blob: 89df8556d1623f0ed715488f44fffe687c847f76 [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",
Frank Barchard8e229db2020-07-06 23:31:35 -0700232 "src/f32-vbinary/gen/vadd-scalar-x1.c",
233 "src/f32-vbinary/gen/vadd-scalar-x2.c",
234 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700235 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
236 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
237 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700238 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
239 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
240 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700241 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
242 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
243 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700244 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
245 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
246 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700247 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
248 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
249 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700250 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
251 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
252 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800253 "src/f32-vbinary/gen/vmax-scalar-x1.c",
254 "src/f32-vbinary/gen/vmax-scalar-x2.c",
255 "src/f32-vbinary/gen/vmax-scalar-x4.c",
256 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
257 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
258 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
259 "src/f32-vbinary/gen/vmin-scalar-x1.c",
260 "src/f32-vbinary/gen/vmin-scalar-x2.c",
261 "src/f32-vbinary/gen/vmin-scalar-x4.c",
262 "src/f32-vbinary/gen/vminc-scalar-x1.c",
263 "src/f32-vbinary/gen/vminc-scalar-x2.c",
264 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700265 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
266 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
267 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700268 "src/f32-vbinary/gen/vmul-scalar-x1.c",
269 "src/f32-vbinary/gen/vmul-scalar-x2.c",
270 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700271 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
272 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
273 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700274 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
275 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
276 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700277 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
278 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
279 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700280 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
281 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
282 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700283 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
284 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
285 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700286 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
287 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
288 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700289 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
290 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
291 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
292 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
293 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
294 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700295 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
296 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
297 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700298 "src/f32-vbinary/gen/vsub-scalar-x1.c",
299 "src/f32-vbinary/gen/vsub-scalar-x2.c",
300 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700301 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
302 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
303 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700304 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
305 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
306 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700307 "src/f32-vlrelu/gen/scalar-x1.c",
308 "src/f32-vlrelu/gen/scalar-x2.c",
309 "src/f32-vlrelu/gen/scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700310 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
311 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
312 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700313 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
314 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
315 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
316 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
317 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
318 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
319 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
320 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
321 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
322 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
323 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
324 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700325 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
326 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
327 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700328 "src/f32-vunary/gen/vabs-scalar-x1.c",
329 "src/f32-vunary/gen/vabs-scalar-x2.c",
330 "src/f32-vunary/gen/vabs-scalar-x4.c",
331 "src/f32-vunary/gen/vneg-scalar-x1.c",
332 "src/f32-vunary/gen/vneg-scalar-x2.c",
333 "src/f32-vunary/gen/vneg-scalar-x4.c",
334 "src/f32-vunary/gen/vsqr-scalar-x1.c",
335 "src/f32-vunary/gen/vsqr-scalar-x2.c",
336 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800337 "src/math/expminus-scalar-lut2048-p1.c",
338 "src/math/expminus-scalar-lut64-p2.c",
339 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700340 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700341 "src/math/roundne-scalar-nearbyint.c",
342 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700343 "src/math/roundd-scalar-addsub.c",
344 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700345 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700346 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700347 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700348 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700349 "src/math/roundz-scalar-addsub.c",
350 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700351 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800352 "src/math/sigmoid-scalar-lut2048-p1-div.c",
353 "src/math/sigmoid-scalar-lut64-p2-div.c",
354 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700355 "src/q8-avgpool/9p8x-minmax-scalar-c1.c",
356 "src/q8-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700357 "src/q8-dwconv/up1x9-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700358 "src/q8-gavgpool/7p7x-minmax-scalar-c1.c",
359 "src/q8-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhande06f492020-04-09 00:19:31 -0700360 "src/q8-gemm/2x2-minmax-scalar.c",
361 "src/q8-igemm/2x2-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700362 "src/q8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700363 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700364 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700365 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700366 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700367 "src/x32-fill/scalar-float.c",
368 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700369 "src/x32-packx/x2-scalar.c",
370 "src/x32-packx/x3-scalar.c",
371 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700372 "src/x32-pad/scalar-int.c",
373 "src/x32-pad/scalar-float.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700374 "src/x32-unpool/scalar.c",
375 "src/x32-zip/x2-scalar.c",
376 "src/x32-zip/x3-scalar.c",
377 "src/x32-zip/x4-scalar.c",
378 "src/x32-zip/xm-scalar.c",
379 "src/x8-lut/scalar.c",
380 "src/x8-zip/x2-scalar.c",
381 "src/x8-zip/x3-scalar.c",
382 "src/x8-zip/x4-scalar.c",
383 "src/x8-zip/xm-scalar.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700384 "src/requantization/precise-scalar.c",
385 "src/requantization/fp32-scalar.c",
386 "src/requantization/q31-scalar.c",
387 "src/requantization/gemmlowp-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700388]
389
Marat Dukhan436ebe62019-12-04 15:10:12 -0800390WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700391 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
392 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700393 "src/f32-clamp/gen/wasm-x1.c",
394 "src/f32-clamp/gen/wasm-x2.c",
395 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700396 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
397 "src/f32-dwconv/gen/up1x4-wasm.c",
398 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
399 "src/f32-dwconv/gen/up1x9-wasm.c",
400 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
401 "src/f32-dwconv/gen/up1x25-wasm.c",
402 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
403 "src/f32-dwconv/gen/up2x4-wasm.c",
404 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
405 "src/f32-dwconv/gen/up2x9-wasm.c",
406 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
407 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700408 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
409 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
410 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
411 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700412 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
413 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700414 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
415 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
416 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
417 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700418 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
419 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700420 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
421 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700422 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
423 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
424 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700425 "src/f32-gemm/gen/1x4-wasm.c",
426 "src/f32-gemm/gen/2x4-wasm.c",
427 "src/f32-gemm/gen/4x2-wasm.c",
428 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700429 "src/f32-gemm/gen/1x4-relu-wasm.c",
430 "src/f32-gemm/gen/2x4-relu-wasm.c",
431 "src/f32-gemm/gen/4x2-relu-wasm.c",
432 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700433 "src/f32-gemm/gen/1x4-minmax-wasm.c",
434 "src/f32-gemm/gen/2x4-minmax-wasm.c",
435 "src/f32-gemm/gen/4x2-minmax-wasm.c",
436 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800437 "src/f32-hswish/gen/wasm-x1.c",
438 "src/f32-hswish/gen/wasm-x2.c",
439 "src/f32-hswish/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700440 "src/f32-igemm/gen/1x4-wasm.c",
441 "src/f32-igemm/gen/2x4-wasm.c",
442 "src/f32-igemm/gen/4x2-wasm.c",
443 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700444 "src/f32-igemm/gen/1x4-relu-wasm.c",
445 "src/f32-igemm/gen/2x4-relu-wasm.c",
446 "src/f32-igemm/gen/4x2-relu-wasm.c",
447 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700448 "src/f32-igemm/gen/1x4-minmax-wasm.c",
449 "src/f32-igemm/gen/2x4-minmax-wasm.c",
450 "src/f32-igemm/gen/4x2-minmax-wasm.c",
451 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700452 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
453 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
454 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700455 "src/f32-prelu/gen/wasm-2x1.c",
456 "src/f32-prelu/gen/wasm-2x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700457 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
458 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
459 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
460 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
461 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
462 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
463 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
464 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
465 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
466 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
467 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
468 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800469 "src/f32-vbinary/gen/vmax-wasm-x1.c",
470 "src/f32-vbinary/gen/vmax-wasm-x2.c",
471 "src/f32-vbinary/gen/vmax-wasm-x4.c",
472 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
473 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
474 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
475 "src/f32-vbinary/gen/vmin-wasm-x1.c",
476 "src/f32-vbinary/gen/vmin-wasm-x2.c",
477 "src/f32-vbinary/gen/vmin-wasm-x4.c",
478 "src/f32-vbinary/gen/vminc-wasm-x1.c",
479 "src/f32-vbinary/gen/vminc-wasm-x2.c",
480 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700481 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
482 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
483 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
484 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
485 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
486 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
487 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
488 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
489 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
490 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
491 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
492 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
493 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
494 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
495 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
496 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
497 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
498 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700499 "src/f32-vlrelu/gen/wasm-x1.c",
500 "src/f32-vlrelu/gen/wasm-x2.c",
501 "src/f32-vlrelu/gen/wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700502 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
503 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
504 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800505]
506
Marat Dukhan290055c2020-06-09 12:24:29 -0700507WASMSIMD_UKERNELS = [
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700508 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
509 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
510 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
511 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
512 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
513 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700514 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
515 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
516 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
517 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
518 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
519 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
520 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
521 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
522 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
523 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
524 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
525 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
526 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
527 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
528 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
529 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
530 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
531 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
532 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
533 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
534 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
535 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
536 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
537 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700538 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
539 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
540 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
541 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat-arm.c",
542 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat-arm.c",
543 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat-x86.c",
544 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700545 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
546 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
547 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
548 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
549 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
550 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
551 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
552 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
553 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
554 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
555 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
556 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
557 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
558 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
559 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
560 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
561 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
562 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
563 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
564 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
565 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
566 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
567 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
568 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
569 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
570 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
571 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
572 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
573 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
574 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
575 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
576 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
577 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
578 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
579 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
580 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
581 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
582 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
583 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
584 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
585 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
586 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
587 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
588 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
589 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
590 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
591 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
592 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
593 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
594 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
595 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
596 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
597 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
598 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
599 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
600 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
601 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
602 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
603 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
604 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700605 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
606 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
607 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
608 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat-arm.c",
609 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat-arm.c",
610 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat-x86.c",
611 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700612 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
613 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
614 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
615 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
616 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
617 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
618 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
619 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
620 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
621 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
622 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
623 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
624 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
625 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
626 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
627 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
628 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
629 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
630 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
631 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
632 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
633 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
634 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
635 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
636 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
637 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
638 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
639 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
640 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
641 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhan9baec802020-06-25 21:34:35 -0700642 "src/f32-hswish/gen/wasmsimd-arm-x4.c",
643 "src/f32-hswish/gen/wasmsimd-arm-x8.c",
Marat Dukhanc303fe62020-06-26 10:09:25 -0700644 "src/f32-hswish/gen/wasmsimd-arm-x16.c",
Marat Dukhan9baec802020-06-25 21:34:35 -0700645 "src/f32-hswish/gen/wasmsimd-x86-x4.c",
646 "src/f32-hswish/gen/wasmsimd-x86-x8.c",
Marat Dukhanc303fe62020-06-26 10:09:25 -0700647 "src/f32-hswish/gen/wasmsimd-x86-x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700648 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
649 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
650 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
651 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700652 "src/f32-rmax/wasmsimd-arm.c",
653 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700654 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
655 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
656 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
657 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700658 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
659 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700660 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
661 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
662 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
663 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700664 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
665 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700666 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
667 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
668 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
669 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700670 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
671 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700672 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
673 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
674 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
675 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700676 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
677 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700678 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
679 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
680 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
681 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700682 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
683 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
684 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
685 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700686 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
687 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
688 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
689 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700690 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
691 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
692 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
693 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700694 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
695 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
696 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
697 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700698 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
699 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700700 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
701 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
702 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
703 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700704 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
705 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700706 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
707 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
708 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
709 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700710 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
711 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700712 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
713 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
714 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
715 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700716 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
717 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700718 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
719 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700720 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
721 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700722 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
723 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
724 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
725 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700726 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
727 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700728 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
729 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
730 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
731 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700732 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
733 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan72b399a2020-06-28 22:41:38 -0700734 "src/f32-vrnd/gen/vrndne-wasmsimd-x4.c",
735 "src/f32-vrnd/gen/vrndne-wasmsimd-x8.c",
736 "src/f32-vrnd/gen/vrndz-wasmsimd-x4.c",
737 "src/f32-vrnd/gen/vrndz-wasmsimd-x8.c",
738 "src/f32-vrnd/gen/vrndu-wasmsimd-x4.c",
739 "src/f32-vrnd/gen/vrndu-wasmsimd-x8.c",
740 "src/f32-vrnd/gen/vrndd-wasmsimd-x4.c",
741 "src/f32-vrnd/gen/vrndd-wasmsimd-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700742 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
743 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700744 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
745 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
746 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
747 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
748 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
749 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700750 "src/math/roundne-wasmsimd-addsub.c",
751 "src/math/roundd-wasmsimd-addsub.c",
752 "src/math/roundu-wasmsimd-addsub.c",
753 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700754 "src/requantization/fp32-wasmsimd.c",
755]
756
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800757PSIMD_FASTMATH_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800758 "src/f32-argmaxpool/4x-psimd-c4.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800759 "src/f32-argmaxpool/9p8x-psimd-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800760 "src/f32-argmaxpool/9x-psimd-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700761 "src/f32-avgpool/9p8x-minmax-psimd-c4.c",
762 "src/f32-avgpool/9x-minmax-psimd-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700763 "src/f32-clamp/gen/psimd-x4.c",
764 "src/f32-clamp/gen/psimd-x8.c",
Erich Elsen0a1970e2020-06-10 09:24:59 -0700765 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700766 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700767 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700768 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700769 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700770 "src/f32-dwconv/gen/up4x25-minmax-psimd-acc2.c",
771 "src/f32-dwconv/gen/up4x25-minmax-psimd.c",
772 "src/f32-dwconv/gen/up4x4-minmax-psimd-acc2.c",
773 "src/f32-dwconv/gen/up4x4-minmax-psimd.c",
774 "src/f32-dwconv/gen/up4x9-minmax-psimd-acc2.c",
775 "src/f32-dwconv/gen/up4x9-minmax-psimd.c",
776 "src/f32-dwconv/gen/up8x25-minmax-psimd-acc2.c",
777 "src/f32-dwconv/gen/up8x25-minmax-psimd.c",
778 "src/f32-dwconv/gen/up8x4-minmax-psimd-acc2.c",
779 "src/f32-dwconv/gen/up8x4-minmax-psimd.c",
780 "src/f32-dwconv/gen/up8x9-minmax-psimd-acc2.c",
781 "src/f32-dwconv/gen/up8x9-minmax-psimd.c",
Erich Elsen6f278b52020-06-10 16:13:11 -0700782 "src/f32-gavgpool-cw/psimd-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700783 "src/f32-gavgpool/7p7x-minmax-psimd-c4.c",
784 "src/f32-gavgpool/7x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700785 "src/f32-gemm/gen/1x8-minmax-psimd-loadsplat.c",
786 "src/f32-gemm/gen/1x8-minmax-psimd-splat.c",
787 "src/f32-gemm/gen/1x8s4-minmax-psimd.c",
788 "src/f32-gemm/gen/4x2c4-minmax-psimd.c",
789 "src/f32-gemm/gen/4x8-minmax-psimd-loadsplat.c",
790 "src/f32-gemm/gen/4x8-minmax-psimd-splat.c",
791 "src/f32-gemm/gen/4x8s4-minmax-psimd.c",
792 "src/f32-gemm/gen/6x8-minmax-psimd-loadsplat.c",
793 "src/f32-gemm/gen/6x8-minmax-psimd-splat.c",
794 "src/f32-gemm/gen/6x8s4-minmax-psimd.c",
795 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-loadsplat.c",
796 "src/f32-gemm/gen-inc/1x8inc-minmax-psimd-splat.c",
797 "src/f32-gemm/gen-inc/1x8s4inc-minmax-psimd.c",
798 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-loadsplat.c",
799 "src/f32-gemm/gen-inc/4x8inc-minmax-psimd-splat.c",
800 "src/f32-gemm/gen-inc/4x8s4inc-minmax-psimd.c",
801 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-loadsplat.c",
802 "src/f32-gemm/gen-inc/6x8inc-minmax-psimd-splat.c",
803 "src/f32-gemm/gen-inc/6x8s4inc-minmax-psimd.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800804 "src/f32-hswish/gen/psimd-x4.c",
805 "src/f32-hswish/gen/psimd-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700806 "src/f32-ibilinear/gen/psimd-c4.c",
807 "src/f32-ibilinear/gen/psimd-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700808 "src/f32-igemm/gen/1x8-minmax-psimd-loadsplat.c",
809 "src/f32-igemm/gen/1x8-minmax-psimd-splat.c",
810 "src/f32-igemm/gen/1x8s4-minmax-psimd.c",
811 "src/f32-igemm/gen/4x2c4-minmax-psimd.c",
812 "src/f32-igemm/gen/4x8-minmax-psimd-loadsplat.c",
813 "src/f32-igemm/gen/4x8-minmax-psimd-splat.c",
814 "src/f32-igemm/gen/4x8s4-minmax-psimd.c",
815 "src/f32-igemm/gen/6x8-minmax-psimd-loadsplat.c",
816 "src/f32-igemm/gen/6x8-minmax-psimd-splat.c",
817 "src/f32-igemm/gen/6x8s4-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700818 "src/f32-maxpool/9p8x-minmax-psimd-c4.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700819 "src/f32-spmm/gen/4x1-minmax-psimd.c",
820 "src/f32-spmm/gen/8x1-minmax-psimd.c",
821 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700822 "src/f32-pavgpool/9p8x-minmax-psimd-c4.c",
823 "src/f32-pavgpool/9x-minmax-psimd-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700824 "src/f32-ppmm/gen/4x8-minmax-psimd.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800825 "src/f32-prelu/gen/psimd-2x4.c",
826 "src/f32-prelu/gen/psimd-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -0800827 "src/f32-rmax/psimd.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700828 "src/f32-vbinary/gen/vadd-minmax-psimd-x4.c",
829 "src/f32-vbinary/gen/vadd-minmax-psimd-x8.c",
830 "src/f32-vbinary/gen/vaddc-minmax-psimd-x4.c",
831 "src/f32-vbinary/gen/vaddc-minmax-psimd-x8.c",
832 "src/f32-vbinary/gen/vdiv-minmax-psimd-x4.c",
833 "src/f32-vbinary/gen/vdiv-minmax-psimd-x8.c",
834 "src/f32-vbinary/gen/vdivc-minmax-psimd-x4.c",
835 "src/f32-vbinary/gen/vdivc-minmax-psimd-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800836 "src/f32-vbinary/gen/vmax-psimd-x4.c",
837 "src/f32-vbinary/gen/vmax-psimd-x8.c",
838 "src/f32-vbinary/gen/vmaxc-psimd-x4.c",
839 "src/f32-vbinary/gen/vmaxc-psimd-x8.c",
840 "src/f32-vbinary/gen/vmin-psimd-x4.c",
841 "src/f32-vbinary/gen/vmin-psimd-x8.c",
842 "src/f32-vbinary/gen/vminc-psimd-x4.c",
843 "src/f32-vbinary/gen/vminc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700844 "src/f32-vbinary/gen/vmul-minmax-psimd-x4.c",
845 "src/f32-vbinary/gen/vmul-minmax-psimd-x8.c",
846 "src/f32-vbinary/gen/vmulc-minmax-psimd-x4.c",
847 "src/f32-vbinary/gen/vmulc-minmax-psimd-x8.c",
848 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x4.c",
849 "src/f32-vbinary/gen/vrdivc-minmax-psimd-x8.c",
