test/global-average-pooling-ncw.cc",
5221 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005222 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005223 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005224)
5225
5226xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005227 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005228 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005229 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005230 "test/hardswish-operator-tester.h",
5231 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005232 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005233)
5234
5235xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005236 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005237 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005238 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005239 "test/leaky-relu-operator-tester.h",
5240 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005241 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005242)
5243
5244xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005245 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005246 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005247 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005248 "test/max-pooling-operator-tester.h",
5249 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005250 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005251)
5252
5253xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005254 name = "maximum_nd_test",
5255 srcs = [
5256 "test/binary-elementwise-operator-tester.h",
5257 "test/maximum-nd.cc",
5258 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005259 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005260)
5261
5262xnnpack_unit_test(
5263 name = "minimum_nd_test",
5264 srcs = [
5265 "test/binary-elementwise-operator-tester.h",
5266 "test/minimum-nd.cc",
5267 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005268 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005269)
5270
5271xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005272 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005273 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005274 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005275 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005276 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005277 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005278)
5279
5280xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005281 name = "negate_nc_test",
5282 srcs = [
5283 "test/negate-nc.cc",
5284 "test/negate-operator-tester.h",
5285 ],
5286 deps = OPERATOR_TEST_DEPS,
5287)
5288
5289xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005290 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005291 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005292 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005293 "test/prelu-operator-tester.h",
5294 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005295 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005296)
5297
5298xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005299 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005300 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005301 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005302 "test/resize-bilinear-operator-tester.h",
5303 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005304 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005305)
5306
5307xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005308 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005309 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005310 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005311 "test/sigmoid-operator-tester.h",
5312 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005313 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005314)
5315
5316xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005317 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005318 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005319 "test/softmax-nc.cc",
5320 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005321 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005322 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005323)
5324
5325xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005326 name = "square_nc_test",
5327 srcs = [
5328 "test/square-nc.cc",
5329 "test/square-operator-tester.h",
5330 ],
5331 deps = OPERATOR_TEST_DEPS,
5332)
5333
5334xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005335 name = "square_root_nc_test",
5336 srcs = [
5337 "test/square-root-nc.cc",
5338 "test/square-root-operator-tester.h",
5339 ],
5340 deps = OPERATOR_TEST_DEPS,
5341)
5342
5343xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005344 name = "squared_difference_nd_test",
5345 srcs = [
5346 "test/binary-elementwise-operator-tester.h",
5347 "test/squared-difference-nd.cc",
5348 ],
5349 deps = OPERATOR_TEST_DEPS,
5350)
5351
5352xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005353 name = "subtract_nd_test",
5354 srcs = [
5355 "test/binary-elementwise-operator-tester.h",
5356 "test/subtract-nd.cc",
5357 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005358 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005359)
5360
5361xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005362 name = "truncation_nc_test",
5363 srcs = [
5364 "test/truncation-nc.cc",
5365 "test/truncation-operator-tester.h",
5366 ],
5367 deps = OPERATOR_TEST_DEPS,
5368)
5369
5370xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005371 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005372 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005373 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005374 "test/unpooling-operator-tester.h",
5375 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005376 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005377)
5378
Chao Mei6ddfc602020-05-13 22:29:36 -07005379############################### Misc unit tests ###############################
5380
5381xnnpack_unit_test(
5382 name = "memory_planner_test",
5383 srcs = [
5384 "test/memory-planner-test.cc",
5385 ],
5386 deps = [
5387 ":XNNPACK",
5388 ":memory_planner",
5389 ],
5390)
5391
Marat Dukhan08c4a432019-10-03 09:29:21 -07005392############################# Build configurations #############################
5393
Marat Dukhanb8642352019-10-30 15:43:02 -07005394# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005395config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005396 name = "xnn_enable_assembly_explicit_true",
5397 define_values = {"xnn_enable_assembly": "true"},
5398)
5399
5400# Disables usage of assembly kernels.
5401config_setting(
5402 name = "xnn_enable_assembly_explicit_false",
5403 define_values = {"xnn_enable_assembly": "false"},
5404)
5405
Marat Dukhan9de90e02020-06-18 16:04:12 -07005406# Enables usage of sparse inference.
5407config_setting(
5408 name = "xnn_enable_sparse_explicit_true",
5409 define_values = {"xnn_enable_sparse": "true"},
5410)
5411
5412# Disables usage of sparse inference.
5413config_setting(
5414 name = "xnn_enable_sparse_explicit_false",
5415 define_values = {"xnn_enable_sparse": "false"},
5416)
5417
Marat Dukhan05702cf2020-03-26 15:41:33 -07005418# Disables usage of HMP-aware optimizations.
