blob: 8ffbba8dd3805ad1a3bac6ffd140264093d54114 [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 Dukhand27202d2020-07-09 23:43:40 -0700101 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700102 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700103 "src/subgraph/square.c",
Marat Dukhan51a01c62020-07-09 03:26:57 -0700104 "src/subgraph/square-root.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700105 "src/subgraph/squared-difference.c",
106 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700107 "src/subgraph/unpooling-2d.c",
108]
109
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800110TABLE_SRCS = [
111 "src/tables/exp2-k-over-64.c",
112 "src/tables/exp2-k-over-2048.c",
113]
114
Marat Dukhan08c4a432019-10-03 09:29:21 -0700115SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800116 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800117 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700119 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
120 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700121 "src/f32-clamp/gen/scalar-x1.c",
122 "src/f32-clamp/gen/scalar-x2.c",
123 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan441e2212019-12-04 18:30:49 -0800124 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800125 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700126 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
127 "src/f32-dwconv-chw/3x3p1-scalar.c",
128 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
129 "src/f32-dwconv-chw/5x5p2-scalar.c",
130 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700131 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
132 "src/f32-dwconv/gen/up1x4-scalar.c",
133 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
134 "src/f32-dwconv/gen/up1x9-scalar.c",
135 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
136 "src/f32-dwconv/gen/up1x25-scalar.c",
137 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
138 "src/f32-dwconv/gen/up2x4-scalar.c",
139 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
140 "src/f32-dwconv/gen/up2x9-scalar.c",
141 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
142 "src/f32-dwconv/gen/up2x25-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700143 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
144 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
145 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
146 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700147 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
148 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
150 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
151 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
152 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700153 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
154 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700155 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700156 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
157 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700158 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
159 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
160 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700161 "src/f32-gemm/gen/1x4-scalar.c",
162 "src/f32-gemm/gen/2x4-scalar.c",
163 "src/f32-gemm/gen/4x2-scalar.c",
164 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700165 "src/f32-gemm/gen/1x4-relu-scalar.c",
166 "src/f32-gemm/gen/2x4-relu-scalar.c",
167 "src/f32-gemm/gen/4x2-relu-scalar.c",
168 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700169 "src/f32-gemm/gen/1x4-minmax-scalar.c",
170 "src/f32-gemm/gen/2x4-minmax-scalar.c",
171 "src/f32-gemm/gen/4x2-minmax-scalar.c",
172 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700173 "src/f32-hswish/gen/hswish-scalar-x1.c",
174 "src/f32-hswish/gen/hswish-scalar-x2.c",
175 "src/f32-hswish/gen/hswish-scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700176 "src/f32-ibilinear/gen/scalar-c1.c",
177 "src/f32-ibilinear/gen/scalar-c2.c",
178 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700179 "src/f32-igemm/gen/1x4-scalar.c",
180 "src/f32-igemm/gen/2x4-scalar.c",
181 "src/f32-igemm/gen/4x2-scalar.c",
182 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700183 "src/f32-igemm/gen/1x4-relu-scalar.c",
184 "src/f32-igemm/gen/2x4-relu-scalar.c",
185 "src/f32-igemm/gen/4x2-relu-scalar.c",
186 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700187 "src/f32-igemm/gen/1x4-minmax-scalar.c",
188 "src/f32-igemm/gen/2x4-minmax-scalar.c",
189 "src/f32-igemm/gen/4x2-minmax-scalar.c",
190 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700191 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
192 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
193 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700194 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
195 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
196 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
197 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800198 "src/f32-prelu/gen/scalar-2x1.c",
199 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800200 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
201 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
202 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
203 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
211 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700212 "src/f32-relu/gen/scalar-x1.c",
213 "src/f32-relu/gen/scalar-x2.c",
214 "src/f32-relu/gen/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700215 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800216 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
217 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
218 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
219 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
220 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
221 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
222 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
223 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
224 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700225 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
226 "src/f32-spmm/gen/1x1-minmax-scalar.c",
227 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
228 "src/f32-spmm/gen/2x1-minmax-scalar.c",
229 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
230 "src/f32-spmm/gen/4x1-minmax-scalar.c",
231 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
232 "src/f32-spmm/gen/8x1-minmax-scalar.c",
233 "src/f32-spmm/gen/8x2-minmax-scalar.c",
234 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700235 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
236 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
237 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700238 "src/f32-vbinary/gen/vadd-scalar-x1.c",
239 "src/f32-vbinary/gen/vadd-scalar-x2.c",
240 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700241 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
242 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
243 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700244 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
245 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
246 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700247 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
248 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
249 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700250 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
251 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
252 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700253 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
254 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
255 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700256 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
257 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
258 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800259 "src/f32-vbinary/gen/vmax-scalar-x1.c",
260 "src/f32-vbinary/gen/vmax-scalar-x2.c",
261 "src/f32-vbinary/gen/vmax-scalar-x4.c",
262 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
263 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
264 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
265 "src/f32-vbinary/gen/vmin-scalar-x1.c",
266 "src/f32-vbinary/gen/vmin-scalar-x2.c",
267 "src/f32-vbinary/gen/vmin-scalar-x4.c",
268 "src/f32-vbinary/gen/vminc-scalar-x1.c",
269 "src/f32-vbinary/gen/vminc-scalar-x2.c",
270 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700271 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
272 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
273 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700274 "src/f32-vbinary/gen/vmul-scalar-x1.c",
275 "src/f32-vbinary/gen/vmul-scalar-x2.c",
276 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700277 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
278 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
279 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700280 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
281 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
282 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700283 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
284 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
285 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700286 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
287 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
288 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700289 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
290 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
291 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700292 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
293 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
294 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700295 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
296 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
297 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
298 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
299 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
300 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700301 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
302 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
303 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700304 "src/f32-vbinary/gen/vsub-scalar-x1.c",
305 "src/f32-vbinary/gen/vsub-scalar-x2.c",
306 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700307 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
308 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
309 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700310 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
311 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
312 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700313 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
314 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
315 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700316 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
317 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
318 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700319 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
320 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
321 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
322 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
323 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
324 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
325 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
326 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
327 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
328 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
329 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
330 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700331 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
332 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
333 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700334 "src/f32-vunary/gen/vabs-scalar-x1.c",
335 "src/f32-vunary/gen/vabs-scalar-x2.c",
336 "src/f32-vunary/gen/vabs-scalar-x4.c",
337 "src/f32-vunary/gen/vneg-scalar-x1.c",
338 "src/f32-vunary/gen/vneg-scalar-x2.c",
339 "src/f32-vunary/gen/vneg-scalar-x4.c",
340 "src/f32-vunary/gen/vsqr-scalar-x1.c",
341 "src/f32-vunary/gen/vsqr-scalar-x2.c",
342 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800343 "src/math/expminus-scalar-lut2048-p1.c",
344 "src/math/expminus-scalar-lut64-p2.c",
345 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700346 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700347 "src/math/roundne-scalar-nearbyint.c",
348 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700349 "src/math/roundd-scalar-addsub.c",
350 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700351 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700352 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700353 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700354 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700355 "src/math/roundz-scalar-addsub.c",
356 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700357 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800358 "src/math/sigmoid-scalar-lut2048-p1-div.c",
359 "src/math/sigmoid-scalar-lut64-p2-div.c",
360 "src/math/sigmoid-scalar-p5-div.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700361 "src/qs8-requantization/fp32-scalar-lrintf.c",
362 "src/qs8-requantization/fp32-scalar-magic.c",
363 "src/qs8-requantization/precise-scalar-signed64.c",
364 "src/qs8-requantization/precise-scalar-unsigned32.c",
365 "src/qs8-requantization/precise-scalar-unsigned64.c",
366 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700367 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
368 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
369 "src/qu8-dwconv/up1x9-minmax-scalar.c",
370 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
371 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
372 "src/qu8-gemm/2x2-minmax-scalar.c",
373 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700374 "src/qu8-requantization/fp32-scalar-lrintf.c",
375 "src/qu8-requantization/fp32-scalar-magic.c",
376 "src/qu8-requantization/precise-scalar-signed64.c",
377 "src/qu8-requantization/precise-scalar-unsigned32.c",
378 "src/qu8-requantization/precise-scalar-unsigned64.c",
379 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700380 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700381 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700382 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700383 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700384 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700385 "src/x32-fill/scalar-float.c",
386 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700387 "src/x32-packx/x2-scalar.c",
388 "src/x32-packx/x3-scalar.c",
389 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700390 "src/x32-pad/scalar-int.c",
391 "src/x32-pad/scalar-float.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700392 "src/x32-unpool/scalar.c",
393 "src/x32-zip/x2-scalar.c",
394 "src/x32-zip/x3-scalar.c",
395 "src/x32-zip/x4-scalar.c",
396 "src/x32-zip/xm-scalar.c",
397 "src/x8-lut/scalar.c",
398 "src/x8-zip/x2-scalar.c",
399 "src/x8-zip/x3-scalar.c",
400 "src/x8-zip/x4-scalar.c",
401 "src/x8-zip/xm-scalar.c",
402]
403
Marat Dukhan436ebe62019-12-04 15:10:12 -0800404WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700405 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
406 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700407 "src/f32-clamp/gen/wasm-x1.c",
408 "src/f32-clamp/gen/wasm-x2.c",
409 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700410 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
411 "src/f32-dwconv/gen/up1x4-wasm.c",
412 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
413 "src/f32-dwconv/gen/up1x9-wasm.c",
414 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
415 "src/f32-dwconv/gen/up1x25-wasm.c",
416 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
417 "src/f32-dwconv/gen/up2x4-wasm.c",
418 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
419 "src/f32-dwconv/gen/up2x9-wasm.c",
420 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
421 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700422 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
423 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
424 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
425 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700426 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
427 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700428 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
429 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
430 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
431 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700432 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
433 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700434 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
435 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700436 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
437 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
438 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700439 "src/f32-gemm/gen/1x4-wasm.c",
440 "src/f32-gemm/gen/2x4-wasm.c",
441 "src/f32-gemm/gen/4x2-wasm.c",
442 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700443 "src/f32-gemm/gen/1x4-relu-wasm.c",
444 "src/f32-gemm/gen/2x4-relu-wasm.c",
445 "src/f32-gemm/gen/4x2-relu-wasm.c",
446 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700447 "src/f32-gemm/gen/1x4-minmax-wasm.c",
448 "src/f32-gemm/gen/2x4-minmax-wasm.c",
449 "src/f32-gemm/gen/4x2-minmax-wasm.c",
450 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700451 "src/f32-hswish/gen/hswish-wasm-x1.c",
452 "src/f32-hswish/gen/hswish-wasm-x2.c",
453 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700454 "src/f32-igemm/gen/1x4-wasm.c",
455 "src/f32-igemm/gen/2x4-wasm.c",
456 "src/f32-igemm/gen/4x2-wasm.c",
457 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700458 "src/f32-igemm/gen/1x4-relu-wasm.c",
459 "src/f32-igemm/gen/2x4-relu-wasm.c",
460 "src/f32-igemm/gen/4x2-relu-wasm.c",
461 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700462 "src/f32-igemm/gen/1x4-minmax-wasm.c",
463 "src/f32-igemm/gen/2x4-minmax-wasm.c",
464 "src/f32-igemm/gen/4x2-minmax-wasm.c",
465 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700466 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
467 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
468 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700469 "src/f32-prelu/gen/wasm-2x1.c",