850 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x4.c",
851 "src/f32-vbinary/gen/vrsubc-minmax-psimd-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700852 "src/f32-vbinary/gen/vsqrdiff-psimd-x4.c",
853 "src/f32-vbinary/gen/vsqrdiff-psimd-x8.c",
854 "src/f32-vbinary/gen/vsqrdiffc-psimd-x4.c",
855 "src/f32-vbinary/gen/vsqrdiffc-psimd-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700856 "src/f32-vbinary/gen/vsub-minmax-psimd-x4.c",
857 "src/f32-vbinary/gen/vsub-minmax-psimd-x8.c",
858 "src/f32-vbinary/gen/vsubc-minmax-psimd-x4.c",
859 "src/f32-vbinary/gen/vsubc-minmax-psimd-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -0700860 "src/f32-vlrelu/gen/psimd-x4.c",
861 "src/f32-vlrelu/gen/psimd-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700862 "src/f32-vmulcaddc/gen/c4-minmax-psimd-2x.c",
863 "src/f32-vmulcaddc/gen/c8-minmax-psimd-2x.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700864 "src/f32-vunary/gen/vabs-psimd-x4.c",
865 "src/f32-vunary/gen/vabs-psimd-x8.c",
866 "src/f32-vunary/gen/vneg-psimd-x4.c",
867 "src/f32-vunary/gen/vneg-psimd-x8.c",
868 "src/f32-vunary/gen/vsqr-psimd-x4.c",
869 "src/f32-vunary/gen/vsqr-psimd-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700870 "src/x32-fill/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700871 "src/x32-packx/x4-psimd.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700872 "src/x32-pad/psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700873 "src/x32-unpool/psimd.c",
874 "src/x32-zip/x2-psimd.c",
875 "src/x32-zip/x3-psimd.c",
876 "src/x32-zip/x4-psimd.c",
877 "src/x32-zip/xm-psimd.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -0700878 "src/requantization/precise-psimd.c",
879 "src/requantization/fp32-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700880]
881
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800882PSIMD_ACCMATH_UKERNELS = [
Marat Dukhanb39689d2020-01-24 13:32:20 -0800883 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x4.c",
884 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8.c",
885 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x8-acc2.c",
886 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12.c",
887 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc2.c",
888 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x12-acc3.c",
889 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16.c",
890 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc2.c",
891 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x16-acc4.c",
892 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20.c",
893 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc2.c",
894 "src/f32-raddstoreexpminusmax/gen/psimd-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800895 "src/f32-sigmoid/gen/psimd-p5-div-x4.c",
896 "src/f32-sigmoid/gen/psimd-p5-div-x8.c",
897 "src/f32-sigmoid/gen/psimd-p5-div-x12.c",
898 "src/f32-sigmoid/gen/psimd-p5-div-x16.c",
899 "src/f32-sigmoid/gen/psimd-p5-div-x20.c",
900 "src/f32-sigmoid/gen/psimd-p5-div-x24.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700901 "src/f32-vrnd/gen/vrndne-psimd-x4.c",
902 "src/f32-vrnd/gen/vrndne-psimd-x8.c",
903 "src/f32-vrnd/gen/vrndz-psimd-x4.c",
904 "src/f32-vrnd/gen/vrndz-psimd-x8.c",
905 "src/f32-vrnd/gen/vrndu-psimd-x4.c",
906 "src/f32-vrnd/gen/vrndu-psimd-x8.c",
907 "src/f32-vrnd/gen/vrndd-psimd-x4.c",
908 "src/f32-vrnd/gen/vrndd-psimd-x8.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700909 "src/math/roundne-psimd-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700910 "src/math/roundd-psimd-addsub.c",
911 "src/math/roundu-psimd-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700912 "src/math/roundz-psimd-addsub.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800913 "src/math/sigmoid-psimd-p5-div.c",
914]
915
Marat Dukhan08c4a432019-10-03 09:29:21 -0700916# ISA-specific micro-kernels
917NEON_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700918 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
919 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700920 "src/f32-clamp/gen/neon-x4.c",
921 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700922 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
923 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700924 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
925 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700926 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
927 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700928 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
929 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700930 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
931 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
932 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
933 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700934 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
935 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
936 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
937 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700938 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
939 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
940 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
941 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700942 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700943 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
944 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700945 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
946 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
947 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
948 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
949 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
950 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
951 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
952 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
953 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
954 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
955 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
956 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
957 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
958 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
959 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
960 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
961 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
962 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
963 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
964 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
965 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
966 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
967 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
968 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
969 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
970 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
971 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
972 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
973 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
974 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
975 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhan662faa02019-12-09 22:48:16 -0800976 "src/f32-hswish/gen/neon-x4.c",
977 "src/f32-hswish/gen/neon-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700978 "src/f32-ibilinear/gen/neon-c4.c",
979 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700980 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
981 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
982 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
983 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
984 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
985 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
986 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
987 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
988 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
989 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
990 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
991 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
992 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
993 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
994 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
995 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700996 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
997 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
998 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700999 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1000 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001001 "src/f32-prelu/gen/neon-2x4.c",
1002 "src/f32-prelu/gen/neon-2x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001003 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1004 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1005 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1006 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1007 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1008 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1009 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1010 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1011 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1012 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1013 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1014 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1015 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1016 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1017 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1018 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1019 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1020 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1021 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1022 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1023 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1024 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1025 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1026 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001027 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001028 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001029 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1030 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1031 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1032 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1033 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1034 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1035 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1036 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1037 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1038 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1039 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1040 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1041 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1042 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1043 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1044 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1045 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1046 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001047 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1048 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1049 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1050 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001051 "src/f32-vbinary/gen/vmax-neon-x4.c",
1052 "src/f32-vbinary/gen/vmax-neon-x8.c",
1053 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1054 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1055 "src/f32-vbinary/gen/vmin-neon-x4.c",
1056 "src/f32-vbinary/gen/vmin-neon-x8.c",
1057 "src/f32-vbinary/gen/vminc-neon-x4.c",
1058 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001059 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1060 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1061 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1062 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1063 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1064 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001065 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1066 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1067 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1068 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001069 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1070 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1071 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1072 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001073 "src/f32-vlrelu/gen/neon-x4.c",
1074 "src/f32-vlrelu/gen/neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001075 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1076 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001077 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1078 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1079 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1080 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1081 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1082 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1083 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1084 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001085 "src/f32-vunary/gen/vabs-neon-x4.c",
1086 "src/f32-vunary/gen/vabs-neon-x8.c",
1087 "src/f32-vunary/gen/vneg-neon-x4.c",
1088 "src/f32-vunary/gen/vneg-neon-x8.c",
1089 "src/f32-vunary/gen/vsqr-neon-x4.c",
1090 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001091 "src/q8-avgpool/9p8x-minmax-neon-c8.c",
1092 "src/q8-avgpool/9x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001093 "src/q8-dwconv/up8x9-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001094 "src/q8-gavgpool/7p7x-minmax-neon-c8.c",
1095 "src/q8-gavgpool/7x-minmax-neon-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001096 "src/q8-gemm/4x8-minmax-neon.c",
1097 "src/q8-gemm/8x8-minmax-neon.c",
1098 "src/q8-igemm/4x8-minmax-neon.c",
1099 "src/q8-igemm/8x8-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001100 "src/q8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001101 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001102 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001103 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001104 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001105 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001106 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001107 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001108 "src/x32-zip/x2-neon.c",
1109 "src/x32-zip/x3-neon.c",
1110 "src/x32-zip/x4-neon.c",
1111 "src/x32-zip/xm-neon.c",
1112 "src/x8-zip/x2-neon.c",
1113 "src/x8-zip/x3-neon.c",
1114 "src/x8-zip/x4-neon.c",
1115 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001116 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001117 "src/math/roundd-neon-addsub.c",
1118 "src/math/roundd-neon-cvt.c",
1119 "src/math/roundu-neon-addsub.c",
1120 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001121 "src/math/roundz-neon-addsub.c",
1122 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001123 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001124 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1125 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1126 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1127 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1128 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1129 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001130 "src/math/sqrt-neon-nr1rsqrts.c",
1131 "src/math/sqrt-neon-nr2rsqrts.c",
1132 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001133 "src/requantization/precise-neon.c",
1134 "src/requantization/fp32-neon.c",
1135 "src/requantization/q31-neon.c",
1136 "src/requantization/gemmlowp-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001137]
1138
1139NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001140 "src/f32-ibilinear/gen/neonfma-c4.c",
1141 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001142 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1143 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1144 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1145 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1146 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1147 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1148 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1149 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1150 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001151 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1152 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1153 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1154 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001155 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1156 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1157 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1158 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001159 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1160 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1161 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1162 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001163 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1164 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1165 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1166 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1167 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1168 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1169 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1170 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1171 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1172 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1173 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1174 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1175 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1176 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1177 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1178 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1179 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1180 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001181 "src/f32-hswish/gen/neonfma-x4.c",
1182 "src/f32-hswish/gen/neonfma-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001183 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1184 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001185 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1186 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1187 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1188 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1189 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1190 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1191 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1192 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1193 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1194 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1195 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1196 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1197 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1198 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1199 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1200 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1201 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1202 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1203 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1204 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1205 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1206 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1207 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1208 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001209 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1210 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1211 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1212 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1213 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1214 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1215 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1216 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1217 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1218 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1219 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1220 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1221 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1222 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1223 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1224 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1225 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1226 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1227 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1228 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1229 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1230 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1231 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1232 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1233 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1234 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1235 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1236 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1237 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1238 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1239 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1240 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1241 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1242 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1243 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1244 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1245 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1246 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1247 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1248 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1249 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1250 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1251 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1252 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1253 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1254 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1255 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1256 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1257 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1258 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1259 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1260 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1261 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1262 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001263 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1264 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001265 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1266 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1267 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1268 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1269 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1270 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1271 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1272 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1273 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1274 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1275 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1276 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1277 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1278 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1279 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1280 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1281 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1282 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1283 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1284 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001285 "src/math/exp-neonfma-lut64-p2.c",
1286 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001287 "src/math/expminus-neonfma-lut2048-p1.c",
1288 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001289 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001290 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1291 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1292 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1293 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1294 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1295 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1296 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1297 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1298 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1299 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1300 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1301 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1302 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1303 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1304 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1305 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1306 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1307 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001308 "src/math/sqrt-neonfma-nr1fma.c",
1309 "src/math/sqrt-neonfma-nr2fma.c",
1310 "src/math/sqrt-neonfma-nr3fma.c",
1311 "src/math/sqrt-neonfma-nr2fma1adj.c",
1312 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001313]
1314
1315AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001316 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1317 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1318 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1319 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1320 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1321 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001322 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1323 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1324 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1325 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1326 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1327 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1328 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1329 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1330 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1331 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1332 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1333 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1334 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1335 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1336 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1337 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1338 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1339 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1340 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1341 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001342 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1343 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001344 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1345 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001346 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1347 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001348 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1349 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001350 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1351 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1352 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1353 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1354 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001355 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1356 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1357 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1358 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1359 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1360 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1361 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1362 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1363 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1364 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1365 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1366 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1367 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1368 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1369 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1370 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1371 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1372 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001373 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1374 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1375 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1376 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1377 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1378 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1379 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1380 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1381 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1382 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1383 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1384 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1385 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1386 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1387 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1388 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1389 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1390 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001391 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1392 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001393 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1394 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1395 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1396 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1397 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1398 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001399]
1400
Marat Dukhan8853b822020-05-07 12:19:01 -07001401NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001402 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1403 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1404 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1405 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1406 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1407 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1408 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1409 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001410 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001411 "src/math/roundd-neonv8.c",
1412 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001413 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001414]
1415
Marat Dukhan08c4a432019-10-03 09:29:21 -07001416AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001417 "src/f16-clamp/gen/neonfp16arith-x16.c",
1418 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001419 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1420 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1421 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1422 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1423 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1424 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1425 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1426 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1427 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1428 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1429 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1430 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001431 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1432 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001433 "src/f16-hswish/gen/neonfp16arith-x16.c",
1434 "src/f16-hswish/gen/neonfp16arith-x8.c",
1435 "src/f16-prelu/gen/neonfp16arith-2x16.c",