5419config_setting(
5420 name = "xnn_enable_hmp_explicit_false",
5421 define_values = {"xnn_enable_hmp": "false"},
5422)
5423
Chao Mei6ddfc602020-05-13 22:29:36 -07005424# Enable usage of optimized memory allocation
5425config_setting(
5426 name = "xnn_enable_memopt_explicit_true",
5427 define_values = {"xnnpack_enable_memopt": "true"},
5428)
5429
5430# Disable usage of optimized memory allocation
5431config_setting(
5432 name = "xnn_enable_memopt_explicit_false",
5433 define_values = {"xnnpack_enable_memopt": "false"},
5434)
5435
Marat Dukhanb8642352019-10-30 15:43:02 -07005436# Builds with -c dbg
5437config_setting(
5438 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005439 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005440 "compilation_mode": "dbg",
5441 },
5442)
5443
5444# Builds with -c opt
5445config_setting(
5446 name = "optimized_build",
5447 values = {
5448 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005449 },
5450)
5451
5452config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005453 name = "linux_k8",
5454 values = {"cpu": "k8"},
5455)
5456
5457config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005458 name = "linux_arm",
5459 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005460)
5461
5462config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005463 name = "linux_armeabi",
5464 values = {"cpu": "armeabi"},
5465)
5466
5467config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005468 name = "linux_armhf",
5469 values = {"cpu": "armhf"},
5470)
5471
5472config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005473 name = "linux_armv7a",
5474 values = {"cpu": "armv7a"},
5475)
5476
5477config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005478 name = "linux_aarch64",
5479 values = {"cpu": "aarch64"},
5480)
5481
5482config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005483 name = "android",
5484 values = {"crosstool_top": "//external:android/crosstool"},
5485)
5486
5487config_setting(
5488 name = "android_armv7",
5489 values = {
5490 "crosstool_top": "//external:android/crosstool",
5491 "cpu": "armeabi-v7a",
5492 },
5493)
5494
5495config_setting(
5496 name = "android_arm64",
5497 values = {
5498 "crosstool_top": "//external:android/crosstool",
5499 "cpu": "arm64-v8a",
5500 },
5501)
5502
5503config_setting(
5504 name = "android_x86",
5505 values = {
5506 "crosstool_top": "//external:android/crosstool",
5507 "cpu": "x86",
5508 },
5509)
5510
5511config_setting(
5512 name = "android_x86_64",
5513 values = {
5514 "crosstool_top": "//external:android/crosstool",
5515 "cpu": "x86_64",
5516 },
5517)
5518
5519config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005520 name = "windows_x86_64",
5521 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005522)
5523
5524config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005525 name = "windows_x86_64_clang",
5526 values = {
5527 "compiler": "clang-cl",
5528 "cpu": "x64_windows",
5529 },
5530)
5531
5532config_setting(
5533 name = "windows_x86_64_mingw",
5534 values = {
5535 "compiler": "mingw-gcc",
5536 "cpu": "x64_windows",
5537 },
5538)
5539
5540config_setting(
5541 name = "windows_x86_64_msys",
5542 values = {
5543 "compiler": "msys-gcc",
5544 "cpu": "x64_windows",
5545 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005546)
5547
5548config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005549 name = "macos_x86_64",
5550 values = {
5551 "apple_platform_type": "macos",
5552 "cpu": "darwin",
5553 },
5554)
5555
5556config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005557 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005558 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005559)
5560
5561config_setting(
5562 name = "emscripten_wasm",
5563 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005564 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005565 "cpu": "wasm",
5566 },
5567)
5568
5569config_setting(
5570 name = "emscripten_wasmsimd",
5571 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005572 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005573 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005574 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005575 },
5576)
5577
5578config_setting(
5579 name = "emscripten_asmjs",
5580 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005581 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005582 "cpu": "asmjs",
5583 },
5584)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005585
5586config_setting(
5587 name = "ios_armv7",
5588 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005589 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005590 "cpu": "ios_armv7",
5591 },
5592)
5593
5594config_setting(
5595 name = "ios_arm64",
5596 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005597 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005598 "cpu": "ios_arm64",
5599 },
5600)
5601
5602config_setting(
5603 name = "ios_arm64e",
5604 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005605 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005606 "cpu": "ios_arm64e",
5607 },
5608)
5609
5610config_setting(
5611 name = "ios_x86",
5612 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005613 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005614 "cpu": "ios_i386",
5615 },
5616)
5617
5618config_setting(
5619 name = "ios_x86_64",
5620 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005621 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005622 "cpu": "ios_x86_64",
5623 },
5624)
5625
5626config_setting(
5627 name = "watchos_armv7k",
5628 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005629 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005630 "cpu": "watchos_armv7k",
5631 },
5632)
5633
5634config_setting(
5635 name = "watchos_arm64_32",
5636 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005637 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005638 "cpu": "watchos_arm64_32",
5639 },
5640)
5641
5642config_setting(
5643 name = "watchos_x86",
5644 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005645 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005646 "cpu": "watchos_i386",
5647 },
5648)
5649
5650config_setting(
5651 name = "watchos_x86_64",
5652 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005653 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005654 "cpu": "watchos_x86_64",
5655 },
5656)
5657
5658config_setting(
5659 name = "tvos_arm64",
5660 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005661 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005662 "cpu": "tvos_arm64",
5663 },
5664)
5665
5666config_setting(
5667 name = "tvos_x86_64",
5668 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005669 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005670 "cpu": "tvos_x86_64",
5671 },
5672)