470 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700471 "src/f32-relu/gen/wasm-x1.c",
472 "src/f32-relu/gen/wasm-x2.c",
473 "src/f32-relu/gen/wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700474 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
475 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
476 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
477 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
478 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
479 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
480 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
481 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
482 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
483 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
484 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
485 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800486 "src/f32-vbinary/gen/vmax-wasm-x1.c",
487 "src/f32-vbinary/gen/vmax-wasm-x2.c",
488 "src/f32-vbinary/gen/vmax-wasm-x4.c",
489 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
490 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
491 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
492 "src/f32-vbinary/gen/vmin-wasm-x1.c",
493 "src/f32-vbinary/gen/vmin-wasm-x2.c",
494 "src/f32-vbinary/gen/vmin-wasm-x4.c",
495 "src/f32-vbinary/gen/vminc-wasm-x1.c",
496 "src/f32-vbinary/gen/vminc-wasm-x2.c",
497 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700498 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
499 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
500 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
501 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
502 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
503 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
504 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
505 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
506 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
507 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
508 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
509 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
510 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
511 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
512 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
513 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
514 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
515 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700516 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
517 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
518 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700519 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
520 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
521 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800522]
523
Marat Dukhan290055c2020-06-09 12:24:29 -0700524WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700525 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
526 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
527 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700528 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
529 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
530 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
531 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700532 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
533 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
534 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
535 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700536 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
537 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
538 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
539 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
540 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
541 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700542 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
543 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
544 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
545 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
546 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
547 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
548 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
549 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
550 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
551 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
552 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
553 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
554 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
555 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
556 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
557 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
558 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
559 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
560 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
561 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
562 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
563 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
564 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
565 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700566 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
567 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
568 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700569 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700570 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
571 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
572 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
573 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700574 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
575 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
576 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
577 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
578 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
579 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
580 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
581 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
582 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
583 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
584 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
585 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
586 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
587 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
588 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
589 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
590 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
591 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
592 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
593 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
594 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
595 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
596 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
597 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
598 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
599 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
600 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
601 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
602 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
603 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700604 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
605 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700606 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
607 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
608 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
609 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
610 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
611 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
612 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
613 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
614 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
615 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
616 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
617 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
618 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
619 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
620 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
621 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
622 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
623 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
624 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
625 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
626 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
627 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
628 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
629 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
630 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
631 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
632 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
633 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
634 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
635 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700636 "src/f32-ibilinear/gen/wasmsimd-c4.c",
637 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700638 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
639 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
640 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700641 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700642 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
643 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
644 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
645 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700646 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
647 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
648 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
649 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
650 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
651 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
652 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
653 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
654 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
655 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
656 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
657 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
658 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
659 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
660 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
661 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
662 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
663 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
664 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
665 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
666 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
667 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
668 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
669 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
670 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
671 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
672 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
673 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
674 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
675 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700676 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
677 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700678 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
679 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
680 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
681 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700682 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
683 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700684 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
685 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
686 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700687 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
688 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700689 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
690 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
691 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
692 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700693 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
694 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700695 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
696 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
697 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700698 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
699 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700700 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
701 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
702 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
703 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
704 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
705 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
706 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700707 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
708 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700709 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
710 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
711 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
712 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700713 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
714 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
715 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
716 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
717 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
718 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
719 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
720 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
721 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
722 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
723 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
724 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700725 "src/f32-relu/gen/wasmsimd-x4.c",
726 "src/f32-relu/gen/wasmsimd-x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700727 "src/f32-rmax/wasmsimd-arm.c",
728 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700729 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
730 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
731 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
732 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
733 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
734 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700735 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
736 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
737 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
738 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
739 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
740 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700741 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
742 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
743 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
744 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700745 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
746 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700747 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
748 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
749 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
750 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700751 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
752 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700753 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
754 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
755 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
756 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700757 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
758 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700759 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
760 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
761 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
762 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700763 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
764 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700765 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
766 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
767 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
768 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700769 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
770 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
771 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
772 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700773 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
774 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
775 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
776 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700777 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
778 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
779 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
780 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700781 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
782 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
783 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
784 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700785 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
786 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700787 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
788 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
789 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
790 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700791 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
792 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700793 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
794 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
795 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
796 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700797 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
798 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700799 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
800 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