1436 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001437 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1438 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1439 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1440 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1441 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1442 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1443 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1444 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1445 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1446 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1447 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1448 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1449 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1450 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1451 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1452 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1453 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1454 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1455 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1456 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1457 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1458 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1459 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1460 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1461 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1462 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1463 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1464 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001465 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1466 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001467 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1468 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1469 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1470 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1471 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1472 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1473 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1474 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001475 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1476 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1477 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1478 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001479 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1480 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1481 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1482 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1483 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1484 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1485 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1486 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1487 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1488 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1489 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1490 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001491 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1492 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1493 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1494 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1495 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1496 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1497 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1498 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001499]
1500
1501SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001502 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1503 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001504 "src/f32-clamp/gen/sse-x4.c",
1505 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001506 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1507 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001508 "src/f32-dwconv-chw/3x3p1-sse.c",
1509 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001510 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1511 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1512 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1513 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1514 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1515 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1516 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1517 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1518 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1519 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1520 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1521 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001522 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001523 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1524 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001525 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1526 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1527 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1528 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1529 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1530 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1531 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1532 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1533 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1534 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1535 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1536 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1537 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001538 "src/f32-hswish/gen/sse-x4.c",
1539 "src/f32-hswish/gen/sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001540 "src/f32-ibilinear/gen/sse-c4.c",
1541 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001542 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1543 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1544 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1545 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1546 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1547 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1548 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001549 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1550 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1551 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001552 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001553 "src/f32-prelu/gen/sse-2x4.c",
1554 "src/f32-prelu/gen/sse-2x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001555 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001556 "src/f32-spmm/gen/4x1-minmax-sse.c",
1557 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001558 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001559 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1560 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1561 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1562 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1563 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1564 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1565 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1566 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001567 "src/f32-vbinary/gen/vmax-sse-x4.c",
1568 "src/f32-vbinary/gen/vmax-sse-x8.c",
1569 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1570 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1571 "src/f32-vbinary/gen/vmin-sse-x4.c",
1572 "src/f32-vbinary/gen/vmin-sse-x8.c",
1573 "src/f32-vbinary/gen/vminc-sse-x4.c",
1574 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001575 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1576 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1577 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1578 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1579 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1580 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1581 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1582 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001583 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1584 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1585 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1586 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001587 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1588 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1589 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1590 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001591 "src/f32-vlrelu/gen/sse-x4.c",
1592 "src/f32-vlrelu/gen/sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001593 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1594 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001595 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1596 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001597 "src/f32-vunary/gen/vabs-sse-x4.c",
1598 "src/f32-vunary/gen/vabs-sse-x8.c",
1599 "src/f32-vunary/gen/vneg-sse-x4.c",
1600 "src/f32-vunary/gen/vneg-sse-x8.c",
1601 "src/f32-vunary/gen/vsqr-sse-x4.c",
1602 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001603 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001604 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001605 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001606 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001607 "src/math/roundd-sse-addsub.c",
1608 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001609 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001610 "src/math/sqrt-sse-hh1mac.c",
1611 "src/math/sqrt-sse-nr1mac.c",
1612 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001613]
1614
1615SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001616 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1617 "src/f32-argmaxpool/4x-sse2-c4.c",
1618 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001619 "src/f32-prelu/gen/sse2-2x4.c",
1620 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001621 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1622 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1623 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1624 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1625 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1626 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1627 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1628 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1629 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1630 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1631 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1632 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001633 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001634 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001635 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001636 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001637 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1638 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan0d3f4672020-06-25 16:42:58 -07001639 "src/f32-vlrelu/gen/sse2-x4.c",
1640 "src/f32-vlrelu/gen/sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001641 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1642 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1643 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1644 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1645 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1646 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1647 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1648 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001649 "src/q8-avgpool/9p8x-minmax-sse2-c8.c",
1650 "src/q8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001651 "src/q8-igemm/4x4c2-minmax-sse2.c",
1652 "src/q8-dwconv/up8x9-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001653 "src/q8-gavgpool/7p7x-minmax-sse2-c8.c",
1654 "src/q8-gavgpool/7x-minmax-sse2-c8.c",
Marat Dukhande06f492020-04-09 00:19:31 -07001655 "src/q8-gemm/2x4c8-minmax-sse2.c",
1656 "src/q8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001657 "src/q8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001658 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001659 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001660 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001661 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001662 "src/x32-zip/x2-sse2.c",
1663 "src/x32-zip/x3-sse2.c",
1664 "src/x32-zip/x4-sse2.c",
1665 "src/x32-zip/xm-sse2.c",
1666 "src/x8-zip/x2-sse2.c",
1667 "src/x8-zip/x3-sse2.c",
1668 "src/x8-zip/x4-sse2.c",
1669 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001670 "src/math/exp-sse2-p5.c",
1671 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001672 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001673 "src/math/roundd-sse2-cvt.c",
1674 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001675 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001676 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001677 "src/requantization/precise-sse2.c",
1678 "src/requantization/fp32-sse2.c",
1679 "src/requantization/q31-sse2.c",
1680 "src/requantization/gemmlowp-sse2.c",
1681]
1682
1683SSSE3_UKERNELS = [
1684 "src/requantization/precise-ssse3.c",
1685 "src/requantization/q31-ssse3.c",
1686 "src/requantization/gemmlowp-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001687]
1688
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001689SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001690 "src/f32-prelu/gen/sse41-2x4.c",
1691 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001692 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001693 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001694 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001695 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001696 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1697 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan0d3f4672020-06-25 16:42:58 -07001698 "src/f32-vlrelu/gen/sse41-x4.c",
1699 "src/f32-vlrelu/gen/sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001700 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1701 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1702 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1703 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1704 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1705 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1706 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1707 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001708 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001709 "src/math/roundd-sse41.c",
1710 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001711 "src/math/roundz-sse41.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001712 "src/requantization/precise-sse4.c",
1713 "src/requantization/q31-sse4.c",
1714 "src/requantization/gemmlowp-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001715]
1716
Marat Dukhan08c4a432019-10-03 09:29:21 -07001717AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001718 "src/f32-clamp/gen/avx-x8.c",
1719 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001720 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1721 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1722 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1723 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1724 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1725 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1726 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1727 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1728 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1729 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1730 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1731 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1732 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1733 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1734 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1735 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1736 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1737 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1738 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1739 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1740 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1741 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1742 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1743 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1744 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1745 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1746 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1747 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1748 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1749 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001750 "src/f32-hswish/gen/avx-x8.c",
1751 "src/f32-hswish/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001752 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1753 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1754 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1755 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1756 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1757 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1758 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1759 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1760 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001761 "src/f32-prelu/gen/avx-2x8.c",
1762 "src/f32-prelu/gen/avx-2x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001763 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001764 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1765 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1766 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1767 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1768 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1769 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1770 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1771 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001772 "src/f32-vbinary/gen/vmax-avx-x8.c",
1773 "src/f32-vbinary/gen/vmax-avx-x16.c",
1774 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1775 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1776 "src/f32-vbinary/gen/vmin-avx-x8.c",
1777 "src/f32-vbinary/gen/vmin-avx-x16.c",
1778 "src/f32-vbinary/gen/vminc-avx-x8.c",
1779 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001780 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1781 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1782 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1783 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1784 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1785 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1786 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1787 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001788 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1789 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1790 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1791 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001792 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1793 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1794 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1795 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07001796 "src/f32-vlrelu/gen/avx-x8.c",
1797 "src/f32-vlrelu/gen/avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001798 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001799 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1800 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1801 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1802 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1803 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1804 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1805 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1806 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001807 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1808 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001809 "src/f32-vunary/gen/vabs-avx-x8.c",
1810 "src/f32-vunary/gen/vabs-avx-x16.c",
1811 "src/f32-vunary/gen/vneg-avx-x8.c",
1812 "src/f32-vunary/gen/vneg-avx-x16.c",
1813 "src/f32-vunary/gen/vsqr-avx-x8.c",
1814 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001815]
1816
Marat Dukhanfda12b82019-11-21 12:27:59 -08001817FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001818 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1819 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1820 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1821 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1822 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1823 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1824 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1825 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1826 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1827 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1828 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1829 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1830 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1831 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1832 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1833 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1834 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1835 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1836 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1837 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1838 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1839 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1840 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1841 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1842 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1843 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1844 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1845 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1846 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1847 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1848 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1849 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1850 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1851 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1852 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1853 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1854 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1855 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1856 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1857 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08001858 "src/f32-hswish/gen/fma3-x8.c",
1859 "src/f32-hswish/gen/fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001860 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1861 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1862 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1863 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1864 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1865 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1866 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1867 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1868 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1869 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1870 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1871 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1872 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1873 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001874 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1875 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1876 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1877 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1878 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1879 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1880 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1881 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001882 "src/math/sqrt-fma3-nr1fma.c",
1883 "src/math/sqrt-fma3-nr2fma.c",
1884 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001885]
1886
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001887AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001888 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1889 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1890 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1891 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1892 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1893 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1894 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1895 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1896 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1897 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1898 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1899 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1900 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1901 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1902 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1903 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1904 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1905 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1906 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1907 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1908 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1909 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1910 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1911 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1912 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1913 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1914 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1915 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1916 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1917 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1918 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1919 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1920 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1921 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1922 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1923 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001924 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1925 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1926 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1927 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1928 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1929 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1930 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1931 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1932 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1933 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1934 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1935 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1936 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1937 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1938 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1939 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1940 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1941 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1942 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1943 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1944 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1945 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1946 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1947 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1948 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1949 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1950 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1951 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1952 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1953 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001954 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1955 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1956 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1957 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1958 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1959 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1960 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1961 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1962 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1963 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1964 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1965 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1966 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1967 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1968 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1969 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1970 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1971 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1972 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1973 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1974 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1975 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1976 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1977 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001978 "src/math/exp-avx2-p5.c",
1979 "src/math/exp-avx2-perm-p3.c",
1980 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001981 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001982 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001983 "src/math/sigmoid-avx2-rr2-p5-div.c",
1984 "src/math/sigmoid-avx2-rr1-p5-div.c",
1985 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1986 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1987 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1988 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001989]
1990
Marat Dukhan08c4a432019-10-03 09:29:21 -07001991AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001992 "src/f32-clamp/gen/avx512f-x16.c",
1993 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001994 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1995 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1996 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1997 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1998 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1999 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2000 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2001 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2002 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2003 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2004 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2005 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2006 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2007 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2008 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2009 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2010 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2011 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2012 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2013 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2014 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2015 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2016 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2017 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhan662faa02019-12-09 22:48:16 -08002018 "src/f32-hswish/gen/avx512f-x16.c",
2019 "src/f32-hswish/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002020 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2021 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2022 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2023 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2024 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2025 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002026 "src/f32-prelu/gen/avx512f-2x16.c",
2027 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002028 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2029 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2030 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2031 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2032 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2033 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2034 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2035 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2036 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2037 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2038 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2039 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2040 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2041 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2042 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2043 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2044 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2045 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2046 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2047 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2048 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2049 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2050 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2051 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2052 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2053 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2054 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2055 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2056 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2057 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2058 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2059 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2060 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2061 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2062 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2063 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002064 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002065 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2066 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2067 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2068 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2069 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2070 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2071 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2072 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002073 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2074 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2075 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2076 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2077 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2078 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2079 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2080 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002081 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2082 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2083 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2084 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2085 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2086 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2087 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2088 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002089 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2090 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2091 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2092 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002093 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2094 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2095 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2096 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07002097 "src/f32-vlrelu/gen/avx512f-x16.c",
2098 "src/f32-vlrelu/gen/avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002099 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002100 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2101 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2102 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2103 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2104 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2105 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2106 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2107 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2108 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2109 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2110 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2111 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2112 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2113 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2114 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2115 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2116 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2117 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2118 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2119 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2120 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2121 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2122 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2123 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002124 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2125 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2126 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2127 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2128 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2129 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2130 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2131 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002132 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2133 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2134 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2135 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2136 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2137 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2138 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2139 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002140 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2141 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2142 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2143 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2144 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2145 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002146 "src/math/exp-avx512f-p5-scalef.c",
2147 "src/math/exp-avx512f-p5.c",
2148 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002149 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002150 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002151 "src/math/sqrt-avx512f-nr1fma.c",
2152 "src/math/sqrt-avx512f-nr2fma.c",
2153 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002154]
2155
2156AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002157 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002158 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002159 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2160 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
2161 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2162 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002163 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2164 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S",
2165 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2166 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002167 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2168 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
2169 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2170 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002171]
2172
2173AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002174 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002175 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2176 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002177 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002178 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2179 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002180 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2181 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2182 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002183 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002184 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2185 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2186 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002187 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002188 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2189 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002190 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002191 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2192 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2193 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2194 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2195 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2196 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2197 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2198 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2199 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2200 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2201 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2202 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2203 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2204 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2205 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2206 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2207 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2208 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002209 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2210 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002211 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002212 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2213 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2214 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2215 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2216 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2217 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2218 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2219 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2220 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2221 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2222 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2223 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2224 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2225 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2226 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2227 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2228 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2229 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002230 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2231 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2232 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2233 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2234 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2235 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2236 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2237 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2238 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2239 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2240 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2241 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2242 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2243 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2244 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2245 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2246 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2247 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002248]
2249
Marat Dukhan1b354632020-03-23 12:50:22 -07002250INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002251 "src/requantization/gemmlowp-requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002252 "src/xnnpack/argmaxpool.h",
2253 "src/xnnpack/avgpool.h",
2254 "src/xnnpack/clamp.h",
2255 "src/xnnpack/common.h",
2256 "src/xnnpack/conv.h",
2257 "src/xnnpack/dwconv.h",
2258 "src/xnnpack/gavgpool.h",
2259 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002260 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002261 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002262 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002263 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002264 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002265 "src/xnnpack/lut.h",
2266 "src/xnnpack/math.h",
2267 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002268 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002269 "src/xnnpack/packx.h",
2270 "src/xnnpack/pad.h",
2271 "src/xnnpack/params.h",
2272 "src/xnnpack/pavgpool.h",
2273 "src/xnnpack/ppmm.h",
2274 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002275 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002276 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002277 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002278 "src/xnnpack/rmax.h",
2279 "src/xnnpack/scalar-utils.h",
2280 "src/xnnpack/spmm.h",
2281 "src/xnnpack/unpool.h",
2282 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002283 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002284 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002285 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002286 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002287 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002288 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002289 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002290]
2291
2292INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002293 "include/xnnpack.h",
2294 "src/xnnpack/allocator.h",
2295 "src/xnnpack/compute.h",
2296 "src/xnnpack/im2col.h",
2297 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002298 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002299 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002300 "src/xnnpack/operator.h",
2301 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002302 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002303 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002304 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002305 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002306]
2307
Marat Dukhan1b354632020-03-23 12:50:22 -07002308ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002309 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002310]
2311
Marat Dukhan1b354632020-03-23 12:50:22 -07002312MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002313 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002314 "include/xnnpack.h",
2315]
2316
Marat Dukhan1b354632020-03-23 12:50:22 -07002317MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002318 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002319 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002320 "src/xnnpack/requantization.h",
2321 "include/xnnpack.h",
2322]
2323
2324OPERATOR_TEST_PARAMS_HDRS = [
2325 "src/xnnpack/params.h",
2326 "src/xnnpack/common.h",
2327]
2328
2329WEIGHTS_PACK_HDRS = [
2330 "src/xnnpack/pack.h",
2331 "src/xnnpack/operator.h",
2332 "src/xnnpack/compute.h",
2333]
2334
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002335LOGGING_COPTS = select({
2336 # No logging in optimized mode
2337 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2338 # Full logging in debug mode
2339 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2340 # Error-only logging in default (fastbuild) mode
2341 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2342})
2343
Marat Dukhan3b59de22020-06-03 20:15:19 -07002344LOGGING_SRCS = select({
2345 # No logging in optimized mode
2346 ":optimized_build": [],
2347 "//conditions:default": [
2348 "src/operator-strings.c",
2349 "src/subgraph-strings.c",
2350 ],
2351})
2352
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002353LOGGING_HDRS = [
2354 "src/xnnpack/log.h",
2355]
2356
Marat Dukhan08c4a432019-10-03 09:29:21 -07002357xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002358 name = "tables",
2359 srcs = TABLE_SRCS,
2360 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002361 gcc_copts = xnnpack_gcc_std_copts(),
2362 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002363)
2364
2365xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002366 name = "scalar_ukernels",
2367 srcs = SCALAR_UKERNELS,
2368 hdrs = INTERNAL_HDRS,
2369 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002370 gcc_copts = xnnpack_gcc_std_copts(),
2371 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002372 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002373 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002374 "@FP16",
2375 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002376 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002377 ],
2378)
2379
2380xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002381 name = "scalar_ukernels_test_mode",
2382 srcs = SCALAR_UKERNELS,
2383 hdrs = INTERNAL_HDRS,
2384 aarch32_copts = ["-marm"],
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 deps = [
2392 ":tables",
2393 "@FP16",
2394 "@FXdiv",
2395 "@pthreadpool",
2396 ],
2397)
2398
2399xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002400 name = "wasm_ukernels",
2401 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002402 gcc_copts = xnnpack_gcc_std_copts(),
2403 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002404 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002405 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002406 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002407 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002408 "@FP16",
2409 "@FXdiv",
2410 "@pthreadpool",
2411 ],
2412)
2413
2414xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002415 name = "wasm_ukernels_test_mode",
2416 hdrs = INTERNAL_HDRS,
2417 copts = [
2418 "-UNDEBUG",
2419 "-DXNN_TEST_MODE=1",
2420 ],
2421 gcc_copts = xnnpack_gcc_std_copts(),
2422 msvc_copts = xnnpack_msvc_std_copts(),
2423 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002424 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002425 deps = [
2426 ":tables",
2427 "@FP16",
2428 "@FXdiv",
2429 "@pthreadpool",
2430 ],
2431)
2432
2433xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002434 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002435 hdrs = INTERNAL_HDRS,
2436 aarch32_copts = [
2437 "-marm",
2438 "-mfpu=neon",
2439 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002440 gcc_copts = xnnpack_gcc_std_copts(),
2441 gcc_x86_copts = ["-msse2"],
2442 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002443 optimized_copts = [
2444 "-O3",
2445 "-ffast-math",
2446 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002447 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002448 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002449 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002450 "@FP16",
2451 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002452 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002453 ],
2454)
2455
2456xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002457 name = "psimd_fastmath_ukernels_test_mode",
2458 hdrs = INTERNAL_HDRS,
2459 aarch32_copts = [
2460 "-marm",
2461 "-mfpu=neon",
2462 ],
2463 copts = [
2464 "-UNDEBUG",
2465 "-DXNN_TEST_MODE=1",
2466 ],
2467 gcc_copts = xnnpack_gcc_std_copts(),
2468 gcc_x86_copts = ["-msse2"],
2469 msvc_copts = xnnpack_msvc_std_copts(),
2470 optimized_copts = [
2471 "-O3",
2472 "-ffast-math",
2473 ],
2474 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2475 deps = [
2476 ":tables",
2477 "@FP16",
2478 "@psimd",
2479 "@pthreadpool",
2480 ],
2481)
2482
2483xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002484 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002485 hdrs = INTERNAL_HDRS,
2486 aarch32_copts = [
2487 "-marm",
2488 "-mfpu=neon",
2489 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002490 gcc_copts = xnnpack_gcc_std_copts(),
2491 gcc_x86_copts = ["-msse2"],
2492 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002493 optimized_copts = [
2494 "-O3",
2495 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002496 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002497 deps = [
2498 ":tables",
2499 "@FP16",
2500 "@psimd",
2501 "@pthreadpool",
2502 ],
2503)
2504
2505xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002506 name = "psimd_accmath_ukernels_test_mode",
2507 hdrs = INTERNAL_HDRS,
2508 aarch32_copts = [
2509 "-marm",
2510 "-mfpu=neon",
2511 ],
2512 copts = [
2513 "-UNDEBUG",
2514 "-DXNN_TEST_MODE=1",
2515 ],
2516 gcc_copts = xnnpack_gcc_std_copts(),
2517 gcc_x86_copts = ["-msse2"],
2518 msvc_copts = xnnpack_msvc_std_copts(),
2519 optimized_copts = [
2520 "-O3",
2521 ],
2522 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2523 deps = [
2524 ":tables",
2525 "@FP16",
2526 "@psimd",
2527 "@pthreadpool",
2528 ],
2529)
2530
2531xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002532 name = "neon_ukernels",
2533 hdrs = INTERNAL_HDRS,
2534 aarch32_copts = [
2535 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002536 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002537 "-mfpu=neon",
2538 ],
2539 aarch32_srcs = NEON_UKERNELS,
2540 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002541 gcc_copts = xnnpack_gcc_std_copts(),
2542 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002543 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002544 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002545 "@FP16",
2546 "@pthreadpool",
2547 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002548)
2549
2550xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002551 name = "neon_ukernels_test_mode",
2552 hdrs = INTERNAL_HDRS,
2553 aarch32_copts = [
2554 "-marm",
2555 "-march=armv7-a",
2556 "-mfpu=neon",
2557 ],
2558 aarch32_srcs = NEON_UKERNELS,
2559 aarch64_srcs = NEON_UKERNELS,
2560 copts = [
2561 "-UNDEBUG",
2562 "-DXNN_TEST_MODE=1",
2563 ],
2564 gcc_copts = xnnpack_gcc_std_copts(),
2565 msvc_copts = xnnpack_msvc_std_copts(),
2566 deps = [
2567 ":tables",
2568 "@FP16",
2569 "@pthreadpool",
2570 ],
2571)
2572
2573xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002574 name = "neonfma_ukernels",
2575 hdrs = INTERNAL_HDRS,
2576 aarch32_copts = [
2577 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002578 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002579 "-mfpu=neon-vfpv4",
2580 ],
2581 aarch32_srcs = NEONFMA_UKERNELS,
2582 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002583 apple_aarch32_copts = [
2584 "-mcpu=swift",
2585 "-mtune=generic",
2586 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002587 gcc_copts = xnnpack_gcc_std_copts(),
2588 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002589 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002590 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002591 "@FP16",
2592 "@pthreadpool",
2593 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002594)
2595
2596xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002597 name = "neonfma_ukernels_test_mode",
2598 hdrs = INTERNAL_HDRS,
2599 aarch32_copts = [
2600 "-marm",
2601 "-march=armv7-a",
2602 "-mfpu=neon-vfpv4",
2603 ],
2604 aarch32_srcs = NEONFMA_UKERNELS,
2605 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002606 apple_aarch32_copts = [
2607 "-mcpu=swift",
2608 "-mtune=generic",
2609 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002610 copts = [
2611 "-UNDEBUG",
2612 "-DXNN_TEST_MODE=1",
2613 ],
2614 gcc_copts = xnnpack_gcc_std_copts(),
2615 msvc_copts = xnnpack_msvc_std_copts(),
2616 deps = [
2617 ":tables",
2618 "@FP16",
2619 "@pthreadpool",
2620 ],
2621)
2622
2623xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002624 name = "neonv8_ukernels",
2625 hdrs = INTERNAL_HDRS,
2626 aarch32_copts = [
2627 "-marm",
2628 "-march=armv8-a",
2629 "-mfpu=neon-fp-armv8",
2630 ],
2631 aarch32_srcs = NEONV8_UKERNELS,
2632 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002633 apple_aarch32_copts = [
2634 "-mcpu=cyclone",
2635 "-mtune=generic",
2636 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002637 gcc_copts = xnnpack_gcc_std_copts(),
2638 msvc_copts = xnnpack_msvc_std_copts(),
2639 deps = [
2640 ":tables",
2641 "@FP16",
2642 "@pthreadpool",
2643 ],
2644)
2645
2646xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002647 name = "neonv8_ukernels_test_mode",
2648 hdrs = INTERNAL_HDRS,
2649 aarch32_copts = [
2650 "-marm",
2651 "-march=armv8-a",
2652 "-mfpu=neon-fp-armv8",
2653 ],
2654 aarch32_srcs = NEONV8_UKERNELS,
2655 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002656 apple_aarch32_copts = [
2657 "-mcpu=cyclone",
2658 "-mtune=generic",
2659 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002660 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 = "neonfp16arith_ukernels",
2675 hdrs = INTERNAL_HDRS,
2676 aarch64_copts = ["-march=armv8.2-a+fp16"],
2677 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002678 gcc_copts = xnnpack_gcc_std_copts(),
2679 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002680 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002681 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002682 "@FP16",
2683 "@pthreadpool",
2684 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002685)
2686
2687xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002688 name = "neonfp16arith_ukernels_test_mode",
2689 hdrs = INTERNAL_HDRS,
2690 aarch64_copts = ["-march=armv8.2-a+fp16"],
2691 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2692 copts = [
2693 "-UNDEBUG",
2694 "-DXNN_TEST_MODE=1",
2695 ],
2696 gcc_copts = xnnpack_gcc_std_copts(),
2697 msvc_copts = xnnpack_msvc_std_copts(),
2698 deps = [
2699 ":tables",
2700 "@FP16",
2701 "@pthreadpool",
2702 ],
2703)
2704
2705xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002706 name = "sse2_ukernels",
2707 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002708 gcc_copts = xnnpack_gcc_std_copts(),
2709 gcc_x86_copts = ["-msse2"],
2710 msvc_copts = xnnpack_msvc_std_copts(),
2711 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002712 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002713 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002714 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002715 "@FP16",
2716 "@pthreadpool",
2717 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002718)
2719
2720xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002721 name = "sse2_ukernels_test_mode",
2722 hdrs = INTERNAL_HDRS,
2723 copts = [
2724 "-UNDEBUG",
2725 "-DXNN_TEST_MODE=1",
2726 ],
2727 gcc_copts = xnnpack_gcc_std_copts(),
2728 gcc_x86_copts = ["-msse2"],
2729 msvc_copts = xnnpack_msvc_std_copts(),
2730 msvc_x86_32_copts = ["/arch:SSE2"],
2731 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2732 deps = [
2733 ":tables",
2734 "@FP16",
2735 "@pthreadpool",
2736 ],
2737)
2738
2739xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002740 name = "ssse3_ukernels",
2741 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002742 gcc_copts = xnnpack_gcc_std_copts(),
2743 gcc_x86_copts = ["-mssse3"],
2744 msvc_copts = xnnpack_msvc_std_copts(),
2745 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002746 x86_srcs = SSSE3_UKERNELS,
2747 deps = [
2748 ":tables",
2749 "@FP16",
2750 "@pthreadpool",
2751 ],
2752)
2753
2754xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002755 name = "ssse3_ukernels_test_mode",
2756 hdrs = INTERNAL_HDRS,
2757 copts = [
2758 "-UNDEBUG",
2759 "-DXNN_TEST_MODE=1",
2760 ],
2761 gcc_copts = xnnpack_gcc_std_copts(),
2762 gcc_x86_copts = ["-mssse3"],
2763 msvc_copts = xnnpack_msvc_std_copts(),
2764 msvc_x86_32_copts = ["/arch:SSE2"],
2765 x86_srcs = SSSE3_UKERNELS,
2766 deps = [
2767 ":tables",
2768 "@FP16",
2769 "@pthreadpool",
2770 ],
2771)
2772
2773xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002774 name = "sse41_ukernels",
2775 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002776 gcc_copts = xnnpack_gcc_std_copts(),
2777 gcc_x86_copts = ["-msse4.1"],
2778 msvc_copts = xnnpack_msvc_std_copts(),
2779 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002780 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002781 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002782 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002783 "@FP16",
2784 "@pthreadpool",
2785 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002786)
2787
2788xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002789 name = "sse41_ukernels_test_mode",
2790 hdrs = INTERNAL_HDRS,
2791 copts = [
2792 "-UNDEBUG",
2793 "-DXNN_TEST_MODE=1",
2794 ],
2795 gcc_copts = xnnpack_gcc_std_copts(),
2796 gcc_x86_copts = ["-msse4.1"],
2797 msvc_copts = xnnpack_msvc_std_copts(),
2798 msvc_x86_32_copts = ["/arch:SSE2"],
2799 x86_srcs = SSE41_UKERNELS,
2800 deps = [
2801 ":tables",
2802 "@FP16",
2803 "@pthreadpool",
2804 ],
2805)
2806
2807xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002808 name = "avx_ukernels",
2809 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002810 gcc_copts = xnnpack_gcc_std_copts(),
2811 gcc_x86_copts = ["-mavx"],
2812 msvc_copts = xnnpack_msvc_std_copts(),
2813 msvc_x86_32_copts = ["/arch:AVX"],
2814 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002815 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002816 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002817 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002818 "@FP16",
2819 "@pthreadpool",
2820 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002821)
2822
2823xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002824 name = "avx_ukernels_test_mode",
2825 hdrs = INTERNAL_HDRS,
2826 copts = [
2827 "-UNDEBUG",
2828 "-DXNN_TEST_MODE=1",
2829 ],
2830 gcc_copts = xnnpack_gcc_std_copts(),
2831 gcc_x86_copts = ["-mavx"],
2832 msvc_copts = xnnpack_msvc_std_copts(),
2833 msvc_x86_32_copts = ["/arch:AVX"],
2834 msvc_x86_64_copts = ["/arch:AVX"],
2835 x86_srcs = AVX_UKERNELS,
2836 deps = [
2837 ":tables",
2838 "@FP16",
2839 "@pthreadpool",
2840 ],
2841)
2842
2843xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002844 name = "fma3_ukernels",
2845 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002846 gcc_copts = xnnpack_gcc_std_copts(),
2847 gcc_x86_copts = ["-mfma"],
2848 msvc_copts = xnnpack_msvc_std_copts(),
2849 msvc_x86_32_copts = ["/arch:AVX"],
2850 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002851 x86_srcs = FMA3_UKERNELS,
2852 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002853 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002854 "@FP16",
2855 "@pthreadpool",
2856 ],
2857)
2858
2859xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002860 name = "fma3_ukernels_test_mode",
2861 hdrs = INTERNAL_HDRS,
2862 copts = [
2863 "-UNDEBUG",
2864 "-DXNN_TEST_MODE=1",
2865 ],
2866 gcc_copts = xnnpack_gcc_std_copts(),
2867 gcc_x86_copts = ["-mfma"],
2868 msvc_copts = xnnpack_msvc_std_copts(),
2869 msvc_x86_32_copts = ["/arch:AVX"],
2870 msvc_x86_64_copts = ["/arch:AVX"],
2871 x86_srcs = FMA3_UKERNELS,
2872 deps = [
2873 ":tables",
2874 "@FP16",
2875 "@pthreadpool",
2876 ],
2877)
2878
2879xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002880 name = "avx2_ukernels",
2881 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002882 gcc_copts = xnnpack_gcc_std_copts(),
2883 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002884 "-mfma",
2885 "-mavx2",
2886 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002887 msvc_copts = xnnpack_msvc_std_copts(),
2888 msvc_x86_32_copts = ["/arch:AVX2"],
2889 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002890 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002891 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002892 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002893 "@FP16",
2894 "@pthreadpool",
2895 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002896)
2897
2898xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002899 name = "avx2_ukernels_test_mode",
2900 hdrs = INTERNAL_HDRS,
2901 copts = [
2902 "-UNDEBUG",
2903 "-DXNN_TEST_MODE=1",
2904 ],
2905 gcc_copts = xnnpack_gcc_std_copts(),
2906 gcc_x86_copts = [
2907 "-mfma",
2908 "-mavx2",
2909 ],
2910 msvc_copts = xnnpack_msvc_std_copts(),
2911 msvc_x86_32_copts = ["/arch:AVX2"],
2912 msvc_x86_64_copts = ["/arch:AVX2"],
2913 x86_srcs = AVX2_UKERNELS,
2914 deps = [
2915 ":tables",
2916 "@FP16",
2917 "@pthreadpool",
2918 ],
2919)
2920
2921xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002922 name = "avx512f_ukernels",
2923 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002924 gcc_copts = xnnpack_gcc_std_copts(),
2925 gcc_x86_copts = ["-mavx512f"],
2926 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2927 msvc_copts = xnnpack_msvc_std_copts(),
2928 msvc_x86_32_copts = ["/arch:AVX512"],
2929 msvc_x86_64_copts = ["/arch:AVX512"],
2930 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002931 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002932 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002933 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002934 "@FP16",
2935 "@pthreadpool",
2936 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002937)
2938
2939xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002940 name = "avx512f_ukernels_test_mode",
2941 hdrs = INTERNAL_HDRS,
2942 copts = [
2943 "-UNDEBUG",
2944 "-DXNN_TEST_MODE=1",
2945 ],
2946 gcc_copts = xnnpack_gcc_std_copts(),
2947 gcc_x86_copts = ["-mavx512f"],
2948 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2949 msvc_copts = xnnpack_msvc_std_copts(),
2950 msvc_x86_32_copts = ["/arch:AVX512"],
2951 msvc_x86_64_copts = ["/arch:AVX512"],
2952 msys_copts = ["-fno-asynchronous-unwind-tables"],
2953 x86_srcs = AVX512F_UKERNELS,
2954 deps = [
2955 ":tables",
2956 "@FP16",
2957 "@pthreadpool",
2958 ],
2959)
2960
2961xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002962 name = "asm_ukernels",
2963 hdrs = ["src/xnnpack/assembly.h"],
2964 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002965 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002966 aarch64_srcs = AARCH64_ASM_UKERNELS,
2967)
2968
Marat Dukhan3b59de22020-06-03 20:15:19 -07002969xnnpack_cc_library(
2970 name = "logging_utils",
2971 srcs = LOGGING_SRCS,
2972 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2973 copts = LOGGING_COPTS + [
2974 "-Isrc",
2975 "-Iinclude",
2976 ] + select({
2977 ":debug_build": [],
2978 "//conditions:default": xnnpack_min_size_copts(),
2979 }),
2980 gcc_copts = xnnpack_gcc_std_copts(),
2981 msvc_copts = xnnpack_msvc_std_copts(),
2982 visibility = xnnpack_visibility(),
2983 deps = [
2984 "@FP16",
2985 "@clog",
2986 "@pthreadpool",
2987 ],
2988)
2989
Marat Dukhan08c4a432019-10-03 09:29:21 -07002990xnnpack_aggregate_library(
2991 name = "ukernels",
2992 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002993 ":neon_ukernels",
2994 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002995 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002996 ":asm_ukernels",
2997 ],
2998 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002999 ":neon_ukernels",
3000 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003001 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003002 ":neonfp16arith_ukernels",
3003 ":asm_ukernels",
3004 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003005 generic_deps = [
3006 ":scalar_ukernels",
3007 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003008 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003009 ":psimd_fastmath_ukernels",
3010 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003011 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003012 wasm_deps = [
3013 ":wasm_ukernels",
3014 ],
3015 wasmsimd_deps = [
3016 ":wasm_ukernels",
3017 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003018 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003019 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003020 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003021 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003022 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003023 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003024 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003025 ":avx512f_ukernels",
3026 ],
3027)
3028
Marat Dukhan33fcf782020-05-24 14:27:15 -07003029xnnpack_aggregate_library(
3030 name = "ukernels_test_mode",
3031 aarch32_deps = [
3032 ":neon_ukernels_test_mode",
3033 ":neonfma_ukernels_test_mode",
3034 ":neonv8_ukernels_test_mode",
3035 ":asm_ukernels",
3036 ],
3037 aarch64_deps = [
3038 ":neon_ukernels_test_mode",
3039 ":neonfma_ukernels_test_mode",
3040 ":neonv8_ukernels_test_mode",
3041 ":neonfp16arith_ukernels_test_mode",
3042 ":asm_ukernels",
3043 ],
3044 generic_deps = [
3045 ":scalar_ukernels_test_mode",
3046 ],
3047 psimd_deps = [
3048 ":psimd_fastmath_ukernels_test_mode",
3049 ":psimd_accmath_ukernels_test_mode",
3050 ],
3051 wasm_deps = [
3052 ":wasm_ukernels_test_mode",
3053 ],
3054 wasmsimd_deps = [
3055 ":wasm_ukernels_test_mode",
3056 ],
3057 x86_deps = [
3058 ":sse2_ukernels_test_mode",
3059 ":ssse3_ukernels_test_mode",
3060 ":sse41_ukernels_test_mode",
3061 ":avx_ukernels_test_mode",
3062 ":fma3_ukernels_test_mode",
3063 ":avx2_ukernels_test_mode",
3064 ":avx512f_ukernels_test_mode",
3065 ],
3066)
3067
Marat Dukhan86df3842020-06-17 14:03:31 -07003068# This target must be compatible with Android NDK r17 and compile with Clang 6.