801 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
802 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700803 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
804 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700805 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
806 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700807 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
808 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700809 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
810 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
811 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
812 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700813 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
814 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700815 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
816 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
817 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
818 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700819 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
820 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700821 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
822 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
823 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
824 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700825 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
826 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
827 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
828 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700829 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
830 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
831 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
832 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
833 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
834 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
835 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
836 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
837 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
838 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
839 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
840 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
841 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
842 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700843 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
844 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700845 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
846 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
847 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
848 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
849 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
850 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700851 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700852 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700853 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700854 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700855 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700856 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700857 "src/x32-zip/x2-wasmsimd.c",
858 "src/x32-zip/x3-wasmsimd.c",
859 "src/x32-zip/x4-wasmsimd.c",
860 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700861 "src/math/roundne-wasmsimd-addsub.c",
862 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700863 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700864 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700865 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700866 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700867 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -0700868 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
869 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700870]
871
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800872PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -0700873 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700874 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700875 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700876 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700877 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700878 "src/f32-spmm/gen/4x1-minmax-psimd.c",
879 "src/f32-spmm/gen/8x1-minmax-psimd.c",
880 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700881]
882
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800883PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700884 "src/qs8-requantization/precise-psimd.c",
885 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700886 "src/qu8-requantization/precise-psimd.c",
887 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800888]
889
Marat Dukhan08c4a432019-10-03 09:29:21 -0700890# ISA-specific micro-kernels
891NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -0700892 "src/f32-argmaxpool/4x-neon-c4.c",
893 "src/f32-argmaxpool/9p8x-neon-c4.c",
894 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700895 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
896 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700897 "src/f32-clamp/gen/neon-x4.c",
898 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700899 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
900 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700901 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
902 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700903 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
904 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700905 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
906 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700907 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
908 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
909 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
910 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700911 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
912 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
913 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
914 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700915 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
916 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
917 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
918 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700919 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700920 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
921 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700922 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
923 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
924 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
925 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
926 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
927 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
928 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
929 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
930 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
931 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
932 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
933 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
934 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
935 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
936 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
937 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
938 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
939 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
940 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
941 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
942 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
943 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
944 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
945 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
946 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
947 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
948 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
949 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
950 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
951 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
952 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700953 "src/f32-hswish/gen/hswish-neon-x4.c",
954 "src/f32-hswish/gen/hswish-neon-x8.c",
955 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700956 "src/f32-ibilinear/gen/neon-c4.c",
957 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700958 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
959 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
960 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
961 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
962 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
963 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
964 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
965 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
966 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
967 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
968 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
969 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
970 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
971 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
972 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
973 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700974 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
975 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
976 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700977 "src/f32-ppmm/gen/4x8-minmax-neon.c",
978 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700979 "src/f32-prelu/gen/neon-1x4.c",
980 "src/f32-prelu/gen/neon-1x8.c",
981 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800982 "src/f32-prelu/gen/neon-2x4.c",
983 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700984 "src/f32-prelu/gen/neon-2x16.c",
985 "src/f32-prelu/gen/neon-4x4.c",
986 "src/f32-prelu/gen/neon-4x8.c",
987 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800988 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
989 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
990 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
991 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
992 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
993 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
994 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
995 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
996 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
997 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
998 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
999 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1000 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1001 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1002 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1003 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1004 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1005 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1006 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1007 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1008 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1009 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1010 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1011 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001012 "src/f32-relu/gen/neon-x4.c",
1013 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001014 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001015 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001016 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1017 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1018 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1019 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1020 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1021 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1022 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1023 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1024 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1025 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1026 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1027 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1028 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1029 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1030 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1031 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1032 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1033 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001034 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1035 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1036 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1037 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001038 "src/f32-vbinary/gen/vmax-neon-x4.c",
1039 "src/f32-vbinary/gen/vmax-neon-x8.c",
1040 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1041 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1042 "src/f32-vbinary/gen/vmin-neon-x4.c",
1043 "src/f32-vbinary/gen/vmin-neon-x8.c",
1044 "src/f32-vbinary/gen/vminc-neon-x4.c",
1045 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001046 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1047 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1048 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1049 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1050 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1051 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001052 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1053 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1054 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1055 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001056 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1057 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1058 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1059 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001060 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1061 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001062 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1063 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001064 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1065 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1066 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1067 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1068 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1069 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1070 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1071 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001072 "src/f32-vunary/gen/vabs-neon-x4.c",
1073 "src/f32-vunary/gen/vabs-neon-x8.c",
1074 "src/f32-vunary/gen/vneg-neon-x4.c",
1075 "src/f32-vunary/gen/vneg-neon-x8.c",
1076 "src/f32-vunary/gen/vsqr-neon-x4.c",
1077 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001078 "src/qs8-requantization/precise-neon.c",
1079 "src/qs8-requantization/fp32-neon.c",
1080 "src/qs8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001081 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1082 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1083 "src/qu8-dwconv/up8x9-minmax-neon.c",
1084 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1085 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1086 "src/qu8-gemm/4x8-minmax-neon.c",
1087 "src/qu8-gemm/8x8-minmax-neon.c",
1088 "src/qu8-igemm/4x8-minmax-neon.c",
1089 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001090 "src/qu8-requantization/precise-neon.c",
1091 "src/qu8-requantization/fp32-neon.c",
1092 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001093 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001094 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001095 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001096 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001097 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001098 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001099 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001100 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001101 "src/x32-zip/x2-neon.c",
1102 "src/x32-zip/x3-neon.c",
1103 "src/x32-zip/x4-neon.c",
1104 "src/x32-zip/xm-neon.c",
1105 "src/x8-zip/x2-neon.c",
1106 "src/x8-zip/x3-neon.c",
1107 "src/x8-zip/x4-neon.c",
1108 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001109 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001110 "src/math/roundd-neon-addsub.c",
1111 "src/math/roundd-neon-cvt.c",
1112 "src/math/roundu-neon-addsub.c",
1113 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001114 "src/math/roundz-neon-addsub.c",
1115 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001116 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001117 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1118 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1119 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1120 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1121 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1122 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001123 "src/math/sqrt-neon-nr1rsqrts.c",
1124 "src/math/sqrt-neon-nr2rsqrts.c",
1125 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001126]
1127
1128NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001129 "src/f32-ibilinear/gen/neonfma-c4.c",