3069# It is similar to :ukernels target, but excludes microkernels which require
3070# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003071xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003072 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003073 aarch32_deps = [
3074 ":neon_ukernels",
3075 ":neonfma_ukernels",
3076 ":neonv8_ukernels",
3077 ":asm_ukernels",
3078 ],
3079 aarch64_deps = [
3080 ":neon_ukernels",
3081 ":neonfma_ukernels",
3082 ":neonv8_ukernels",
3083 ":asm_ukernels",
3084 ],
3085 generic_deps = [
3086 ":scalar_ukernels",
3087 ],
3088 psimd_deps = [
3089 ":psimd_fastmath_ukernels",
3090 ":psimd_accmath_ukernels",
3091 ],
3092 wasm_deps = [
3093 ":wasm_ukernels",
3094 ],
3095 wasmsimd_deps = [
3096 ":wasm_ukernels",
3097 ],
3098 x86_deps = [
3099 ":sse2_ukernels",
3100 ":ssse3_ukernels",
3101 ":sse41_ukernels",
3102 ":avx_ukernels",
3103 ":fma3_ukernels",
3104 ":avx2_ukernels",
3105 ":avx512f_ukernels",
3106 ],
3107)
3108
Marat Dukhan08c4a432019-10-03 09:29:21 -07003109xnnpack_cc_library(
3110 name = "im2col",
3111 srcs = ["src/im2col.c"],
3112 hdrs = [
3113 "src/xnnpack/common.h",
3114 "src/xnnpack/im2col.h",
3115 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003116 gcc_copts = xnnpack_gcc_std_copts(),
3117 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003118)
3119
3120xnnpack_cc_library(
3121 name = "indirection",
3122 srcs = ["src/indirection.c"],
3123 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003124 gcc_copts = xnnpack_gcc_std_copts(),
3125 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003126 deps = [
3127 "@FP16",
3128 "@FXdiv",
3129 "@pthreadpool",
3130 ],
3131)
3132
3133xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003134 name = "indirection_test_mode",
3135 srcs = ["src/indirection.c"],
3136 hdrs = INTERNAL_HDRS,
3137 copts = [
3138 "-UNDEBUG",
3139 "-DXNN_TEST_MODE=1",
3140 ],
3141 gcc_copts = xnnpack_gcc_std_copts(),
3142 msvc_copts = xnnpack_msvc_std_copts(),
3143 deps = [
3144 "@FP16",
3145 "@FXdiv",
3146 "@pthreadpool",
3147 ],
3148)
3149
3150xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003151 name = "packing",
3152 srcs = ["src/packing.c"],
3153 hdrs = INTERNAL_HDRS,
3154 gcc_copts = xnnpack_gcc_std_copts(),
3155 msvc_copts = xnnpack_msvc_std_copts(),
3156 deps = [
3157 "@FP16",
3158 "@FXdiv",
3159 "@pthreadpool",
3160 ],
3161)
3162
3163xnnpack_cc_library(
3164 name = "packing_test_mode",
3165 srcs = ["src/packing.c"],
3166 hdrs = INTERNAL_HDRS,
3167 copts = [
3168 "-UNDEBUG",
3169 "-DXNN_TEST_MODE=1",
3170 ],
3171 gcc_copts = xnnpack_gcc_std_copts(),
3172 msvc_copts = xnnpack_msvc_std_copts(),
3173 deps = [
3174 "@FP16",
3175 "@FXdiv",
3176 "@pthreadpool",
3177 ],
3178)
3179
3180xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003181 name = "operator_run",
3182 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003183 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003184 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003185 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3186 "//conditions:default": [],
3187 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003188 gcc_copts = xnnpack_gcc_std_copts(),
3189 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003190 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003191 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003192 "@FP16",
3193 "@FXdiv",
3194 "@clog",
3195 "@pthreadpool",
3196 ],
3197)
3198
Chao Mei6ddfc602020-05-13 22:29:36 -07003199xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003200 name = "operator_run_test_mode",
3201 srcs = ["src/operator-run.c"],
3202 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3203 copts = LOGGING_COPTS + [
3204 "-UNDEBUG",
3205 "-DXNN_TEST_MODE=1",
3206 ] + select({
3207 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3208 "//conditions:default": [],
3209 }),
3210 gcc_copts = xnnpack_gcc_std_copts(),
3211 msvc_copts = xnnpack_msvc_std_copts(),
3212 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003213 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003214 "@FP16",
3215 "@FXdiv",
3216 "@clog",
3217 "@pthreadpool",
3218 ],
3219)
3220
3221xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003222 name = "memory_planner",
3223 srcs = ["src/memory-planner.c"],
3224 hdrs = INTERNAL_HDRS,
3225 defines = select({
3226 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3227 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3228 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3229 }),
3230 gcc_copts = xnnpack_gcc_std_copts(),
3231 msvc_copts = xnnpack_msvc_std_copts(),
3232 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003233 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003234 "@pthreadpool",
3235 ],
3236)
3237
Marat Dukhan33fcf782020-05-24 14:27:15 -07003238xnnpack_cc_library(
3239 name = "memory_planner_test_mode",
3240 srcs = ["src/memory-planner.c"],
3241 hdrs = INTERNAL_HDRS,
3242 copts = [
3243 "-UNDEBUG",
3244 "-DXNN_TEST_MODE=1",
3245 ],
3246 defines = select({
3247 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3248 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3249 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3250 }),
3251 gcc_copts = xnnpack_gcc_std_copts(),
3252 msvc_copts = xnnpack_msvc_std_copts(),
3253 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003254 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003255 "@pthreadpool",
3256 ],
3257)
3258
Marat Dukhan08c4a432019-10-03 09:29:21 -07003259cc_library(
3260 name = "enable_assembly",
3261 defines = select({
3262 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3263 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003264 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003265 }),
3266)
3267
Marat Dukhan9de90e02020-06-18 16:04:12 -07003268cc_library(
3269 name = "enable_sparse",
3270 defines = select({
3271 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3272 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003273 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003274 }),
3275)
3276
Marat Dukhancf056b22019-10-07 10:26:29 -07003277xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003278 name = "operators",
3279 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003280 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003281 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003282 ],
3283 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003284 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -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 Dukhan08c4a432019-10-03 09:29:21 -07003293 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003294 gcc_copts = xnnpack_gcc_std_copts(),
3295 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003296 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003297 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003298 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003299 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003300 "@FP16",
3301 "@FXdiv",
3302 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003303 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003304 ],
3305)
3306
Marat Dukhan10a38082020-04-17 03:58:35 -07003307xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003308 name = "operators_test_mode",
3309 srcs = OPERATOR_SRCS + [
3310 "src/memory.c",
3311 "src/operator-delete.c",
3312 ],
3313 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3314 copts = LOGGING_COPTS + [
3315 "-Isrc",
3316 "-Iinclude",
3317 "-UNDEBUG",
3318 "-DXNN_TEST_MODE=1",
3319 ] + select({
3320 ":debug_build": [],
3321 "//conditions:default": xnnpack_min_size_copts(),
3322 }) + select({
3323 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3324 "//conditions:default": [],
3325 }),
3326 gcc_copts = xnnpack_gcc_std_copts(),
3327 msvc_copts = xnnpack_msvc_std_copts(),
3328 deps = [
3329 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003330 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003331 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003332 "@FP16",
3333 "@FXdiv",
3334 "@clog",
3335 "@pthreadpool",
3336 ],
3337)
3338
3339xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003340 name = "XNNPACK",
3341 srcs = [
3342 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003343 "src/runtime.c",
3344 "src/subgraph.c",
3345 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003346 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003347 hdrs = ["include/xnnpack.h"],
3348 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003349 "-Isrc",
3350 "-Iinclude",
3351 ] + select({
3352 ":debug_build": [],
3353 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003354 }) + select({
3355 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3356 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003357 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003358 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003359 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003360 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003361 visibility = xnnpack_visibility(),
3362 deps = [
3363 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003364 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003365 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003366 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003367 ":operator_run",
3368 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003369 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003370 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003371 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003372 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003373 ] + select({
3374 ":emscripten": [],
3375 "//conditions:default": ["@cpuinfo"],
3376 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003377)
3378
Marat Dukhan10a38082020-04-17 03:58:35 -07003379xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003380 name = "XNNPACK_test_mode",
3381 srcs = [
3382 "src/init.c",
3383 "src/runtime.c",
3384 "src/subgraph.c",
3385 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003386 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003387 hdrs = ["include/xnnpack.h"],
3388 copts = LOGGING_COPTS + [
3389 "-Isrc",
3390 "-Iinclude",
3391 "-UNDEBUG",
3392 "-DXNN_TEST_MODE=1",
3393 ] + select({
3394 ":debug_build": [],
3395 "//conditions:default": xnnpack_min_size_copts(),
3396 }) + select({
3397 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3398 "//conditions:default": [],
3399 }),
3400 gcc_copts = xnnpack_gcc_std_copts(),
3401 includes = ["include"],
3402 msvc_copts = xnnpack_msvc_std_copts(),
3403 visibility = xnnpack_visibility(),
3404 deps = [
3405 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003406 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003407 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003408 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003409 ":operator_run_test_mode",
3410 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003411 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003412 "@clog",
3413 "@FP16",
3414 "@pthreadpool",
3415 ] + select({
3416 ":emscripten": [],
3417 "//conditions:default": ["@cpuinfo"],
3418 }),
3419)
3420
3421xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003422 name = "xnnpack_f32",
3423 srcs = [
3424 "src/init.c",
3425 "src/runtime.c",
3426 "src/subgraph.c",
3427 "src/tensor.c",
3428 ] + SUBGRAPH_SRCS,
3429 hdrs = ["include/xnnpack.h"],
3430 copts = LOGGING_COPTS + [
3431 "-Isrc",
3432 "-Iinclude",
3433 ] + select({
3434 ":debug_build": [],
3435 "//conditions:default": xnnpack_min_size_copts(),
3436 }) + select({
3437 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3438 "//conditions:default": [],
3439 }),
3440 defines = [
3441 "XNN_NO_Q8_OPERATORS",
3442 "XNN_NO_U8_OPERATORS",
3443 "XNN_NO_X8_OPERATORS",
3444 "XNN_NO_F16_OPERATORS",
3445 "XNN_NO_X16_OPERATORS",
3446 ],
3447 gcc_copts = xnnpack_gcc_std_copts(),
3448 includes = ["include"],
3449 msvc_copts = xnnpack_msvc_std_copts(),
3450 visibility = xnnpack_visibility(),
3451 deps = [
3452 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003453 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003454 ":logging_utils",
3455 ":memory_planner",
3456 ":operator_run",
3457 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003458 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003459 "@clog",
3460 "@FP16",
3461 "@pthreadpool",
3462 ] + select({
3463 ":emscripten": [],
3464 "//conditions:default": ["@cpuinfo"],
3465 }),
3466)
3467
3468xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003469 name = "xnnpack_operators_nhwc_f32",
3470 srcs = [
3471 "src/init.c",
3472 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003473 hdrs = ["include/xnnpack.h"],
3474 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003475 "-Isrc",
3476 "-Iinclude",
3477 ] + select({
3478 ":debug_build": [],
3479 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003480 }) + select({
3481 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3482 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003483 }),
3484 defines = [
3485 "XNN_NO_Q8_OPERATORS",
3486 "XNN_NO_U8_OPERATORS",
3487 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003488 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003489 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003490 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003491 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003492 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003493 visibility = xnnpack_visibility(),
3494 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003495 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003496 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003497 ":operator_run",
3498 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003499 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003500 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003501 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003502 ] + select({
3503 ":emscripten": [],
3504 "//conditions:default": ["@cpuinfo"],
3505 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003506)
3507
Marat Dukhancf056b22019-10-07 10:26:29 -07003508xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003509 name = "bench_utils",
3510 srcs = ["bench/utils.cc"],
3511 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003512 deps = [
3513 "@com_google_benchmark//:benchmark",
3514 "@cpuinfo",
3515 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003516)
3517
Frank Barchard7e955972019-10-11 10:34:25 -07003518######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003519
3520xnnpack_benchmark(
3521 name = "q8_gemm_bench",
3522 srcs = [
3523 "bench/gemm.h",
3524 "bench/q8-gemm.cc",
3525 "src/xnnpack/AlignedAllocator.h",
3526 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003527 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003528 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003529)
3530
3531xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003532 name = "f16_igemm_bench",
3533 srcs = [
3534 "bench/f16-igemm.cc",
3535 "bench/conv.h",
3536 "bench/google/conv.h",
3537 "src/xnnpack/AlignedAllocator.h",
3538 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003539 deps = MICROKERNEL_BENCHMARK_DEPS + [
3540 ":indirection",
3541 ":packing",
3542 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003543)
3544
3545xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003546 name = "f16_gemm_bench",
3547 srcs = [
3548 "bench/f16-gemm.cc",
3549 "bench/gemm.h",
3550 "src/xnnpack/AlignedAllocator.h",
3551 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003552 deps = MICROKERNEL_BENCHMARK_DEPS + [
3553 ":packing",
3554 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003555)
3556
3557xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003558 name = "f16_spmm_bench",
3559 srcs = [
3560 "bench/f16-spmm.cc",
3561 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003562 "src/xnnpack/AlignedAllocator.h",
3563 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003564 deps = MICROKERNEL_BENCHMARK_DEPS,
3565)
3566
3567xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003568 name = "f32_igemm_bench",
3569 srcs = [
3570 "bench/f32-igemm.cc",
3571 "bench/conv.h",
3572 "src/xnnpack/AlignedAllocator.h",
3573 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003574 deps = MICROKERNEL_BENCHMARK_DEPS + [
3575 ":indirection",
3576 ":packing",
3577 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003578)
3579
3580xnnpack_benchmark(
3581 name = "f32_conv_hwc_bench",
3582 srcs = [
3583 "bench/f32-conv-hwc.cc",
3584 "bench/dconv.h",
3585 "src/xnnpack/AlignedAllocator.h",
3586 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003587 deps = MICROKERNEL_BENCHMARK_DEPS + [
3588 ":packing",
3589 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003590)
3591
3592xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003593 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003594 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003595 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003596 "bench/dconv.h",
3597 "src/xnnpack/AlignedAllocator.h",
3598 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003599 deps = MICROKERNEL_BENCHMARK_DEPS + [
3600 ":packing",
3601 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003602)
3603
3604xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003605 name = "f16_dwconv_bench",
3606 srcs = [
3607 "bench/f16-dwconv.cc",
3608 "bench/dwconv.h",
3609 "bench/google/dwconv.h",
3610 "src/xnnpack/AlignedAllocator.h",
3611 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003612 deps = MICROKERNEL_BENCHMARK_DEPS + [
3613 ":indirection",
3614 ":packing",
3615 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003616)
3617
3618xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003619 name = "f32_dwconv_bench",
3620 srcs = [
3621 "bench/f32-dwconv.cc",
3622 "bench/dwconv.h",
3623 "src/xnnpack/AlignedAllocator.h",
3624 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003625 deps = MICROKERNEL_BENCHMARK_DEPS + [
3626 ":indirection",
3627 ":packing",
3628 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003629)
3630
3631xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003632 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003633 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003634 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003635 "bench/dwconv.h",
3636 "src/xnnpack/AlignedAllocator.h",
3637 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003638 deps = MICROKERNEL_BENCHMARK_DEPS + [
3639 ":indirection",
3640 ":packing",
3641 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003642)
3643
3644xnnpack_benchmark(
3645 name = "f32_gemm_bench",
3646 srcs = [
3647 "bench/f32-gemm.cc",
3648 "bench/gemm.h",
3649 "src/xnnpack/AlignedAllocator.h",
3650 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003651 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003652 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003653)
3654
3655xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003656 name = "f32_hswish_bench",
3657 srcs = [
3658 "bench/f32-hswish.cc",
3659 "src/xnnpack/AlignedAllocator.h",
3660 ] + MICROKERNEL_BENCHMARK_HDRS,
3661 deps = MICROKERNEL_BENCHMARK_DEPS,
3662)
3663
3664xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003665 name = "f32_raddexpminusmax_bench",
3666 srcs = [
3667 "bench/f32-raddexpminusmax.cc",
3668 "src/xnnpack/AlignedAllocator.h",
3669 ] + MICROKERNEL_BENCHMARK_HDRS,
3670 deps = MICROKERNEL_BENCHMARK_DEPS,
3671)
3672
3673xnnpack_benchmark(
3674 name = "f32_raddextexp_bench",
3675 srcs = [
3676 "bench/f32-raddextexp.cc",
3677 "src/xnnpack/AlignedAllocator.h",
3678 ] + MICROKERNEL_BENCHMARK_HDRS,
3679 deps = MICROKERNEL_BENCHMARK_DEPS,
3680)
3681
3682xnnpack_benchmark(
3683 name = "f32_raddstoreexpminusmax_bench",
3684 srcs = [
3685 "bench/f32-raddstoreexpminusmax.cc",
3686 "src/xnnpack/AlignedAllocator.h",
3687 ] + MICROKERNEL_BENCHMARK_HDRS,
3688 deps = MICROKERNEL_BENCHMARK_DEPS,
3689)
3690
3691xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003692 name = "f32_rmax_bench",
3693 srcs = [
3694 "bench/f32-rmax.cc",
3695 "src/xnnpack/AlignedAllocator.h",
3696 ] + MICROKERNEL_BENCHMARK_HDRS,
3697 deps = MICROKERNEL_BENCHMARK_DEPS,
3698)
3699
3700xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003701 name = "f32_sigmoid_bench",
3702 srcs = [
3703 "bench/f32-sigmoid.cc",
3704 "src/xnnpack/AlignedAllocator.h",
3705 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003706 deps = MICROKERNEL_BENCHMARK_DEPS,
3707)
3708
3709xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003710 name = "f32_spmm_bench",
3711 srcs = [
3712 "bench/f32-spmm.cc",
3713 "bench/gemm.h",
3714 "src/xnnpack/AlignedAllocator.h",
3715 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003716 deps = MICROKERNEL_BENCHMARK_DEPS,
3717)
3718
3719xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003720 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003721 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003722 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003723 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003724 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003725 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003726)
3727
3728xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003729 name = "f32_vscaleexpminusmax_bench",
3730 srcs = [
3731 "bench/f32-vscaleexpminusmax.cc",
3732 "src/xnnpack/AlignedAllocator.h",
3733 ] + MICROKERNEL_BENCHMARK_HDRS,
3734 deps = MICROKERNEL_BENCHMARK_DEPS,
3735)
3736
3737xnnpack_benchmark(
3738 name = "f32_vscaleextexp_bench",
3739 srcs = [
3740 "bench/f32-vscaleextexp.cc",
3741 "src/xnnpack/AlignedAllocator.h",
3742 ] + MICROKERNEL_BENCHMARK_HDRS,
3743 deps = MICROKERNEL_BENCHMARK_DEPS,
3744)
3745