1130 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001131 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1132 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1133 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1134 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1135 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1136 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1137 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1138 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1139 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001140 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1141 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1142 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1143 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001144 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1145 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1146 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1147 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001148 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1149 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1150 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1151 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001152 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1153 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1154 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1155 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1156 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1157 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1158 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1159 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1160 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1161 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1162 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1163 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1164 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1165 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1166 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1167 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1168 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1169 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001170 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1171 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001172 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1173 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1174 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1175 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1176 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1177 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1178 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1179 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1180 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1181 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1182 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1183 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1184 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1185 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1186 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1187 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1188 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1189 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1190 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1191 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1192 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1193 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1194 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1195 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001196 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1197 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1198 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1199 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1200 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1201 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1202 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1203 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1204 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1205 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1206 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1207 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1208 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1209 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1210 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1211 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1212 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1213 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1214 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1215 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1216 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1217 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1218 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1219 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1220 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1221 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1222 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1223 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1224 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1225 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1226 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1227 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1228 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1229 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1230 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1231 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1232 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1233 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1234 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1235 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1236 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1237 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1238 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1239 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1240 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1241 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1242 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1243 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1244 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1245 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1246 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1247 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1248 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1249 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001250 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1251 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001252 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1253 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1254 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1255 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1256 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1257 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1258 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1259 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1260 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1261 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1262 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1263 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1264 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1265 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1266 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1267 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1268 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1269 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1270 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1271 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001272 "src/math/exp-neonfma-lut64-p2.c",
1273 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001274 "src/math/expminus-neonfma-lut2048-p1.c",
1275 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001276 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001277 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1278 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1279 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1280 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1281 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1282 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1283 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1284 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1285 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1286 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1287 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1288 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1289 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1290 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1291 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1292 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1293 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1294 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001295 "src/math/sqrt-neonfma-nr1fma.c",
1296 "src/math/sqrt-neonfma-nr2fma.c",
1297 "src/math/sqrt-neonfma-nr3fma.c",
1298 "src/math/sqrt-neonfma-nr2fma1adj.c",
1299 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001300]
1301
1302AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001303 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1304 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1305 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1306 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1307 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1308 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001309 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1310 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1311 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1312 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1313 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1314 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1315 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1316 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1317 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1318 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1319 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1320 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1321 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1322 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1323 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1324 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1325 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1326 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1327 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1328 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001329 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1330 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001331 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1332 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001333 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1334 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001335 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1336 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001337 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1338 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1339 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1340 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1341 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001342 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1343 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1344 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1345 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1346 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1347 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1348 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1349 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1350 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1351 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1352 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1353 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1354 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1355 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1356 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1357 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1358 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1359 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001360 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1361 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1362 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1363 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1364 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1365 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1366 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1367 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1368 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1369 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1370 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1371 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1372 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1373 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1374 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1375 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1376 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1377 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001378 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1379 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001380 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1381 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1382 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1383 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1384 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1385 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001386]
1387
Marat Dukhan8853b822020-05-07 12:19:01 -07001388NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001389 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1390 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1391 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1392 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1393 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1394 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1395 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1396 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001397 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001398 "src/math/roundd-neonv8.c",
1399 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001400 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001401]
1402
Marat Dukhan08c4a432019-10-03 09:29:21 -07001403AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001404 "src/f16-clamp/gen/neonfp16arith-x16.c",
1405 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001406 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1407 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1408 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1409 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1410 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1411 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1412 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1413 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1414 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1415 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1416 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1417 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001418 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1419 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001420 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1421 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001422 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001423 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001424 "src/f16-relu/gen/neonfp16arith-x16.c",
1425 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001426 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1427 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1428 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1429 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1430 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1431 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1432 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1433 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1434 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1435 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1436 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1437 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1438 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1439 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1440 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1441 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1442 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1443 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1444 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1445 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1446 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1447 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1448 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1449 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1450 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1451 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1452 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1453 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001454 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1455 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001456 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1457 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1458 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1459 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1460 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1461 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1462 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1463 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001464 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1465 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1466 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1467 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001468 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1469 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1470 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1471 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1472 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1473 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1474 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1475 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1476 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1477 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1478 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1479 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001480 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1481 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1482 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1483 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1484 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1485 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1486 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1487 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001488]