3746xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003747 name = "f32_vsqrt_bench",
3748 srcs = [
3749 "bench/f32-vsqrt.cc",
3750 "src/xnnpack/AlignedAllocator.h",
3751 ] + MICROKERNEL_BENCHMARK_HDRS,
3752 deps = MICROKERNEL_BENCHMARK_DEPS,
3753)
3754
3755xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003756 name = "f32_im2col_gemm_bench",
3757 srcs = [
3758 "bench/f32-im2col-gemm.cc",
3759 "bench/conv.h",
3760 "src/xnnpack/AlignedAllocator.h",
3761 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003762 deps = MICROKERNEL_BENCHMARK_DEPS + [
3763 ":im2col",
3764 ":packing",
3765 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003766)
3767
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003768xnnpack_benchmark(
3769 name = "requantization_bench",
3770 srcs = [
3771 "bench/requantization.cc",
3772 "src/xnnpack/requantization-stubs.h",
3773 "src/xnnpack/AlignedAllocator.h",
3774 ] + MICROKERNEL_BENCHMARK_HDRS,
3775 deps = MICROKERNEL_BENCHMARK_DEPS,
3776)
3777
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003778xnnpack_benchmark(
3779 name = "rounding_bench",
3780 srcs = [
3781 "bench/rounding.cc",
3782 "src/xnnpack/math-stubs.h",
3783 "src/xnnpack/AlignedAllocator.h",
3784 ] + MICROKERNEL_BENCHMARK_HDRS,
3785 deps = MICROKERNEL_BENCHMARK_DEPS,
3786)
3787
Marat Dukhan08c4a432019-10-03 09:29:21 -07003788########################### Benchmarks for operators ###########################
3789
3790xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003791 name = "average_pooling_bench",
3792 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003793 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003794 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003795 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003796)
3797
3798xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003799 name = "bankers_rounding_bench",
3800 srcs = ["bench/bankers-rounding.cc"],
3801 copts = xnnpack_optional_tflite_copts(),
3802 tags = ["nowin32"],
3803 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3804)
3805
3806xnnpack_benchmark(
3807 name = "ceiling_bench",
3808 srcs = ["bench/ceiling.cc"],
3809 copts = xnnpack_optional_tflite_copts(),
3810 tags = ["nowin32"],
3811 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3812)
3813
3814xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003815 name = "channel_shuffle_bench",
3816 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003817 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003818)
3819
3820xnnpack_benchmark(
3821 name = "convolution_bench",
3822 srcs = ["bench/convolution.cc"],
3823 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003824 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003825 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003826)
3827
3828xnnpack_benchmark(
3829 name = "deconvolution_bench",
3830 srcs = ["bench/deconvolution.cc"],
3831 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003832 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003833 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003834)
3835
3836xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003837 name = "floor_bench",
3838 srcs = ["bench/floor.cc"],
3839 copts = xnnpack_optional_tflite_copts(),
3840 tags = ["nowin32"],
3841 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3842)
3843
3844xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003845 name = "global_average_pooling_bench",
3846 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003847 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003848)
3849
3850xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07003851 name = "hardswish_bench",
3852 srcs = ["bench/hardswish.cc"],
3853 copts = xnnpack_optional_tflite_copts(),
3854 tags = ["nowin32"],
3855 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3856)
3857
3858xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003859 name = "max_pooling_bench",
3860 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003861 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003862)
3863
3864xnnpack_benchmark(
3865 name = "sigmoid_bench",
3866 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003867 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003868 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003869 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003870)
3871
3872xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003873 name = "prelu_bench",
3874 srcs = ["bench/prelu.cc"],
3875 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003876 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003877 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003878)
3879
3880xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003881 name = "softmax_bench",
3882 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003883 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003884 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003885 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003886)
3887
Marat Dukhan87727142020-06-24 15:24:10 -07003888xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07003889 name = "square_root_bench",
3890 srcs = ["bench/square-root.cc"],
3891 copts = xnnpack_optional_tflite_copts(),
3892 tags = ["nowin32"],
3893 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3894)
3895
3896xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003897 name = "truncation_bench",
3898 srcs = ["bench/truncation.cc"],
3899 deps = OPERATOR_BENCHMARK_DEPS,
3900)
3901
Marat Dukhanc068bb62019-10-04 13:24:39 -07003902############################# End-to-end benchmarks ############################
3903
3904cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003905 name = "fp32_mobilenet_v1",
3906 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003907 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003908 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003909 linkstatic = True,
3910 deps = [
3911 ":XNNPACK",
3912 "@pthreadpool",
3913 ],
3914)
3915
3916cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003917 name = "fp16_mobilenet_v1",
3918 srcs = ["models/fp16-mobilenet-v1.cc"],
3919 hdrs = ["models/models.h"],
3920 copts = xnnpack_std_cxxopts(),
3921 linkstatic = True,
3922 deps = [
3923 ":XNNPACK",
3924 "@FP16",
3925 "@pthreadpool",
3926 ],
3927)
3928
3929cc_library(
3930 name = "fp32_mobilenet_v2",
3931 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003932 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003933 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003934 linkstatic = True,
3935 deps = [
3936 ":XNNPACK",
3937 "@pthreadpool",
3938 ],
3939)
3940
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003941cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003942 name = "fp16_mobilenet_v2",
3943 srcs = ["models/fp16-mobilenet-v2.cc"],
3944 hdrs = ["models/models.h"],
3945 copts = xnnpack_std_cxxopts(),
3946 linkstatic = True,
3947 deps = [
3948 ":XNNPACK",
3949 "@FP16",
3950 "@pthreadpool",
3951 ],
3952)
3953
3954cc_library(
3955 name = "fp32_mobilenet_v3_large",
3956 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003957 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003958 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003959 linkstatic = True,
3960 deps = [
3961 ":XNNPACK",
3962 "@pthreadpool",
3963 ],
3964)
3965
3966cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003967 name = "fp32_mobilenet_v3_small",
3968 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003969 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003970 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003971 linkstatic = True,
3972 deps = [
3973 ":XNNPACK",
3974 "@pthreadpool",
3975 ],
3976)
3977
Marat Dukhanc068bb62019-10-04 13:24:39 -07003978xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07003979 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003980 srcs = [
3981 "bench/f32-dwconv-e2e.cc",
3982 "bench/end2end.h",
3983 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07003984 deps = MICROKERNEL_BENCHMARK_DEPS + [
3985 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003986 ":fp32_mobilenet_v1",
3987 ":fp32_mobilenet_v2",
3988 ":fp32_mobilenet_v3_large",
3989 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07003990 ],
3991)
3992
3993xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07003994 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003995 srcs = [
3996 "bench/f32-gemm-e2e.cc",
3997 "bench/end2end.h",
3998 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07003999 deps = MICROKERNEL_BENCHMARK_DEPS + [
4000 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004001 ":fp32_mobilenet_v1",
4002 ":fp32_mobilenet_v2",
4003 ":fp32_mobilenet_v3_large",
4004 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004005 ],
4006)
4007
4008xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004009 name = "end2end_bench",
4010 srcs = ["bench/end2end.cc"],
4011 deps = [
4012 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004013 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004014 ":fp16_mobilenet_v1",
4015 ":fp16_mobilenet_v2",
4016 ":fp32_mobilenet_v1",
4017 ":fp32_mobilenet_v2",
4018 ":fp32_mobilenet_v3_large",
4019 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004020 "@pthreadpool",
4021 ],
4022)
4023
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004024#################### Accuracy evaluation for math functions ####################
4025
4026xnnpack_benchmark(
4027 name = "f32_exp_eval",
4028 srcs = [
4029 "eval/f32-exp.cc",
4030 "src/xnnpack/AlignedAllocator.h",
4031 ] + ACCURACY_EVAL_HDRS,
4032 deps = ACCURACY_EVAL_DEPS,
4033)
4034
Marat Dukhan515c9772019-10-17 18:07:57 -07004035xnnpack_benchmark(
4036 name = "f32_expminus_eval",
4037 srcs = [
4038 "eval/f32-expminus.cc",
4039 "src/xnnpack/AlignedAllocator.h",
4040 ] + ACCURACY_EVAL_HDRS,
4041 deps = ACCURACY_EVAL_DEPS,
4042)
4043
Marat Dukhan98ba4412019-10-23 02:14:28 -07004044xnnpack_benchmark(
4045 name = "f32_extexp_eval",
4046 srcs = [
4047 "eval/f32-extexp.cc",
4048 "src/xnnpack/AlignedAllocator.h",
4049 ] + ACCURACY_EVAL_HDRS,
4050 deps = ACCURACY_EVAL_DEPS,
4051)
4052
Marat Dukhan8853b822020-05-07 12:19:01 -07004053xnnpack_unit_test(
4054 name = "f32_roundne_eval",
4055 srcs = [
4056 "eval/f32-roundne.cc",
4057 "src/xnnpack/AlignedAllocator.h",
4058 "src/xnnpack/math-stubs.h",
4059 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004060 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004061 deps = MICROKERNEL_TEST_DEPS,
4062)
4063
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004064xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004065 name = "f32_roundd_eval",
4066 srcs = [
4067 "eval/f32-roundd.cc",
4068 "src/xnnpack/AlignedAllocator.h",
4069 "src/xnnpack/math-stubs.h",
4070 ] + MICROKERNEL_TEST_HDRS,
4071 automatic = False,
4072 deps = MICROKERNEL_TEST_DEPS,
4073)
4074
4075xnnpack_unit_test(
4076 name = "f32_roundu_eval",
4077 srcs = [
4078 "eval/f32-roundu.cc",
4079 "src/xnnpack/AlignedAllocator.h",
4080 "src/xnnpack/math-stubs.h",
4081 ] + MICROKERNEL_TEST_HDRS,
4082 automatic = False,
4083 deps = MICROKERNEL_TEST_DEPS,
4084)
4085
4086xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004087 name = "f32_roundz_eval",
4088 srcs = [
4089 "eval/f32-roundz.cc",
4090 "src/xnnpack/AlignedAllocator.h",
4091 "src/xnnpack/math-stubs.h",
4092 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004093 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004094 deps = MICROKERNEL_TEST_DEPS,
4095)
4096
Marat Dukhan346a9e52019-11-15 09:06:30 -08004097xnnpack_benchmark(
4098 name = "f32_sigmoid_eval",
4099 srcs = [
4100 "eval/f32-sigmoid.cc",
4101 "src/xnnpack/AlignedAllocator.h",
4102 ] + ACCURACY_EVAL_HDRS,
4103 deps = ACCURACY_EVAL_DEPS,
4104)
4105
Marat Dukhan84000762020-06-29 18:38:43 -07004106xnnpack_benchmark(
4107 name = "f32_sqrt_eval",
4108 srcs = [
4109 "eval/f32-sqrt.cc",
4110 "src/xnnpack/AlignedAllocator.h",
4111 ] + ACCURACY_EVAL_HDRS,
4112 deps = ACCURACY_EVAL_DEPS + [
4113 ":bench_utils",
4114 ],
4115)
4116
Marat Dukhan08c4a432019-10-03 09:29:21 -07004117######################### Unit tests for micro-kernels #########################
4118
4119xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004120 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004121 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004122 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004123 "test/gemm-microkernel-tester.h",
4124 "src/xnnpack/AlignedAllocator.h",
4125 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004126 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004127)
4128
4129xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004130 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004131 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004132 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004133 "test/spmm-microkernel-tester.h",
4134 "src/xnnpack/AlignedAllocator.h",
4135 ] + MICROKERNEL_TEST_HDRS,
4136 deps = MICROKERNEL_TEST_DEPS,
4137)
4138
4139xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004140 name = "f32_argmaxpool_test",
4141 srcs = [
4142 "test/f32-argmaxpool.cc",
4143 "test/argmaxpool-microkernel-tester.h",
4144 "src/xnnpack/AlignedAllocator.h",
4145 ] + MICROKERNEL_TEST_HDRS,
4146 deps = MICROKERNEL_TEST_DEPS,
4147)
4148
4149xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004150 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004151 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004152 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004153 "test/avgpool-microkernel-tester.h",
4154 "src/xnnpack/AlignedAllocator.h",
4155 ] + MICROKERNEL_TEST_HDRS,
4156 deps = MICROKERNEL_TEST_DEPS,
4157)
4158
4159xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004160 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004161 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004162 "test/f32-ibilinear.cc",
4163 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004164 "src/xnnpack/AlignedAllocator.h",
4165 ] + MICROKERNEL_TEST_HDRS,
4166 deps = MICROKERNEL_TEST_DEPS,
4167)
4168
4169xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004170 name = "f16_clamp_test",
4171 srcs = [
4172 "test/f16-clamp.cc",
4173 "test/clamp-microkernel-tester.h",
4174 ] + MICROKERNEL_TEST_HDRS,
4175 deps = MICROKERNEL_TEST_DEPS,
4176)
4177
4178xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004179 name = "f32_clamp_test",
4180 srcs = [
4181 "test/f32-clamp.cc",
4182 "test/clamp-microkernel-tester.h",
4183 ] + MICROKERNEL_TEST_HDRS,
4184 deps = MICROKERNEL_TEST_DEPS,
4185)
4186
4187xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004188 name = "f32_igemm_test",
4189 srcs = [
4190 "test/f32-igemm.cc",
4191 "test/gemm-microkernel-tester.h",
4192 "src/xnnpack/AlignedAllocator.h",
4193 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004194 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004195)
4196
4197xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004198 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004199 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004200 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004201 "test/gemm-microkernel-tester.h",
4202 "src/xnnpack/AlignedAllocator.h",
4203 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004204 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004205)
4206
4207xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004208 name = "f32_igemm_minmax_test",
4209 srcs = [
4210 "test/f32-igemm-minmax.cc",
4211 "test/gemm-microkernel-tester.h",
4212 "src/xnnpack/AlignedAllocator.h",
4213 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004214 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004215)
4216
4217xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004218 name = "f16_igemm_minmax_test",
4219 srcs = [
4220 "test/f16-igemm-minmax.cc",
4221 "test/gemm-microkernel-tester.h",
4222 "src/xnnpack/AlignedAllocator.h",
4223 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004224 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004225)
4226
4227xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004228 name = "f32_conv_hwc_test",
4229 srcs = [
4230 "test/f32-conv-hwc.cc",
4231 "test/conv-hwc-microkernel-tester.h",
4232 "src/xnnpack/AlignedAllocator.h",
4233 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004234 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004235)
4236
4237xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004238 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004239 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004240 "test/f32-conv-hwc2chw.cc",
4241 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004242 "src/xnnpack/AlignedAllocator.h",
4243 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004244 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004245)
4246
4247xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004248 name = "f32_dwconv_test",
4249 srcs = [
4250 "test/f32-dwconv.cc",
4251 "test/dwconv-microkernel-tester.h",
4252 "src/xnnpack/AlignedAllocator.h",
4253 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004254 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004255)
4256
4257xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004258 name = "f16_dwconv_minmax_test",
4259 srcs = [
4260 "test/f16-dwconv-minmax.cc",
4261 "test/dwconv-microkernel-tester.h",
4262 "src/xnnpack/AlignedAllocator.h",
4263 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004264 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004265)
4266
4267xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004268 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004269 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004270 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004271 "test/dwconv-microkernel-tester.h",
4272 "src/xnnpack/AlignedAllocator.h",
4273 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004274 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004275)
4276
4277xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004278 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004279 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004280 "test/f32-dwconv-chw.cc",
4281 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004282 "src/xnnpack/AlignedAllocator.h",
4283 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004284 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004285)
4286
4287xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004288 name = "f16_gavgpool_minmax_test",
4289 srcs = [
4290 "test/f16-gavgpool-minmax.cc",
4291 "test/gavgpool-microkernel-tester.h",
4292 "src/xnnpack/AlignedAllocator.h",
4293 ] + MICROKERNEL_TEST_HDRS,
4294 deps = MICROKERNEL_TEST_DEPS,
4295)
4296
4297xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004298 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004299 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004300 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004301 "test/gavgpool-microkernel-tester.h",
4302 "src/xnnpack/AlignedAllocator.h",
4303 ] + MICROKERNEL_TEST_HDRS,
4304 deps = MICROKERNEL_TEST_DEPS,
4305)
4306
4307xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004308 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004309 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004310 "test/f32-gavgpool-cw.cc",
4311 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004312 "src/xnnpack/AlignedAllocator.h",
4313 ] + MICROKERNEL_TEST_HDRS,
4314 deps = MICROKERNEL_TEST_DEPS,
4315)
4316
4317xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004318 name = "f32_gemm_test",
4319 srcs = [
4320 "test/f32-gemm.cc",
4321 "test/gemm-microkernel-tester.h",
4322 "src/xnnpack/AlignedAllocator.h",
4323 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004324 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004325)
4326
4327xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004328 name = "f32_gemm_relu_test",
4329 srcs = [
4330 "test/f32-gemm-relu.cc",
4331 "test/gemm-microkernel-tester.h",
4332 "src/xnnpack/AlignedAllocator.h",
4333 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004334 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004335)
4336
4337xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004338 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004339 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004340 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004341 "test/gemm-microkernel-tester.h",
4342 "src/xnnpack/AlignedAllocator.h",
4343 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004344 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004345)
4346
4347xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004348 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004349 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004350 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004351 "test/gemm-microkernel-tester.h",
4352 "src/xnnpack/AlignedAllocator.h",
4353 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004354 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004355)
4356
4357xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004358 name = "f16_hswish_test",
4359 srcs = [
4360 "test/f16-hswish.cc",
4361 "test/hswish-microkernel-tester.h",
4362 ] + MICROKERNEL_TEST_HDRS,
4363 deps = MICROKERNEL_TEST_DEPS,
4364)
4365
4366xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004367 name = "f32_hswish_test",
4368 srcs = [
4369 "test/f32-hswish.cc",
4370 "test/hswish-microkernel-tester.h",
4371 ] + MICROKERNEL_TEST_HDRS,
4372 deps = MICROKERNEL_TEST_DEPS,
4373)
4374
4375xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004376 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004377 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004378 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004379 "test/maxpool-microkernel-tester.h",