1489
1490SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001491 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1492 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001493 "src/f32-clamp/gen/sse-x4.c",
1494 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001495 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1496 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001497 "src/f32-dwconv-chw/3x3p1-sse.c",
1498 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001499 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1500 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1501 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1502 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1503 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1504 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1505 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1506 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1507 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1508 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1509 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1510 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001511 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001512 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1513 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001514 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1515 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1516 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1517 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1518 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1519 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1520 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1521 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1522 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1523 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1524 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1525 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1526 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001527 "src/f32-hswish/gen/hswish-sse-x4.c",
1528 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001529 "src/f32-ibilinear/gen/sse-c4.c",
1530 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001531 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1532 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1533 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1534 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1535 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1536 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1537 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001538 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1539 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1540 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001541 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001542 "src/f32-prelu/gen/sse-2x4.c",
1543 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001544 "src/f32-relu/gen/sse-x4.c",
1545 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001546 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001547 "src/f32-spmm/gen/4x1-minmax-sse.c",
1548 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001549 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001550 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1551 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1552 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1553 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1554 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1555 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1556 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1557 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001558 "src/f32-vbinary/gen/vmax-sse-x4.c",
1559 "src/f32-vbinary/gen/vmax-sse-x8.c",
1560 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1561 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1562 "src/f32-vbinary/gen/vmin-sse-x4.c",
1563 "src/f32-vbinary/gen/vmin-sse-x8.c",
1564 "src/f32-vbinary/gen/vminc-sse-x4.c",
1565 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001566 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1567 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1568 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1569 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1570 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1571 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1572 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1573 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001574 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1575 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1576 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1577 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001578 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1579 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1580 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1581 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001582 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1583 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001584 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1585 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001586 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1587 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001588 "src/f32-vunary/gen/vabs-sse-x4.c",
1589 "src/f32-vunary/gen/vabs-sse-x8.c",
1590 "src/f32-vunary/gen/vneg-sse-x4.c",
1591 "src/f32-vunary/gen/vneg-sse-x8.c",
1592 "src/f32-vunary/gen/vsqr-sse-x4.c",
1593 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001594 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001595 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001596 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001597 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001598 "src/math/roundd-sse-addsub.c",
1599 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001600 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001601 "src/math/sqrt-sse-hh1mac.c",
1602 "src/math/sqrt-sse-nr1mac.c",
1603 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001604]
1605
1606SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001607 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1608 "src/f32-argmaxpool/4x-sse2-c4.c",
1609 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001610 "src/f32-prelu/gen/sse2-2x4.c",
1611 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001612 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1613 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1614 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1615 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1616 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1617 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1618 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1619 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1620 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1621 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1622 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1623 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001624 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001625 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001626 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001627 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001628 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1629 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001630 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1631 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001632 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1633 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1634 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1635 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1636 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1637 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1638 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1639 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001640 "src/qs8-requantization/precise-sse2.c",
1641 "src/qs8-requantization/fp32-sse2.c",
1642 "src/qs8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001643 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1644 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1645 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1646 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1647 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1648 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1649 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1650 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001651 "src/qu8-requantization/precise-sse2.c",
1652 "src/qu8-requantization/fp32-sse2.c",
1653 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001654 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001655 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001656 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001657 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001658 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001659 "src/x32-zip/x2-sse2.c",
1660 "src/x32-zip/x3-sse2.c",
1661 "src/x32-zip/x4-sse2.c",
1662 "src/x32-zip/xm-sse2.c",
1663 "src/x8-zip/x2-sse2.c",
1664 "src/x8-zip/x3-sse2.c",
1665 "src/x8-zip/x4-sse2.c",
1666 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001667 "src/math/exp-sse2-p5.c",
1668 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001669 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001670 "src/math/roundd-sse2-cvt.c",
1671 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001672 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001673 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001674]
1675
1676SSSE3_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001677 "src/qs8-requantization/precise-ssse3.c",
1678 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001679 "src/qu8-requantization/precise-ssse3.c",
1680 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001681]
1682
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001683SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001684 "src/f32-prelu/gen/sse41-2x4.c",
1685 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001686 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001687 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001688 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001689 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001690 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1691 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001692 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1693 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001694 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1695 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1696 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1697 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1698 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1699 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1700 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1701 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001702 "src/qs8-requantization/fp32-sse4.c",
1703 "src/qs8-requantization/precise-sse4.c",
1704 "src/qs8-requantization/q31-sse4.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001705 "src/qu8-requantization/precise-sse4.c",
1706 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001707 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001708 "src/math/roundd-sse41.c",
1709 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001710 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001711]
1712
Marat Dukhan08c4a432019-10-03 09:29:21 -07001713AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001714 "src/f32-clamp/gen/avx-x8.c",
1715 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001716 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1717 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1718 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1719 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1720 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1721 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1722 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1723 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1724 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1725 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1726 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1727 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1728 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1729 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1730 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1731 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1732 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1733 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1734 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1735 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1736 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1737 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1738 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1739 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1740 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1741 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1742 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1743 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1744 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1745 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001746 "src/f32-hswish/gen/hswish-avx-x8.c",
1747 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001748 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1749 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1750 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1751 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1752 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1753 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1754 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1755 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1756 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001757 "src/f32-prelu/gen/avx-2x8.c",
1758 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001759 "src/f32-relu/gen/avx-x16.c",
1760 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001761 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001762 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1763 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1764 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1765 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1766 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1767 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1768 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1769 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001770 "src/f32-vbinary/gen/vmax-avx-x8.c",
1771 "src/f32-vbinary/gen/vmax-avx-x16.c",
1772 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1773 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1774 "src/f32-vbinary/gen/vmin-avx-x8.c",
1775 "src/f32-vbinary/gen/vmin-avx-x16.c",
1776 "src/f32-vbinary/gen/vminc-avx-x8.c",
1777 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001778 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1779 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1780 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1781 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1782 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1783 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1784 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1785 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001786 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1787 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1788 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1789 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001790 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1791 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1792 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1793 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001794 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1795 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001796 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001797 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1798 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1799 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1800 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1801 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1802 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1803 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1804 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001805 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1806 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001807 "src/f32-vunary/gen/vabs-avx-x8.c",
1808 "src/f32-vunary/gen/vabs-avx-x16.c",
1809 "src/f32-vunary/gen/vneg-avx-x8.c",
1810 "src/f32-vunary/gen/vneg-avx-x16.c",
1811 "src/f32-vunary/gen/vsqr-avx-x8.c",
1812 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001813]
1814