4380 ] + MICROKERNEL_TEST_HDRS,
4381 deps = MICROKERNEL_TEST_DEPS,
4382)
4383
4384xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004385 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004386 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004387 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004388 "test/avgpool-microkernel-tester.h",
4389 "src/xnnpack/AlignedAllocator.h",
4390 ] + MICROKERNEL_TEST_HDRS,
4391 deps = MICROKERNEL_TEST_DEPS,
4392)
4393
4394xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004395 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004396 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004397 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004398 "test/gemm-microkernel-tester.h",
4399 "src/xnnpack/AlignedAllocator.h",
4400 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004401 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004402)
4403
4404xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004405 name = "f16_prelu_test",
4406 srcs = [
4407 "test/f16-prelu.cc",
4408 "test/prelu-microkernel-tester.h",
4409 "src/xnnpack/AlignedAllocator.h",
4410 ] + MICROKERNEL_TEST_HDRS,
4411 deps = MICROKERNEL_TEST_DEPS,
4412)
4413
4414xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004415 name = "f32_prelu_test",
4416 srcs = [
4417 "test/f32-prelu.cc",
4418 "test/prelu-microkernel-tester.h",
4419 "src/xnnpack/AlignedAllocator.h",
4420 ] + MICROKERNEL_TEST_HDRS,
4421 deps = MICROKERNEL_TEST_DEPS,
4422)
4423
4424xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004425 name = "f32_raddexpminusmax_test",
4426 srcs = [
4427 "test/f32-raddexpminusmax.cc",
4428 "test/raddexpminusmax-microkernel-tester.h",
4429 ] + MICROKERNEL_TEST_HDRS,
4430 deps = MICROKERNEL_TEST_DEPS,
4431)
4432
4433xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004434 name = "f32_raddextexp_test",
4435 srcs = [
4436 "test/f32-raddextexp.cc",
4437 "test/raddextexp-microkernel-tester.h",
4438 ] + MICROKERNEL_TEST_HDRS,
4439 deps = MICROKERNEL_TEST_DEPS,
4440)
4441
4442xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004443 name = "f32_raddstoreexpminusmax_test",
4444 srcs = [
4445 "test/f32-raddstoreexpminusmax.cc",
4446 "test/raddstoreexpminusmax-microkernel-tester.h",
4447 ] + MICROKERNEL_TEST_HDRS,
4448 deps = MICROKERNEL_TEST_DEPS,
4449)
4450
4451xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004452 name = "f32_rmax_test",
4453 srcs = [
4454 "test/f32-rmax.cc",
4455 "test/rmax-microkernel-tester.h",
4456 ] + MICROKERNEL_TEST_HDRS,
4457 deps = MICROKERNEL_TEST_DEPS,
4458)
4459
4460xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004461 name = "f32_sigmoid_test",
4462 srcs = [
4463 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004464 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004465 ] + MICROKERNEL_TEST_HDRS,
4466 deps = MICROKERNEL_TEST_DEPS,
4467)
4468
4469xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004470 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004471 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004472 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004473 "test/spmm-microkernel-tester.h",
4474 "src/xnnpack/AlignedAllocator.h",
4475 ] + MICROKERNEL_TEST_HDRS,
4476 deps = MICROKERNEL_TEST_DEPS,
4477)
4478
4479xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004480 name = "f16_vadd_minmax_test",
4481 srcs = [
4482 "test/f16-vadd-minmax.cc",
4483 "test/vbinary-microkernel-tester.h",
4484 ] + MICROKERNEL_TEST_HDRS,
4485 deps = MICROKERNEL_TEST_DEPS,
4486)
4487
4488xnnpack_unit_test(
4489 name = "f16_vaddc_minmax_test",
4490 srcs = [
4491 "test/f16-vaddc-minmax.cc",
4492 "test/vbinaryc-microkernel-tester.h",
4493 ] + MICROKERNEL_TEST_HDRS,
4494 deps = MICROKERNEL_TEST_DEPS,
4495)
4496
4497xnnpack_unit_test(
4498 name = "f16_vdiv_minmax_test",
4499 srcs = [
4500 "test/f16-vdiv-minmax.cc",
4501 "test/vbinary-microkernel-tester.h",
4502 ] + MICROKERNEL_TEST_HDRS,
4503 deps = MICROKERNEL_TEST_DEPS,
4504)
4505
4506xnnpack_unit_test(
4507 name = "f16_vdivc_minmax_test",
4508 srcs = [
4509 "test/f16-vdivc-minmax.cc",
4510 "test/vbinaryc-microkernel-tester.h",
4511 ] + MICROKERNEL_TEST_HDRS,
4512 deps = MICROKERNEL_TEST_DEPS,
4513)
4514
4515xnnpack_unit_test(
4516 name = "f16_vrdivc_minmax_test",
4517 srcs = [
4518 "test/f16-vrdivc-minmax.cc",
4519 "test/vbinaryc-microkernel-tester.h",
4520 ] + MICROKERNEL_TEST_HDRS,
4521 deps = MICROKERNEL_TEST_DEPS,
4522)
4523
4524xnnpack_unit_test(
4525 name = "f16_vmax_test",
4526 srcs = [
4527 "test/f16-vmax.cc",
4528 "test/vbinary-microkernel-tester.h",
4529 ] + MICROKERNEL_TEST_HDRS,
4530 deps = MICROKERNEL_TEST_DEPS,
4531)
4532
4533xnnpack_unit_test(
4534 name = "f16_vmaxc_test",
4535 srcs = [
4536 "test/f16-vmaxc.cc",
4537 "test/vbinaryc-microkernel-tester.h",
4538 ] + MICROKERNEL_TEST_HDRS,
4539 deps = MICROKERNEL_TEST_DEPS,
4540)
4541
4542xnnpack_unit_test(
4543 name = "f16_vmin_test",
4544 srcs = [
4545 "test/f16-vmin.cc",
4546 "test/vbinary-microkernel-tester.h",
4547 ] + MICROKERNEL_TEST_HDRS,
4548 deps = MICROKERNEL_TEST_DEPS,
4549)
4550
4551xnnpack_unit_test(
4552 name = "f16_vminc_test",
4553 srcs = [
4554 "test/f16-vminc.cc",
4555 "test/vbinaryc-microkernel-tester.h",
4556 ] + MICROKERNEL_TEST_HDRS,
4557 deps = MICROKERNEL_TEST_DEPS,
4558)
4559
4560xnnpack_unit_test(
4561 name = "f16_vmul_minmax_test",
4562 srcs = [
4563 "test/f16-vmul-minmax.cc",
4564 "test/vbinary-microkernel-tester.h",
4565 ] + MICROKERNEL_TEST_HDRS,
4566 deps = MICROKERNEL_TEST_DEPS,
4567)
4568
4569xnnpack_unit_test(
4570 name = "f16_vmulc_minmax_test",
4571 srcs = [
4572 "test/f16-vmulc-minmax.cc",
4573 "test/vbinaryc-microkernel-tester.h",
4574 ] + MICROKERNEL_TEST_HDRS,
4575 deps = MICROKERNEL_TEST_DEPS,
4576)
4577
4578xnnpack_unit_test(
4579 name = "f16_vsub_minmax_test",
4580 srcs = [
4581 "test/f16-vsub-minmax.cc",
4582 "test/vbinary-microkernel-tester.h",
4583 ] + MICROKERNEL_TEST_HDRS,
4584 deps = MICROKERNEL_TEST_DEPS,
4585)
4586
4587xnnpack_unit_test(
4588 name = "f16_vsubc_minmax_test",
4589 srcs = [
4590 "test/f16-vsubc-minmax.cc",
4591 "test/vbinaryc-microkernel-tester.h",
4592 ] + MICROKERNEL_TEST_HDRS,
4593 deps = MICROKERNEL_TEST_DEPS,
4594)
4595
4596xnnpack_unit_test(
4597 name = "f16_vrsubc_minmax_test",
4598 srcs = [
4599 "test/f16-vrsubc-minmax.cc",
4600 "test/vbinaryc-microkernel-tester.h",
4601 ] + MICROKERNEL_TEST_HDRS,
4602 deps = MICROKERNEL_TEST_DEPS,
4603)
4604
4605xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004606 name = "f32_vabs_test",
4607 srcs = [
4608 "test/f32-vabs.cc",
4609 "test/vunary-microkernel-tester.h",
4610 ] + MICROKERNEL_TEST_HDRS,
4611 deps = MICROKERNEL_TEST_DEPS,
4612)
4613
4614xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004615 name = "f32_vadd_test",
4616 srcs = [
4617 "test/f32-vadd.cc",
4618 "test/vbinary-microkernel-tester.h",
4619 ] + MICROKERNEL_TEST_HDRS,
4620 deps = MICROKERNEL_TEST_DEPS,
4621)
4622
4623xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004624 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004625 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004626 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004627 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004628 ] + MICROKERNEL_TEST_HDRS,
4629 deps = MICROKERNEL_TEST_DEPS,
4630)
4631
4632xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004633 name = "f32_vaddc_test",
4634 srcs = [
4635 "test/f32-vaddc.cc",
4636 "test/vbinaryc-microkernel-tester.h",
4637 ] + MICROKERNEL_TEST_HDRS,
4638 deps = MICROKERNEL_TEST_DEPS,
4639)
4640
4641xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004642 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004643 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004644 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004645 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004646 ] + MICROKERNEL_TEST_HDRS,
4647 deps = MICROKERNEL_TEST_DEPS,
4648)
4649
4650xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004651 name = "f32_vdiv_test",
4652 srcs = [
4653 "test/f32-vdiv.cc",
4654 "test/vbinary-microkernel-tester.h",
4655 ] + MICROKERNEL_TEST_HDRS,
4656 deps = MICROKERNEL_TEST_DEPS,
4657)
4658
4659xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004660 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004661 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004662 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004663 "test/vbinary-microkernel-tester.h",
4664 ] + MICROKERNEL_TEST_HDRS,
4665 deps = MICROKERNEL_TEST_DEPS,
4666)
4667
4668xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004669 name = "f32_vdivc_test",
4670 srcs = [
4671 "test/f32-vdivc.cc",
4672 "test/vbinaryc-microkernel-tester.h",
4673 ] + MICROKERNEL_TEST_HDRS,
4674 deps = MICROKERNEL_TEST_DEPS,
4675)
4676
4677xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004678 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004679 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004680 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004681 "test/vbinaryc-microkernel-tester.h",
4682 ] + MICROKERNEL_TEST_HDRS,
4683 deps = MICROKERNEL_TEST_DEPS,
4684)
4685
4686xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004687 name = "f32_vrdivc_test",
4688 srcs = [
4689 "test/f32-vrdivc.cc",
4690 "test/vbinaryc-microkernel-tester.h",
4691 ] + MICROKERNEL_TEST_HDRS,
4692 deps = MICROKERNEL_TEST_DEPS,
4693)
4694
4695xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004696 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004697 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004698 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004699 "test/vbinaryc-microkernel-tester.h",
4700 ] + MICROKERNEL_TEST_HDRS,
4701 deps = MICROKERNEL_TEST_DEPS,
4702)
4703
4704xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004705 name = "f32_vmax_test",
4706 srcs = [
4707 "test/f32-vmax.cc",
4708 "test/vbinary-microkernel-tester.h",
4709 ] + MICROKERNEL_TEST_HDRS,
4710 deps = MICROKERNEL_TEST_DEPS,
4711)
4712
4713xnnpack_unit_test(
4714 name = "f32_vmaxc_test",
4715 srcs = [
4716 "test/f32-vmaxc.cc",
4717 "test/vbinaryc-microkernel-tester.h",
4718 ] + MICROKERNEL_TEST_HDRS,
4719 deps = MICROKERNEL_TEST_DEPS,
4720)
4721
4722xnnpack_unit_test(
4723 name = "f32_vmin_test",
4724 srcs = [
4725 "test/f32-vmin.cc",
4726 "test/vbinary-microkernel-tester.h",
4727 ] + MICROKERNEL_TEST_HDRS,
4728 deps = MICROKERNEL_TEST_DEPS,
4729)
4730
4731xnnpack_unit_test(
4732 name = "f32_vminc_test",
4733 srcs = [
4734 "test/f32-vminc.cc",
4735 "test/vbinaryc-microkernel-tester.h",
4736 ] + MICROKERNEL_TEST_HDRS,
4737 deps = MICROKERNEL_TEST_DEPS,
4738)
4739
4740xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004741 name = "f32_vmul_test",
4742 srcs = [
4743 "test/f32-vmul.cc",
4744 "test/vbinary-microkernel-tester.h",
4745 ] + MICROKERNEL_TEST_HDRS,
4746 deps = MICROKERNEL_TEST_DEPS,
4747)
4748
4749xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004750 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004751 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004752 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004753 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004754 ] + MICROKERNEL_TEST_HDRS,
4755 deps = MICROKERNEL_TEST_DEPS,
4756)
4757
4758xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004759 name = "f32_vmulc_test",
4760 srcs = [
4761 "test/f32-vmulc.cc",
4762 "test/vbinaryc-microkernel-tester.h",
4763 ] + MICROKERNEL_TEST_HDRS,
4764 deps = MICROKERNEL_TEST_DEPS,
4765)
4766
4767xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004768 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004769 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004770 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004771 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004772 ] + MICROKERNEL_TEST_HDRS,
4773 deps = MICROKERNEL_TEST_DEPS,
4774)
4775
4776xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004777 name = "f16_vmulcaddc_minmax_test",
4778 srcs = [
4779 "test/f16-vmulcaddc-minmax.cc",
4780 "test/vmulcaddc-microkernel-tester.h",
4781 "src/xnnpack/AlignedAllocator.h",
4782 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004783 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004784)
4785
4786xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004787 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004788 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004789 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004790 "test/vmulcaddc-microkernel-tester.h",
4791 "src/xnnpack/AlignedAllocator.h",
4792 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004793 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004794)
4795
4796xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004797 name = "f32_vlrelu_test",
4798 srcs = [
4799 "test/f32-vlrelu.cc",
4800 "test/vunary-microkernel-tester.h",
4801 ] + MICROKERNEL_TEST_HDRS,
4802 deps = MICROKERNEL_TEST_DEPS,
4803)
4804
4805xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004806 name = "f32_vneg_test",
4807 srcs = [
4808 "test/f32-vneg.cc",
4809 "test/vunary-microkernel-tester.h",
4810 ] + MICROKERNEL_TEST_HDRS,
4811 deps = MICROKERNEL_TEST_DEPS,
4812)
4813
4814xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004815 name = "f32_vrndne_test",
4816 srcs = [
4817 "test/f32-vrndne.cc",
4818 "test/vunary-microkernel-tester.h",
4819 ] + MICROKERNEL_TEST_HDRS,
4820 deps = MICROKERNEL_TEST_DEPS,
4821)
4822
4823xnnpack_unit_test(
4824 name = "f32_vrndz_test",
4825 srcs = [
4826 "test/f32-vrndz.cc",
4827 "test/vunary-microkernel-tester.h",
4828 ] + MICROKERNEL_TEST_HDRS,
4829 deps = MICROKERNEL_TEST_DEPS,
4830)
4831
4832xnnpack_unit_test(
4833 name = "f32_vrndu_test",
4834 srcs = [
4835 "test/f32-vrndu.cc",
4836 "test/vunary-microkernel-tester.h",
4837 ] + MICROKERNEL_TEST_HDRS,
4838 deps = MICROKERNEL_TEST_DEPS,
4839)
4840
4841xnnpack_unit_test(
4842 name = "f32_vrndd_test",
4843 srcs = [
4844 "test/f32-vrndd.cc",
4845 "test/vunary-microkernel-tester.h",
4846 ] + MICROKERNEL_TEST_HDRS,
4847 deps = MICROKERNEL_TEST_DEPS,
4848)
4849
4850xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004851 name = "f32_vscale_test",
4852 srcs = [
4853 "test/f32-vscale.cc",
4854 "test/vscale-microkernel-tester.h",
4855 ] + MICROKERNEL_TEST_HDRS,
4856 deps = MICROKERNEL_TEST_DEPS,
4857)
4858
4859xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004860 name = "f32_vscaleexpminusmax_test",
4861 srcs = [
4862 "test/f32-vscaleexpminusmax.cc",
4863 "test/vscaleexpminusmax-microkernel-tester.h",
4864 ] + MICROKERNEL_TEST_HDRS,
4865 deps = MICROKERNEL_TEST_DEPS,
4866)
4867
4868xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004869 name = "f32_vscaleextexp_test",
4870 srcs = [
4871 "test/f32-vscaleextexp.cc",
4872 "test/vscaleextexp-microkernel-tester.h",
4873 ] + MICROKERNEL_TEST_HDRS,
4874 deps = MICROKERNEL_TEST_DEPS,
4875)
4876
4877xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004878 name = "f32_vsqr_test",
4879 srcs = [
4880 "test/f32-vsqr.cc",
4881 "test/vunary-microkernel-tester.h",
4882 ] + MICROKERNEL_TEST_HDRS,
4883 deps = MICROKERNEL_TEST_DEPS,
4884)
4885
4886xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004887 name = "f32_vsqrdiff_test",
4888 srcs = [
4889 "test/f32-vsqrdiff.cc",
4890 "test/vbinary-microkernel-tester.h",
4891 ] + MICROKERNEL_TEST_HDRS,
4892 deps = MICROKERNEL_TEST_DEPS,
4893)
4894
4895xnnpack_unit_test(
4896 name = "f32_vsqrdiffc_test",
4897 srcs = [
4898 "test/f32-vsqrdiffc.cc",
4899 "test/vbinaryc-microkernel-tester.h",
4900 ] + MICROKERNEL_TEST_HDRS,
4901 deps = MICROKERNEL_TEST_DEPS,
4902)
4903
4904xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004905 name = "f32_vsqrt_test",
4906 srcs = [
4907 "test/f32-vsqrt.cc",
4908 "test/vunary-microkernel-tester.h",
4909 ] + MICROKERNEL_TEST_HDRS,
4910 deps = MICROKERNEL_TEST_DEPS,
4911)
4912
4913xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004914 name = "f32_vsub_test",
4915 srcs = [
4916 "test/f32-vsub.cc",
4917 "test/vbinary-microkernel-tester.h",
4918 ] + MICROKERNEL_TEST_HDRS,
4919 deps = MICROKERNEL_TEST_DEPS,
4920)
4921
4922xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004923 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07004924 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004925 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004926 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004927 ] + MICROKERNEL_TEST_HDRS,
4928 deps = MICROKERNEL_TEST_DEPS,
4929)
4930
4931xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004932 name = "f32_vsubc_test",
4933 srcs = [
4934 "test/f32-vsubc.cc",
4935 "test/vbinaryc-microkernel-tester.h",
4936 ] + MICROKERNEL_TEST_HDRS,
4937 deps = MICROKERNEL_TEST_DEPS,
4938)
4939
4940xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004941 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004942 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004943 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004944 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004945 ] + MICROKERNEL_TEST_HDRS,
4946 deps = MICROKERNEL_TEST_DEPS,
4947)
4948
4949xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004950 name = "f32_vrsubc_test",
4951 srcs = [
4952 "test/f32-vrsubc.cc",
4953 "test/vbinaryc-microkernel-tester.h",
4954 ] + MICROKERNEL_TEST_HDRS,
4955 deps = MICROKERNEL_TEST_DEPS,
4956)
4957
4958xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004959 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004960 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004961 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004962 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07004963 ] + MICROKERNEL_TEST_HDRS,
4964 deps = MICROKERNEL_TEST_DEPS,
4965)
4966
4967xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004968 name = "q8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004969 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004970 "test/q8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004971 "test/avgpool-microkernel-tester.h",
4972 "src/xnnpack/AlignedAllocator.h",
4973 ] + MICROKERNEL_TEST_HDRS,
4974 deps = MICROKERNEL_TEST_DEPS,
4975)
4976
4977xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004978 name = "q8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004979 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004980 "test/q8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004981 "test/gemm-microkernel-tester.h",
4982 "src/xnnpack/AlignedAllocator.h",
4983 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004984 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004985)
4986
4987xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004988 name = "q8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004989 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004990 "test/q8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004991 "test/dwconv-microkernel-tester.h",
4992 "src/xnnpack/AlignedAllocator.h",
4993 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004994 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004995)
4996
4997xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004998 name = "q8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004999 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005000 "test/q8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005001 "test/gavgpool-microkernel-tester.h",
5002 "src/xnnpack/AlignedAllocator.h",
5003 ] + MICROKERNEL_TEST_HDRS,
5004 deps = MICROKERNEL_TEST_DEPS,
5005)
5006
5007xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005008 name = "q8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005009 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005010 "test/q8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005011 "test/gemm-microkernel-tester.h",
5012 "src/xnnpack/AlignedAllocator.h",
5013 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005014 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005015)
5016
5017xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005018 name = "q8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005019 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005020 "test/q8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005021 "test/vadd-microkernel-tester.h",
5022 ] + MICROKERNEL_TEST_HDRS,
5023 deps = MICROKERNEL_TEST_DEPS,
5024)
5025
5026xnnpack_unit_test(
5027 name = "u8_clamp_test",
5028 srcs = [
5029 "test/u8-clamp.cc",
5030 "test/clamp-microkernel-tester.h",
5031 ] + MICROKERNEL_TEST_HDRS,
5032 deps = MICROKERNEL_TEST_DEPS,
5033)
5034
5035xnnpack_unit_test(
5036 name = "u8_lut32norm_test",
5037 srcs = [
5038 "test/u8-lut32norm.cc",
5039 "test/lut-norm-microkernel-tester.h",
5040 ] + MICROKERNEL_TEST_HDRS,
5041 deps = MICROKERNEL_TEST_DEPS,
5042)
5043
5044xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005045 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005046 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005047 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005048 "test/maxpool-microkernel-tester.h",
5049 ] + MICROKERNEL_TEST_HDRS,
5050 deps = MICROKERNEL_TEST_DEPS,
5051)
5052
5053xnnpack_unit_test(
5054 name = "u8_rmax_test",
5055 srcs = [
5056 "test/u8-rmax.cc",
5057 "test/rmax-microkernel-tester.h",
5058 ] + MICROKERNEL_TEST_HDRS,
5059 deps = MICROKERNEL_TEST_DEPS,
5060)
5061
5062xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005063 name = "x32_fill_test",
5064 srcs = [
5065 "test/x32-fill.cc",
5066 "test/fill-microkernel-tester.h",
5067 ] + MICROKERNEL_TEST_HDRS,
5068 deps = MICROKERNEL_TEST_DEPS,
5069)
5070
5071xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005072 name = "x32_packx_test",
5073 srcs = [
5074 "test/x32-packx.cc",
5075 "test/pack-microkernel-tester.h",
5076 "src/xnnpack/AlignedAllocator.h",
5077 ] + MICROKERNEL_TEST_HDRS,
5078 deps = MICROKERNEL_TEST_DEPS,
5079)
5080
5081xnnpack_unit_test(
5082 name = "x32_pad_test",