Marat Dukhanfda12b82019-11-21 12:27:59 -08001815FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001816 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1817 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1818 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1819 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1820 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1821 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1822 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1823 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1824 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1825 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1826 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1827 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1828 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1829 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1830 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1831 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1832 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1833 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1834 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1835 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1836 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1837 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1838 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1839 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1840 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1841 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1842 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1843 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1844 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1845 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1846 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1847 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1848 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1849 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1850 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1851 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1852 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1853 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1854 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1855 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001856 "src/f32-hswish/gen/hswish-fma3-x8.c",
1857 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001858 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1859 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1860 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1861 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1862 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1863 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1864 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1865 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1866 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1867 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1868 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1869 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1870 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1871 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001872 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1873 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1874 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1875 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1876 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1877 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1878 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1879 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001880 "src/math/sqrt-fma3-nr1fma.c",
1881 "src/math/sqrt-fma3-nr2fma.c",
1882 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001883]
1884
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001885AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001886 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1887 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1888 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1889 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1890 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1891 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1892 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1893 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1894 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1895 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1896 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1897 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1898 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1899 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1900 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1901 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1902 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1903 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1904 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1905 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1906 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1907 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1908 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1909 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1910 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1911 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1912 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1913 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1914 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1915 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1916 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1917 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1918 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1919 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1920 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1921 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001922 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1923 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1924 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1925 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1926 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1927 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1928 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1929 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1930 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1931 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1932 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1933 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1934 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1935 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1936 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1937 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1938 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1939 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1940 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1941 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1942 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1943 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1944 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1945 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1946 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1947 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1948 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1949 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1950 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1951 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001952 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1953 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1954 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1955 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1956 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1957 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1958 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1959 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1960 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1961 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1962 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1963 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1964 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1965 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1966 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1967 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1968 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1969 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1970 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1971 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1972 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1973 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1974 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1975 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001976 "src/math/exp-avx2-p5.c",
1977 "src/math/exp-avx2-perm-p3.c",
1978 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001979 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001980 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001981 "src/math/sigmoid-avx2-rr2-p5-div.c",
1982 "src/math/sigmoid-avx2-rr1-p5-div.c",
1983 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1984 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1985 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1986 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001987]
1988
Marat Dukhan08c4a432019-10-03 09:29:21 -07001989AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001990 "src/f32-clamp/gen/avx512f-x16.c",
1991 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001992 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1993 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1994 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1995 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1996 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1997 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1998 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1999 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2000 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2001 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2002 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2003 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2004 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2005 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2006 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2007 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2008 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2009 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2010 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2011 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2012 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2013 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2014 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2015 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002016 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2017 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002018 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2019 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2020 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2021 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2022 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2023 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002024 "src/f32-prelu/gen/avx512f-2x16.c",
2025 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002026 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2027 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2028 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2029 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2030 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2031 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2032 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2033 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2034 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2035 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2036 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2037 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2038 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2039 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2040 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2041 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2042 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2043 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2044 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2045 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2046 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2047 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2048 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2049 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2050 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2051 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2052 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2053 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2054 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2055 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2056 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2057 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2058 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2059 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2060 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2061 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002062 "src/f32-relu/gen/avx512f-x16.c",
2063 "src/f32-relu/gen/avx512f-x32.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 Dukhan19dd91d2020-07-16 11:12:44 -07002097 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2098 "src/f32-vlrelu/gen/vlrelu-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",
Frank Barchard490febe2020-07-16 18:42:17 -07002161 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002162 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002163 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002164 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002165 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2166 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002167 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2168 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2169 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2170 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.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 Dukhan08c4a432019-10-03 09:29:21 -07002251 "src/xnnpack/argmaxpool.h",
2252 "src/xnnpack/avgpool.h",
2253 "src/xnnpack/clamp.h",
2254 "src/xnnpack/common.h",
2255 "src/xnnpack/conv.h",
2256 "src/xnnpack/dwconv.h",
2257 "src/xnnpack/gavgpool.h",
2258 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002259 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002260 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002261 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002262 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002263 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002264 "src/xnnpack/lut.h",
2265 "src/xnnpack/math.h",
2266 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002267 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002268 "src/xnnpack/packx.h",
2269 "src/xnnpack/pad.h",
2270 "src/xnnpack/params.h",
2271 "src/xnnpack/pavgpool.h",
2272 "src/xnnpack/ppmm.h",
2273 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002274 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002275 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002276 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002277 "src/xnnpack/rmax.h",
2278 "src/xnnpack/scalar-utils.h",
2279 "src/xnnpack/spmm.h",
2280 "src/xnnpack/unpool.h",
2281 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002282 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002283 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002284 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002285 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002286 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002287 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002288 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002289]
2290
2291INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002292 "include/xnnpack.h",
2293 "src/xnnpack/allocator.h",
2294 "src/xnnpack/compute.h",
2295 "src/xnnpack/im2col.h",
2296 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002297 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002298 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002299 "src/xnnpack/operator.h",
2300 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002301 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002302 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002303 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002304 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002305]
2306
Marat Dukhan1b354632020-03-23 12:50:22 -07002307ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002308 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002309]
2310
Marat Dukhan1b354632020-03-23 12:50:22 -07002311MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002312 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002313 "include/xnnpack.h",
2314]
2315
Marat Dukhan1b354632020-03-23 12:50:22 -07002316MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002317 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002318 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002319 "src/xnnpack/requantization.h",
2320 "include/xnnpack.h",
2321]
2322
2323OPERATOR_TEST_PARAMS_HDRS = [
2324 "src/xnnpack/params.h",
2325 "src/xnnpack/common.h",
2326]
2327
2328WEIGHTS_PACK_HDRS = [
2329 "src/xnnpack/pack.h",
2330 "src/xnnpack/operator.h",
2331 "src/xnnpack/compute.h",
2332]
2333
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002334LOGGING_COPTS = select({
2335 # No logging in optimized mode
2336 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2337 # Full logging in debug mode
2338 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2339 # Error-only logging in default (fastbuild) mode
2340 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2341})
2342
Marat Dukhan3b59de22020-06-03 20:15:19 -07002343LOGGING_SRCS = select({
2344 # No logging in optimized mode
2345 ":optimized_build": [],
2346 "//conditions:default": [
2347 "src/operator-strings.c",
2348 "src/subgraph-strings.c",
2349 ],
2350})
2351
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002352LOGGING_HDRS = [
2353 "src/xnnpack/log.h",
2354]
2355
Marat Dukhan08c4a432019-10-03 09:29:21 -07002356xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002357 name = "tables",
2358 srcs = TABLE_SRCS,
2359 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002360 gcc_copts = xnnpack_gcc_std_copts(),
2361 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002362)
2363
2364xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002365 name = "scalar_ukernels",
2366 srcs = SCALAR_UKERNELS,
2367 hdrs = INTERNAL_HDRS,
2368 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002369 gcc_copts = xnnpack_gcc_std_copts(),
2370 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002371 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002372 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002373 "@FP16",
2374 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002375 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002376 ],
2377)
2378
2379xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002380 name = "scalar_ukernels_test_mode",