5083 srcs = [
5084 "test/x32-pad.cc",
5085 "test/pad-microkernel-tester.h",
5086 ] + MICROKERNEL_TEST_HDRS,
5087 deps = MICROKERNEL_TEST_DEPS,
5088)
5089
5090xnnpack_unit_test(
5091 name = "x32_unpool_test",
5092 srcs = [
5093 "test/x32-unpool.cc",
5094 "test/unpool-microkernel-tester.h",
5095 ] + MICROKERNEL_TEST_HDRS,
5096 deps = MICROKERNEL_TEST_DEPS,
5097)
5098
5099xnnpack_unit_test(
5100 name = "x32_zip_test",
5101 srcs = [
5102 "test/x32-zip.cc",
5103 "test/zip-microkernel-tester.h",
5104 ] + MICROKERNEL_TEST_HDRS,
5105 deps = MICROKERNEL_TEST_DEPS,
5106)
5107
5108xnnpack_unit_test(
5109 name = "x8_lut_test",
5110 srcs = [
5111 "test/x8-lut.cc",
5112 "test/lut-microkernel-tester.h",
5113 ] + MICROKERNEL_TEST_HDRS,
5114 deps = MICROKERNEL_TEST_DEPS,
5115)
5116
5117xnnpack_unit_test(
5118 name = "x8_zip_test",
5119 srcs = [
5120 "test/x8-zip.cc",
5121 "test/zip-microkernel-tester.h",
5122 ] + MICROKERNEL_TEST_HDRS,
5123 deps = MICROKERNEL_TEST_DEPS,
5124)
5125
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005126xnnpack_unit_test(
5127 name = "requantization_test",
5128 srcs = [
5129 "src/xnnpack/requantization-stubs.h",
5130 "test/requantization.cc",
5131 "test/requantization-tester.h",
5132 ] + MICROKERNEL_TEST_HDRS,
5133 deps = MICROKERNEL_TEST_DEPS,
5134)
5135
Marat Dukhan20c3b922020-03-10 03:45:06 -07005136########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005137
5138xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005139 name = "operator_size_test",
5140 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005141 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005142)
5143
Marat Dukhan20c3b922020-03-10 03:45:06 -07005144xnnpack_binary(
5145 name = "subgraph_size_test",
5146 srcs = ["test/subgraph-size.c"],
5147 deps = [":XNNPACK"],
5148)
5149
5150########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005151
5152xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005153 name = "abs_nc_test",
5154 srcs = [
5155 "test/abs-nc.cc",
5156 "test/abs-operator-tester.h",
5157 ],
5158 deps = OPERATOR_TEST_DEPS,
5159)
5160
5161xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005162 name = "add_nd_test",
5163 srcs = [
5164 "test/add-nd.cc",
5165 "test/binary-elementwise-operator-tester.h",
5166 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005167 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005168)
5169
5170xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005171 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005172 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005173 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005174 "test/argmax-pooling-operator-tester.h",
5175 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005176 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005177)
5178
5179xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005180 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005181 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005182 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005183 "test/average-pooling-operator-tester.h",
5184 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005185 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005186)
5187
5188xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005189 name = "bankers_rounding_nc_test",
5190 srcs = [
5191 "test/bankers-rounding-nc.cc",
5192 "test/bankers-rounding-operator-tester.h",
5193 ],
5194 deps = OPERATOR_TEST_DEPS,
5195)
5196
5197xnnpack_unit_test(
5198 name = "ceiling_nc_test",
5199 srcs = [
5200 "test/ceiling-nc.cc",
5201 "test/ceiling-operator-tester.h",
5202 ],
5203 deps = OPERATOR_TEST_DEPS,
5204)
5205
5206xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005207 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005208 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005209 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005210 "test/channel-shuffle-operator-tester.h",
5211 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005212 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005213)
5214
5215xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005216 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005217 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005218 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005219 "test/clamp-operator-tester.h",
5220 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005221 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005222)
5223
5224xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005225 name = "constant_pad_nd_test",
5226 srcs = [
5227 "test/constant-pad-nd.cc",
5228 "test/constant-pad-operator-tester.h",
5229 ],
5230 deps = OPERATOR_TEST_DEPS,
5231)
5232
5233xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005234 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005235 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005236 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005237 "test/convolution-operator-tester.h",
5238 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005239 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005240)
5241
5242xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005243 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005244 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005245 "test/convolution-nchw.cc",
5246 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005247 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005248 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005249)
5250
5251xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005252 name = "copy_nc_test",
5253 srcs = [
5254 "test/copy-nc.cc",
5255 "test/copy-operator-tester.h",
5256 ],
5257 deps = OPERATOR_TEST_DEPS,
5258)
5259
5260xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005261 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005262 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005263 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005264 "test/deconvolution-operator-tester.h",
5265 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005266 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005267)
5268
5269xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005270 name = "divide_nd_test",
5271 srcs = [
5272 "test/binary-elementwise-operator-tester.h",
5273 "test/divide-nd.cc",
5274 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005275 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005276)
5277
5278xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005279 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005280 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005281 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005282 "test/fully-connected-operator-tester.h",
5283 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005284 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005285)
5286
5287xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005288 name = "floor_nc_test",
5289 srcs = [
5290 "test/floor-nc.cc",
5291 "test/floor-operator-tester.h",
5292 ],
5293 deps = OPERATOR_TEST_DEPS,
5294)
5295
5296xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005297 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005298 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005299 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005300 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005301 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005302 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005303)
5304
5305xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005306 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005307 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005308 "test/global-average-pooling-ncw.cc",
5309 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005310 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005311 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005312)
5313
5314xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005315 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005316 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005317 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005318 "test/hardswish-operator-tester.h",
5319 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005320 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005321)
5322
5323xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005324 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005325 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005326 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005327 "test/leaky-relu-operator-tester.h",
5328 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005329 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005330)
5331
5332xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005333 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005334 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005335 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005336 "test/max-pooling-operator-tester.h",
5337 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005338 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005339)
5340
5341xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005342 name = "maximum_nd_test",
5343 srcs = [
5344 "test/binary-elementwise-operator-tester.h",
5345 "test/maximum-nd.cc",
5346 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005347 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005348)
5349
5350xnnpack_unit_test(
5351 name = "minimum_nd_test",
5352 srcs = [
5353 "test/binary-elementwise-operator-tester.h",
5354 "test/minimum-nd.cc",
5355 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005356 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005357)
5358
5359xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005360 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005361 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005362 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005363 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005364 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005365 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005366)
5367
5368xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005369 name = "negate_nc_test",
5370 srcs = [
5371 "test/negate-nc.cc",
5372 "test/negate-operator-tester.h",
5373 ],
5374 deps = OPERATOR_TEST_DEPS,
5375)
5376
5377xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005378 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005379 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005380 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005381 "test/prelu-operator-tester.h",
5382 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005383 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005384)
5385
5386xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005387 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005388 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005389 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005390 "test/resize-bilinear-operator-tester.h",
5391 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005392 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005393)
5394
5395xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005396 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005397 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005398 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005399 "test/sigmoid-operator-tester.h",
5400 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005401 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005402)
5403
5404xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005405 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005406 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005407 "test/softmax-nc.cc",
5408 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005409 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005410 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005411)
5412
5413xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005414 name = "square_nc_test",
5415 srcs = [
5416 "test/square-nc.cc",
5417 "test/square-operator-tester.h",
5418 ],
5419 deps = OPERATOR_TEST_DEPS,
5420)
5421
5422xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005423 name = "square_root_nc_test",
5424 srcs = [
5425 "test/square-root-nc.cc",
5426 "test/square-root-operator-tester.h",
5427 ],
5428 deps = OPERATOR_TEST_DEPS,
5429)
5430
5431xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005432 name = "squared_difference_nd_test",
5433 srcs = [
5434 "test/binary-elementwise-operator-tester.h",
5435 "test/squared-difference-nd.cc",
5436 ],
5437 deps = OPERATOR_TEST_DEPS,
5438)
5439
5440xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005441 name = "subtract_nd_test",
5442 srcs = [
5443 "test/binary-elementwise-operator-tester.h",
5444 "test/subtract-nd.cc",
5445 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005446 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005447)
5448
5449xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005450 name = "truncation_nc_test",
5451 srcs = [
5452 "test/truncation-nc.cc",
5453 "test/truncation-operator-tester.h",
5454 ],
5455 deps = OPERATOR_TEST_DEPS,
5456)
5457
5458xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005459 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005460 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005461 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005462 "test/unpooling-operator-tester.h",
5463 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005464 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005465)
5466
Chao Mei6ddfc602020-05-13 22:29:36 -07005467############################### Misc unit tests ###############################
5468
5469xnnpack_unit_test(
5470 name = "memory_planner_test",
5471 srcs = [
5472 "test/memory-planner-test.cc",
5473 ],
5474 deps = [
5475 ":XNNPACK",
5476 ":memory_planner",
5477 ],
5478)
5479
Marat Dukhan08c4a432019-10-03 09:29:21 -07005480############################# Build configurations #############################
5481
Marat Dukhanb8642352019-10-30 15:43:02 -07005482# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005483config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005484 name = "xnn_enable_assembly_explicit_true",
5485 define_values = {"xnn_enable_assembly": "true"},
5486)
5487
5488# Disables usage of assembly kernels.
5489config_setting(
5490 name = "xnn_enable_assembly_explicit_false",
5491 define_values = {"xnn_enable_assembly": "false"},
5492)
5493
Marat Dukhan9de90e02020-06-18 16:04:12 -07005494# Enables usage of sparse inference.
5495config_setting(
5496 name = "xnn_enable_sparse_explicit_true",
5497 define_values = {"xnn_enable_sparse": "true"},
5498)
5499
5500# Disables usage of sparse inference.
5501config_setting(
5502 name = "xnn_enable_sparse_explicit_false",
5503 define_values = {"xnn_enable_sparse": "false"},
5504)
5505
Marat Dukhan05702cf2020-03-26 15:41:33 -07005506# Disables usage of HMP-aware optimizations.
5507config_setting(
5508 name = "xnn_enable_hmp_explicit_false",
5509 define_values = {"xnn_enable_hmp": "false"},
5510)
5511
Chao Mei6ddfc602020-05-13 22:29:36 -07005512# Enable usage of optimized memory allocation
5513config_setting(
5514 name = "xnn_enable_memopt_explicit_true",
5515 define_values = {"xnnpack_enable_memopt": "true"},
5516)
5517
5518# Disable usage of optimized memory allocation
5519config_setting(
5520 name = "xnn_enable_memopt_explicit_false",
5521 define_values = {"xnnpack_enable_memopt": "false"},
5522)
5523
Marat Dukhanb8642352019-10-30 15:43:02 -07005524# Builds with -c dbg
5525config_setting(
5526 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005527 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005528 "compilation_mode": "dbg",
5529 },
5530)
5531
5532# Builds with -c opt
5533config_setting(
5534 name = "optimized_build",
5535 values = {
5536 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005537 },
5538)
5539
5540config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005541 name = "linux_k8",
5542 values = {"cpu": "k8"},
5543)
5544
5545config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005546 name = "linux_arm",
5547 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005548)
5549
5550config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005551 name = "linux_armeabi",
5552 values = {"cpu": "armeabi"},
5553)
5554
5555config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005556 name = "linux_armhf",
5557 values = {"cpu": "armhf"},
5558)
5559
5560config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005561 name = "linux_armv7a",
5562 values = {"cpu": "armv7a"},
5563)
5564
5565config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005566 name = "linux_aarch64",
5567 values = {"cpu": "aarch64"},
5568)
5569
5570config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005571 name = "android",
5572 values = {"crosstool_top": "//external:android/crosstool"},
5573)
5574
5575config_setting(
5576 name = "android_armv7",
5577 values = {
5578 "crosstool_top": "//external:android/crosstool",
5579 "cpu": "armeabi-v7a",
5580 },
5581)
5582
5583config_setting(
5584 name = "android_arm64",
5585 values = {
5586 "crosstool_top": "//external:android/crosstool",
5587 "cpu": "arm64-v8a",
5588 },
5589)
5590
5591config_setting(
5592 name = "android_x86",
5593 values = {
5594 "crosstool_top": "//external:android/crosstool",
5595 "cpu": "x86",
5596 },
5597)
5598
5599config_setting(
5600 name = "android_x86_64",
5601 values = {
5602 "crosstool_top": "//external:android/crosstool",
5603 "cpu": "x86_64",
5604 },
5605)
5606
5607config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005608 name = "windows_x86_64",
5609 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005610)
5611
5612config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005613 name = "windows_x86_64_clang",
5614 values = {
5615 "compiler": "clang-cl",
5616 "cpu": "x64_windows",
5617 },
5618)
5619
5620config_setting(
5621 name = "windows_x86_64_mingw",
5622 values = {
5623 "compiler": "mingw-gcc",
5624 "cpu": "x64_windows",
5625 },
5626)
5627
5628config_setting(
5629 name = "windows_x86_64_msys",
5630 values = {
5631 "compiler": "msys-gcc",
5632 "cpu": "x64_windows",
5633 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005634)
5635
5636config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005637 name = "macos_x86_64",
5638 values = {
5639 "apple_platform_type": "macos",
5640 "cpu": "darwin",
5641 },
5642)
5643
5644config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005645 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005646 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005647)
5648
5649config_setting(
5650 name = "emscripten_wasm",
5651 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005652 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005653 "cpu": "wasm",
5654 },
5655)
5656
5657config_setting(
5658 name = "emscripten_wasmsimd",
5659 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005660 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005661 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005662 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005663 },
5664)
5665
5666config_setting(
5667 name = "emscripten_asmjs",
5668 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005669 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005670 "cpu": "asmjs",
5671 },
5672)
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005673
5674config_setting(
5675 name = "ios_armv7",
5676 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005677 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005678 "cpu": "ios_armv7",
5679 },
5680)
5681
5682config_setting(
5683 name = "ios_arm64",
5684 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005685 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005686 "cpu": "ios_arm64",
5687 },
5688)
5689
5690config_setting(
5691 name = "ios_arm64e",
5692 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005693 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005694 "cpu": "ios_arm64e",
5695 },
5696)
5697
5698config_setting(
5699 name = "ios_x86",
5700 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005701 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005702 "cpu": "ios_i386",
5703 },
5704)
5705
5706config_setting(
5707 name = "ios_x86_64",
5708 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005709 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005710 "cpu": "ios_x86_64",
5711 },
5712)
5713
5714config_setting(
5715 name = "watchos_armv7k",
5716 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005717 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005718 "cpu": "watchos_armv7k",
5719 },
5720)
5721
5722config_setting(
5723 name = "watchos_arm64_32",
5724 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005725 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005726 "cpu": "watchos_arm64_32",
5727 },
5728)
5729
5730config_setting(
5731 name = "watchos_x86",
5732 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005733 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005734 "cpu": "watchos_i386",
5735 },
5736)
5737
5738config_setting(
5739 name = "watchos_x86_64",
5740 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005741 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005742 "cpu": "watchos_x86_64",
5743 },
5744)
5745
5746config_setting(
5747 name = "tvos_arm64",
5748 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005749 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005750 "cpu": "tvos_arm64",
5751 },
5752)
5753
5754config_setting(
5755 name = "tvos_x86_64",
5756 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005757 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005758 "cpu": "tvos_x86_64",
5759 },
5760)