2381 srcs = SCALAR_UKERNELS,
2382 hdrs = INTERNAL_HDRS,
2383 aarch32_copts = ["-marm"],
2384 copts = [
2385 "-UNDEBUG",
2386 "-DXNN_TEST_MODE=1",
2387 ],
2388 gcc_copts = xnnpack_gcc_std_copts(),
2389 msvc_copts = xnnpack_msvc_std_copts(),
2390 deps = [
2391 ":tables",
2392 "@FP16",
2393 "@FXdiv",
2394 "@pthreadpool",
2395 ],
2396)
2397
2398xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002399 name = "wasm_ukernels",
2400 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002401 gcc_copts = xnnpack_gcc_std_copts(),
2402 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002403 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002404 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002405 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002406 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002407 "@FP16",
2408 "@FXdiv",
2409 "@pthreadpool",
2410 ],
2411)
2412
2413xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002414 name = "wasm_ukernels_test_mode",
2415 hdrs = INTERNAL_HDRS,
2416 copts = [
2417 "-UNDEBUG",
2418 "-DXNN_TEST_MODE=1",
2419 ],
2420 gcc_copts = xnnpack_gcc_std_copts(),
2421 msvc_copts = xnnpack_msvc_std_copts(),
2422 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002423 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002424 deps = [
2425 ":tables",
2426 "@FP16",
2427 "@FXdiv",
2428 "@pthreadpool",
2429 ],
2430)
2431
2432xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002433 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002434 hdrs = INTERNAL_HDRS,
2435 aarch32_copts = [
2436 "-marm",
2437 "-mfpu=neon",
2438 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002439 gcc_copts = xnnpack_gcc_std_copts(),
2440 gcc_x86_copts = ["-msse2"],
2441 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002442 optimized_copts = [
2443 "-O3",
2444 "-ffast-math",
2445 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002446 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002447 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002448 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002449 "@FP16",
2450 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002451 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002452 ],
2453)
2454
2455xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002456 name = "psimd_fastmath_ukernels_test_mode",
2457 hdrs = INTERNAL_HDRS,
2458 aarch32_copts = [
2459 "-marm",
2460 "-mfpu=neon",
2461 ],
2462 copts = [
2463 "-UNDEBUG",
2464 "-DXNN_TEST_MODE=1",
2465 ],
2466 gcc_copts = xnnpack_gcc_std_copts(),
2467 gcc_x86_copts = ["-msse2"],
2468 msvc_copts = xnnpack_msvc_std_copts(),
2469 optimized_copts = [
2470 "-O3",
2471 "-ffast-math",
2472 ],
2473 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2474 deps = [
2475 ":tables",
2476 "@FP16",
2477 "@psimd",
2478 "@pthreadpool",
2479 ],
2480)
2481
2482xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002483 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002484 hdrs = INTERNAL_HDRS,
2485 aarch32_copts = [
2486 "-marm",
2487 "-mfpu=neon",
2488 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002489 gcc_copts = xnnpack_gcc_std_copts(),
2490 gcc_x86_copts = ["-msse2"],
2491 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002492 optimized_copts = [
2493 "-O3",
2494 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002495 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002496 deps = [
2497 ":tables",
2498 "@FP16",
2499 "@psimd",
2500 "@pthreadpool",
2501 ],
2502)
2503
2504xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002505 name = "psimd_accmath_ukernels_test_mode",
2506 hdrs = INTERNAL_HDRS,
2507 aarch32_copts = [
2508 "-marm",
2509 "-mfpu=neon",
2510 ],
2511 copts = [
2512 "-UNDEBUG",
2513 "-DXNN_TEST_MODE=1",
2514 ],
2515 gcc_copts = xnnpack_gcc_std_copts(),
2516 gcc_x86_copts = ["-msse2"],
2517 msvc_copts = xnnpack_msvc_std_copts(),
2518 optimized_copts = [
2519 "-O3",
2520 ],
2521 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2522 deps = [
2523 ":tables",
2524 "@FP16",
2525 "@psimd",
2526 "@pthreadpool",
2527 ],
2528)
2529
2530xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002531 name = "neon_ukernels",
2532 hdrs = INTERNAL_HDRS,
2533 aarch32_copts = [
2534 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002535 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002536 "-mfpu=neon",
2537 ],
2538 aarch32_srcs = NEON_UKERNELS,
2539 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002540 gcc_copts = xnnpack_gcc_std_copts(),
2541 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002542 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002543 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002544 "@FP16",
2545 "@pthreadpool",
2546 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002547)
2548
2549xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002550 name = "neon_ukernels_test_mode",
2551 hdrs = INTERNAL_HDRS,
2552 aarch32_copts = [
2553 "-marm",
2554 "-march=armv7-a",
2555 "-mfpu=neon",
2556 ],
2557 aarch32_srcs = NEON_UKERNELS,
2558 aarch64_srcs = NEON_UKERNELS,
2559 copts = [
2560 "-UNDEBUG",
2561 "-DXNN_TEST_MODE=1",
2562 ],
2563 gcc_copts = xnnpack_gcc_std_copts(),
2564 msvc_copts = xnnpack_msvc_std_copts(),
2565 deps = [
2566 ":tables",
2567 "@FP16",
2568 "@pthreadpool",
2569 ],
2570)
2571
2572xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002573 name = "neonfma_ukernels",
2574 hdrs = INTERNAL_HDRS,
2575 aarch32_copts = [
2576 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002577 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002578 "-mfpu=neon-vfpv4",
2579 ],
2580 aarch32_srcs = NEONFMA_UKERNELS,
2581 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002582 apple_aarch32_copts = [
2583 "-mcpu=swift",
2584 "-mtune=generic",
2585 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002586 gcc_copts = xnnpack_gcc_std_copts(),
2587 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002588 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002589 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002590 "@FP16",
2591 "@pthreadpool",
2592 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002593)
2594
2595xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002596 name = "neonfma_ukernels_test_mode",
2597 hdrs = INTERNAL_HDRS,
2598 aarch32_copts = [
2599 "-marm",
2600 "-march=armv7-a",
2601 "-mfpu=neon-vfpv4",
2602 ],
2603 aarch32_srcs = NEONFMA_UKERNELS,
2604 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002605 apple_aarch32_copts = [
2606 "-mcpu=swift",
2607 "-mtune=generic",
2608 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002609 copts = [
2610 "-UNDEBUG",
2611 "-DXNN_TEST_MODE=1",
2612 ],
2613 gcc_copts = xnnpack_gcc_std_copts(),
2614 msvc_copts = xnnpack_msvc_std_copts(),
2615 deps = [
2616 ":tables",
2617 "@FP16",
2618 "@pthreadpool",
2619 ],
2620)
2621
2622xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002623 name = "neonv8_ukernels",
2624 hdrs = INTERNAL_HDRS,
2625 aarch32_copts = [
2626 "-marm",
2627 "-march=armv8-a",
2628 "-mfpu=neon-fp-armv8",
2629 ],
2630 aarch32_srcs = NEONV8_UKERNELS,
2631 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002632 apple_aarch32_copts = [
2633 "-mcpu=cyclone",
2634 "-mtune=generic",
2635 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002636 gcc_copts = xnnpack_gcc_std_copts(),
2637 msvc_copts = xnnpack_msvc_std_copts(),
2638 deps = [
2639 ":tables",
2640 "@FP16",
2641 "@pthreadpool",
2642 ],
2643)
2644
2645xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002646 name = "neonv8_ukernels_test_mode",
2647 hdrs = INTERNAL_HDRS,
2648 aarch32_copts = [
2649 "-marm",
2650 "-march=armv8-a",
2651 "-mfpu=neon-fp-armv8",
2652 ],
2653 aarch32_srcs = NEONV8_UKERNELS,
2654 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002655 apple_aarch32_copts = [
2656 "-mcpu=cyclone",
2657 "-mtune=generic",
2658 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002659 copts = [
2660 "-UNDEBUG",
2661 "-DXNN_TEST_MODE=1",
2662 ],
2663 gcc_copts = xnnpack_gcc_std_copts(),
2664 msvc_copts = xnnpack_msvc_std_copts(),
2665 deps = [
2666 ":tables",
2667 "@FP16",
2668 "@pthreadpool",
2669 ],
2670)
2671
2672xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002673 name = "neonfp16arith_ukernels",
2674 hdrs = INTERNAL_HDRS,
2675 aarch64_copts = ["-march=armv8.2-a+fp16"],
2676 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002677 gcc_copts = xnnpack_gcc_std_copts(),
2678 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002679 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002680 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002681 "@FP16",
2682 "@pthreadpool",
2683 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002684)
2685
2686xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002687 name = "neonfp16arith_ukernels_test_mode",
2688 hdrs = INTERNAL_HDRS,
2689 aarch64_copts = ["-march=armv8.2-a+fp16"],
2690 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2691 copts = [
2692 "-UNDEBUG",
2693 "-DXNN_TEST_MODE=1",
2694 ],
2695 gcc_copts = xnnpack_gcc_std_copts(),
2696 msvc_copts = xnnpack_msvc_std_copts(),
2697 deps = [
2698 ":tables",
2699 "@FP16",
2700 "@pthreadpool",
2701 ],
2702)
2703
2704xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002705 name = "sse2_ukernels",
2706 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002707 gcc_copts = xnnpack_gcc_std_copts(),
2708 gcc_x86_copts = ["-msse2"],
2709 msvc_copts = xnnpack_msvc_std_copts(),
2710 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002711 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002712 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002713 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002714 "@FP16",
2715 "@pthreadpool",
2716 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002717)
2718
2719xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002720 name = "sse2_ukernels_test_mode",
2721 hdrs = INTERNAL_HDRS,
2722 copts = [
2723 "-UNDEBUG",
2724 "-DXNN_TEST_MODE=1",
2725 ],
2726 gcc_copts = xnnpack_gcc_std_copts(),
2727 gcc_x86_copts = ["-msse2"],
2728 msvc_copts = xnnpack_msvc_std_copts(),
2729 msvc_x86_32_copts = ["/arch:SSE2"],
2730 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2731 deps = [
2732 ":tables",
2733 "@FP16",
2734 "@pthreadpool",
2735 ],
2736)
2737
2738xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002739 name = "ssse3_ukernels",
2740 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002741 gcc_copts = xnnpack_gcc_std_copts(),
2742 gcc_x86_copts = ["-mssse3"],
2743 msvc_copts = xnnpack_msvc_std_copts(),
2744 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002745 x86_srcs = SSSE3_UKERNELS,
2746 deps = [
2747 ":tables",
2748 "@FP16",
2749 "@pthreadpool",
2750 ],
2751)
2752
2753xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002754 name = "ssse3_ukernels_test_mode",
2755 hdrs = INTERNAL_HDRS,
2756 copts = [
2757 "-UNDEBUG",
2758 "-DXNN_TEST_MODE=1",
2759 ],
2760 gcc_copts = xnnpack_gcc_std_copts(),
2761 gcc_x86_copts = ["-mssse3"],
2762 msvc_copts = xnnpack_msvc_std_copts(),
2763 msvc_x86_32_copts = ["/arch:SSE2"],
2764 x86_srcs = SSSE3_UKERNELS,
2765 deps = [
2766 ":tables",
2767 "@FP16",
2768 "@pthreadpool",
2769 ],
2770)
2771
2772xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002773 name = "sse41_ukernels",
2774 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002775 gcc_copts = xnnpack_gcc_std_copts(),
2776 gcc_x86_copts = ["-msse4.1"],
2777 msvc_copts = xnnpack_msvc_std_copts(),
2778 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002779 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002780 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002781 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002782 "@FP16",
2783 "@pthreadpool",
2784 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002785)
2786
2787xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002788 name = "sse41_ukernels_test_mode",
2789 hdrs = INTERNAL_HDRS,
2790 copts = [
2791 "-UNDEBUG",
2792 "-DXNN_TEST_MODE=1",
2793 ],
2794 gcc_copts = xnnpack_gcc_std_copts(),
2795 gcc_x86_copts = ["-msse4.1"],
2796 msvc_copts = xnnpack_msvc_std_copts(),
2797 msvc_x86_32_copts = ["/arch:SSE2"],
2798 x86_srcs = SSE41_UKERNELS,
2799 deps = [
2800 ":tables",
2801 "@FP16",
2802 "@pthreadpool",
2803 ],
2804)
2805
2806xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002807 name = "avx_ukernels",
2808 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002809 gcc_copts = xnnpack_gcc_std_copts(),
2810 gcc_x86_copts = ["-mavx"],
2811 msvc_copts = xnnpack_msvc_std_copts(),
2812 msvc_x86_32_copts = ["/arch:AVX"],
2813 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002814 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002815 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002816 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002817 "@FP16",
2818 "@pthreadpool",
2819 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002820)
2821
2822xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002823 name = "avx_ukernels_test_mode",
2824 hdrs = INTERNAL_HDRS,
2825 copts = [
2826 "-UNDEBUG",
2827 "-DXNN_TEST_MODE=1",
2828 ],
2829 gcc_copts = xnnpack_gcc_std_copts(),
2830 gcc_x86_copts = ["-mavx"],
2831 msvc_copts = xnnpack_msvc_std_copts(),
2832 msvc_x86_32_copts = ["/arch:AVX"],
2833 msvc_x86_64_copts = ["/arch:AVX"],
2834 x86_srcs = AVX_UKERNELS,
2835 deps = [
2836 ":tables",
2837 "@FP16",
2838 "@pthreadpool",
2839 ],
2840)
2841
2842xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002843 name = "fma3_ukernels",
2844 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002845 gcc_copts = xnnpack_gcc_std_copts(),
2846 gcc_x86_copts = ["-mfma"],
2847 msvc_copts = xnnpack_msvc_std_copts(),
2848 msvc_x86_32_copts = ["/arch:AVX"],
2849 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002850 x86_srcs = FMA3_UKERNELS,
2851 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002852 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002853 "@FP16",
2854 "@pthreadpool",
2855 ],
2856)
2857
2858xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002859 name = "fma3_ukernels_test_mode",
2860 hdrs = INTERNAL_HDRS,
2861 copts = [
2862 "-UNDEBUG",
2863 "-DXNN_TEST_MODE=1",
2864 ],
2865 gcc_copts = xnnpack_gcc_std_copts(),
2866 gcc_x86_copts = ["-mfma"],
2867 msvc_copts = xnnpack_msvc_std_copts(),
2868 msvc_x86_32_copts = ["/arch:AVX"],
2869 msvc_x86_64_copts = ["/arch:AVX"],
2870 x86_srcs = FMA3_UKERNELS,
2871 deps = [
2872 ":tables",
2873 "@FP16",
2874 "@pthreadpool",
2875 ],
2876)
2877
2878xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002879 name = "avx2_ukernels",
2880 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002881 gcc_copts = xnnpack_gcc_std_copts(),
2882 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002883 "-mfma",
2884 "-mavx2",
2885 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002886 msvc_copts = xnnpack_msvc_std_copts(),
2887 msvc_x86_32_copts = ["/arch:AVX2"],
2888 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002889 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002890 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002891 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002892 "@FP16",
2893 "@pthreadpool",
2894 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002895)
2896
2897xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002898 name = "avx2_ukernels_test_mode",
2899 hdrs = INTERNAL_HDRS,
2900 copts = [
2901 "-UNDEBUG",
2902 "-DXNN_TEST_MODE=1",
2903 ],
2904 gcc_copts = xnnpack_gcc_std_copts(),
2905 gcc_x86_copts = [
2906 "-mfma",
2907 "-mavx2",
2908 ],
2909 msvc_copts = xnnpack_msvc_std_copts(),
2910 msvc_x86_32_copts = ["/arch:AVX2"],
2911 msvc_x86_64_copts = ["/arch:AVX2"],
2912 x86_srcs = AVX2_UKERNELS,
2913 deps = [
2914 ":tables",
2915 "@FP16",
2916 "@pthreadpool",
2917 ],
2918)
2919
2920xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002921 name = "avx512f_ukernels",
2922 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002923 gcc_copts = xnnpack_gcc_std_copts(),
2924 gcc_x86_copts = ["-mavx512f"],
2925 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2926 msvc_copts = xnnpack_msvc_std_copts(),
2927 msvc_x86_32_copts = ["/arch:AVX512"],
2928 msvc_x86_64_copts = ["/arch:AVX512"],
2929 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002930 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002931 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002932 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002933 "@FP16",
2934 "@pthreadpool",
2935 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002936)
2937
2938xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002939 name = "avx512f_ukernels_test_mode",
2940 hdrs = INTERNAL_HDRS,
2941 copts = [
2942 "-UNDEBUG",
2943 "-DXNN_TEST_MODE=1",
2944 ],
2945 gcc_copts = xnnpack_gcc_std_copts(),
2946 gcc_x86_copts = ["-mavx512f"],
2947 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2948 msvc_copts = xnnpack_msvc_std_copts(),
2949 msvc_x86_32_copts = ["/arch:AVX512"],
2950 msvc_x86_64_copts = ["/arch:AVX512"],
2951 msys_copts = ["-fno-asynchronous-unwind-tables"],
2952 x86_srcs = AVX512F_UKERNELS,
2953 deps = [
2954 ":tables",
2955 "@FP16",
2956 "@pthreadpool",
2957 ],
2958)
2959
2960xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002961 name = "asm_ukernels",
2962 hdrs = ["src/xnnpack/assembly.h"],
2963 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002964 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002965 aarch64_srcs = AARCH64_ASM_UKERNELS,
2966)
2967
Marat Dukhan3b59de22020-06-03 20:15:19 -07002968xnnpack_cc_library(
2969 name = "logging_utils",
2970 srcs = LOGGING_SRCS,
2971 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2972 copts = LOGGING_COPTS + [
2973 "-Isrc",
2974 "-Iinclude",
2975 ] + select({
2976 ":debug_build": [],
2977 "//conditions:default": xnnpack_min_size_copts(),
2978 }),
2979 gcc_copts = xnnpack_gcc_std_copts(),
2980 msvc_copts = xnnpack_msvc_std_copts(),
2981 visibility = xnnpack_visibility(),
2982 deps = [
2983 "@FP16",
2984 "@clog",
2985 "@pthreadpool",
2986 ],
2987)
2988
Marat Dukhan08c4a432019-10-03 09:29:21 -07002989xnnpack_aggregate_library(
2990 name = "ukernels",
2991 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002992 ":neon_ukernels",
2993 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002994 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002995 ":asm_ukernels",
2996 ],
2997 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002998 ":neon_ukernels",
2999 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003000 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003001 ":neonfp16arith_ukernels",
3002 ":asm_ukernels",
3003 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003004 generic_deps = [
3005 ":scalar_ukernels",
3006 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003007 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003008 ":psimd_fastmath_ukernels",
3009 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003010 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003011 wasm_deps = [
3012 ":wasm_ukernels",
3013 ],
3014 wasmsimd_deps = [
3015 ":wasm_ukernels",
3016 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003017 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003018 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003019 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003020 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003021 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003022 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003023 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003024 ":avx512f_ukernels",
3025 ],
3026)
3027
Marat Dukhan33fcf782020-05-24 14:27:15 -07003028xnnpack_aggregate_library(
3029 name = "ukernels_test_mode",
3030 aarch32_deps = [
3031 ":neon_ukernels_test_mode",
3032 ":neonfma_ukernels_test_mode",
3033 ":neonv8_ukernels_test_mode",
3034 ":asm_ukernels",
3035 ],
3036 aarch64_deps = [
3037 ":neon_ukernels_test_mode",
3038 ":neonfma_ukernels_test_mode",
3039 ":neonv8_ukernels_test_mode",
3040 ":neonfp16arith_ukernels_test_mode",
3041 ":asm_ukernels",
3042 ],
3043 generic_deps = [
3044 ":scalar_ukernels_test_mode",
3045 ],
3046 psimd_deps = [
3047 ":psimd_fastmath_ukernels_test_mode",
3048 ":psimd_accmath_ukernels_test_mode",
3049 ],
3050 wasm_deps = [
3051 ":wasm_ukernels_test_mode",
3052 ],
3053 wasmsimd_deps = [
3054 ":wasm_ukernels_test_mode",
3055 ],
3056 x86_deps = [
3057 ":sse2_ukernels_test_mode",
3058 ":ssse3_ukernels_test_mode",
3059 ":sse41_ukernels_test_mode",
3060 ":avx_ukernels_test_mode",
3061 ":fma3_ukernels_test_mode",
3062 ":avx2_ukernels_test_mode",
3063 ":avx512f_ukernels_test_mode",
3064 ],
3065)
3066
Marat Dukhan86df3842020-06-17 14:03:31 -07003067# This target must be compatible with Android NDK r17 and compile with Clang 6.
3068# It is similar to :ukernels target, but excludes microkernels which require
3069# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003070xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003071 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003072 aarch32_deps = [
3073 ":neon_ukernels",
3074 ":neonfma_ukernels",
3075 ":neonv8_ukernels",
3076 ":asm_ukernels",
3077 ],
3078 aarch64_deps = [
3079 ":neon_ukernels",
3080 ":neonfma_ukernels",
3081 ":neonv8_ukernels",
3082 ":asm_ukernels",
3083 ],
3084 generic_deps = [
3085 ":scalar_ukernels",
3086 ],
3087 psimd_deps = [
3088 ":psimd_fastmath_ukernels",
3089 ":psimd_accmath_ukernels",
3090 ],
3091 wasm_deps = [
3092 ":wasm_ukernels",
3093 ],
3094 wasmsimd_deps = [
3095 ":wasm_ukernels",
3096 ],
3097 x86_deps = [
3098 ":sse2_ukernels",
3099 ":ssse3_ukernels",
3100 ":sse41_ukernels",
3101 ":avx_ukernels",
3102 ":fma3_ukernels",
3103 ":avx2_ukernels",
3104 ":avx512f_ukernels",
3105 ],
3106)
3107
Marat Dukhan08c4a432019-10-03 09:29:21 -07003108xnnpack_cc_library(
3109 name = "im2col",
3110 srcs = ["src/im2col.c"],
3111 hdrs = [
3112 "src/xnnpack/common.h",
3113 "src/xnnpack/im2col.h",
3114 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003115 gcc_copts = xnnpack_gcc_std_copts(),
3116 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003117)
3118
3119xnnpack_cc_library(
3120 name = "indirection",
3121 srcs = ["src/indirection.c"],
3122 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003123 gcc_copts = xnnpack_gcc_std_copts(),
3124 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003125 deps = [
3126 "@FP16",
3127 "@FXdiv",
3128 "@pthreadpool",
3129 ],
3130)
3131
3132xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003133 name = "indirection_test_mode",
3134 srcs = ["src/indirection.c"],
3135 hdrs = INTERNAL_HDRS,
3136 copts = [
3137 "-UNDEBUG",
3138 "-DXNN_TEST_MODE=1",
3139 ],
3140 gcc_copts = xnnpack_gcc_std_copts(),
3141 msvc_copts = xnnpack_msvc_std_copts(),
3142 deps = [
3143 "@FP16",
3144 "@FXdiv",
3145 "@pthreadpool",
3146 ],
3147)
3148
3149xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003150 name = "packing",
3151 srcs = ["src/packing.c"],
3152 hdrs = INTERNAL_HDRS,
3153 gcc_copts = xnnpack_gcc_std_copts(),
3154 msvc_copts = xnnpack_msvc_std_copts(),
3155 deps = [
3156 "@FP16",
3157 "@FXdiv",
3158 "@pthreadpool",
3159 ],
3160)
3161
3162xnnpack_cc_library(
3163 name = "packing_test_mode",
3164 srcs = ["src/packing.c"],
3165 hdrs = INTERNAL_HDRS,
3166 copts = [
3167 "-UNDEBUG",
3168 "-DXNN_TEST_MODE=1",
3169 ],
3170 gcc_copts = xnnpack_gcc_std_copts(),
3171 msvc_copts = xnnpack_msvc_std_copts(),
3172 deps = [
3173 "@FP16",
3174 "@FXdiv",
3175 "@pthreadpool",
3176 ],
3177)
3178
3179xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003180 name = "operator_run",
3181 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003182 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003183 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003184 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3185 "//conditions:default": [],
3186 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003187 gcc_copts = xnnpack_gcc_std_copts(),
3188 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003189 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003190 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003191 "@FP16",
3192 "@FXdiv",
3193 "@clog",
3194 "@pthreadpool",
3195 ],
3196)
3197
Chao Mei6ddfc602020-05-13 22:29:36 -07003198xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003199 name = "operator_run_test_mode",
3200 srcs = ["src/operator-run.c"],
3201 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3202 copts = LOGGING_COPTS + [
3203 "-UNDEBUG",
3204 "-DXNN_TEST_MODE=1",
3205 ] + select({
3206 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3207 "//conditions:default": [],
3208 }),
3209 gcc_copts = xnnpack_gcc_std_copts(),
3210 msvc_copts = xnnpack_msvc_std_copts(),
3211 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003212 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003213 "@FP16",
3214 "@FXdiv",
3215 "@clog",
3216 "@pthreadpool",
3217 ],
3218)
3219
3220xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003221 name = "memory_planner",
3222 srcs = ["src/memory-planner.c"],
3223 hdrs = INTERNAL_HDRS,
3224 defines = select({
3225 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3226 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3227 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3228 }),
3229 gcc_copts = xnnpack_gcc_std_copts(),
3230 msvc_copts = xnnpack_msvc_std_copts(),
3231 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003232 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003233 "@pthreadpool",
3234 ],
3235)
3236
Marat Dukhan33fcf782020-05-24 14:27:15 -07003237xnnpack_cc_library(
3238 name = "memory_planner_test_mode",
3239 srcs = ["src/memory-planner.c"],
3240 hdrs = INTERNAL_HDRS,
3241 copts = [
3242 "-UNDEBUG",
3243 "-DXNN_TEST_MODE=1",
3244 ],
3245 defines = select({
3246 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3247 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3248 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3249 }),
3250 gcc_copts = xnnpack_gcc_std_copts(),
3251 msvc_copts = xnnpack_msvc_std_copts(),
3252 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003253 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003254 "@pthreadpool",
3255 ],
3256)
3257
Marat Dukhan08c4a432019-10-03 09:29:21 -07003258cc_library(
3259 name = "enable_assembly",
3260 defines = select({
3261 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3262 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003263 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003264 }),
3265)
3266
Marat Dukhan9de90e02020-06-18 16:04:12 -07003267cc_library(
3268 name = "enable_sparse",
3269 defines = select({
3270 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3271 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003272 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003273 }),
3274)
3275
Marat Dukhancf056b22019-10-07 10:26:29 -07003276xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003277 name = "operators",
3278 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003279 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003280 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003281 ],
3282 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003283 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003284 "-Isrc",
3285 "-Iinclude",
3286 ] + select({
3287 ":debug_build": [],
3288 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003289 }) + select({
3290 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3291 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003292 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003293 gcc_copts = xnnpack_gcc_std_copts(),
3294 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003295 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003296 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003297 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003298 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003299 "@FP16",
3300 "@FXdiv",
3301 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003302 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003303 ],
3304)
3305
Marat Dukhan10a38082020-04-17 03:58:35 -07003306xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003307 name = "operators_test_mode",
3308 srcs = OPERATOR_SRCS + [
3309 "src/memory.c",
3310 "src/operator-delete.c",
3311 ],
3312 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3313 copts = LOGGING_COPTS + [
3314 "-Isrc",
3315 "-Iinclude",
3316 "-UNDEBUG",
3317 "-DXNN_TEST_MODE=1",
3318 ] + select({
3319 ":debug_build": [],
3320 "//conditions:default": xnnpack_min_size_copts(),
3321 }) + select({
3322 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3323 "//conditions:default": [],
3324 }),
3325 gcc_copts = xnnpack_gcc_std_copts(),
3326 msvc_copts = xnnpack_msvc_std_copts(),
3327 deps = [
3328 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003329 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003330 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003331 "@FP16",
3332 "@FXdiv",
3333 "@clog",
3334 "@pthreadpool",
3335 ],
3336)
3337
3338xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003339 name = "XNNPACK",
3340 srcs = [
3341 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003342 "src/runtime.c",
3343 "src/subgraph.c",
3344 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003345 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003346 hdrs = ["include/xnnpack.h"],
3347 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003348 "-Isrc",
3349 "-Iinclude",
3350 ] + select({
3351 ":debug_build": [],
3352 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003353 }) + select({
3354 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3355 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003356 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003357 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003358 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003359 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003360 visibility = xnnpack_visibility(),
3361 deps = [
3362 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003363 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003364 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003365 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003366 ":operator_run",
3367 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003368 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003369 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003370 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003371 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003372 ] + select({
3373 ":emscripten": [],
3374 "//conditions:default": ["@cpuinfo"],
3375 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003376)
3377
Marat Dukhan10a38082020-04-17 03:58:35 -07003378xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003379 name = "XNNPACK_test_mode",
3380 srcs = [
3381 "src/init.c",
3382 "src/runtime.c",
3383 "src/subgraph.c",
3384 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003385 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003386 hdrs = ["include/xnnpack.h"],
3387 copts = LOGGING_COPTS + [
3388 "-Isrc",
3389 "-Iinclude",
3390 "-UNDEBUG",
3391 "-DXNN_TEST_MODE=1",
3392 ] + select({
3393 ":debug_build": [],
3394 "//conditions:default": xnnpack_min_size_copts(),
3395 }) + select({
3396 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3397 "//conditions:default": [],
3398 }),
3399 gcc_copts = xnnpack_gcc_std_copts(),
3400 includes = ["include"],
3401 msvc_copts = xnnpack_msvc_std_copts(),
3402 visibility = xnnpack_visibility(),
3403 deps = [
3404 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003405 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003406 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003407 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003408 ":operator_run_test_mode",
3409 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003410 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003411 "@clog",
3412 "@FP16",
3413 "@pthreadpool",
3414 ] + select({
3415 ":emscripten": [],
3416 "//conditions:default": ["@cpuinfo"],
3417 }),
3418)
3419
3420xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003421 name = "xnnpack_f32",
3422 srcs = [
3423 "src/init.c",
3424 "src/runtime.c",
3425 "src/subgraph.c",
3426 "src/tensor.c",
3427 ] + SUBGRAPH_SRCS,
3428 hdrs = ["include/xnnpack.h"],
3429 copts = LOGGING_COPTS + [
3430 "-Isrc",
3431 "-Iinclude",
3432 ] + select({
3433 ":debug_build": [],
3434 "//conditions:default": xnnpack_min_size_copts(),
3435 }) + select({
3436 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3437 "//conditions:default": [],
3438 }),
3439 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003440 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003441 "XNN_NO_U8_OPERATORS",
3442 "XNN_NO_X8_OPERATORS",
3443 "XNN_NO_F16_OPERATORS",
3444 "XNN_NO_X16_OPERATORS",
3445 ],
3446 gcc_copts = xnnpack_gcc_std_copts(),
3447 includes = ["include"],
3448 msvc_copts = xnnpack_msvc_std_copts(),
3449 visibility = xnnpack_visibility(),
3450 deps = [
3451 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003452 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003453 ":logging_utils",
3454 ":memory_planner",
3455 ":operator_run",
3456 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003457 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003458 "@clog",
3459 "@FP16",
3460 "@pthreadpool",
3461 ] + select({
3462 ":emscripten": [],
3463 "//conditions:default": ["@cpuinfo"],
3464 }),
3465)
3466
3467xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003468 name = "xnnpack_operators_nhwc_f32",
3469 srcs = [
3470 "src/init.c",
3471 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003472 hdrs = ["include/xnnpack.h"],
3473 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003474 "-Isrc",
3475 "-Iinclude",
3476 ] + select({
3477 ":debug_build": [],
3478 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003479 }) + select({
3480 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3481 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003482 }),
3483 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003484 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003485 "XNN_NO_U8_OPERATORS",
3486 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003487 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003488 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003489 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003490 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003491 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003492 visibility = xnnpack_visibility(),
3493 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003494 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003495 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003496 ":operator_run",
3497 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003498 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003499 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003500 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003501 ] + select({
3502 ":emscripten": [],
3503 "//conditions:default": ["@cpuinfo"],
3504 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003505)
3506
Marat Dukhancf056b22019-10-07 10:26:29 -07003507xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003508 name = "bench_utils",
3509 srcs = ["bench/utils.cc"],
3510 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003511 deps = [
3512 "@com_google_benchmark//:benchmark",
3513 "@cpuinfo",
3514 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003515)
3516
Frank Barchard7e955972019-10-11 10:34:25 -07003517######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003518
3519xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003520 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003521 srcs = [
3522 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003523 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003524 "src/xnnpack/AlignedAllocator.h",
3525 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003526 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003527 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003528)
3529
3530xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07003531 name = "qu8_requantization_bench",
3532 srcs = [
3533 "bench/qu8-requantization.cc",
3534 "src/xnnpack/requantization-stubs.h",
3535 "src/xnnpack/AlignedAllocator.h",
3536 ] + MICROKERNEL_BENCHMARK_HDRS,
3537 deps = MICROKERNEL_BENCHMARK_DEPS,
3538)
3539
3540xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003541 name = "f16_igemm_bench",
3542 srcs = [
3543 "bench/f16-igemm.cc",
3544 "bench/conv.h",
3545 "bench/google/conv.h",
3546 "src/xnnpack/AlignedAllocator.h",
3547 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003548 deps = MICROKERNEL_BENCHMARK_DEPS + [
3549 ":indirection",
3550 ":packing",
3551 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003552)
3553
3554xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003555 name = "f16_gemm_bench",
3556 srcs = [
3557 "bench/f16-gemm.cc",
3558 "bench/gemm.h",
3559 "src/xnnpack/AlignedAllocator.h",
3560 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003561 deps = MICROKERNEL_BENCHMARK_DEPS + [
3562 ":packing",
3563 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003564)
3565
3566xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003567 name = "f16_spmm_bench",
3568 srcs = [
3569 "bench/f16-spmm.cc",
3570 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003571 "src/xnnpack/AlignedAllocator.h",
3572 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003573 deps = MICROKERNEL_BENCHMARK_DEPS,
3574)
3575
3576xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003577 name = "f32_igemm_bench",
3578 srcs = [
3579 "bench/f32-igemm.cc",
3580 "bench/conv.h",
3581 "src/xnnpack/AlignedAllocator.h",
3582 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003583 deps = MICROKERNEL_BENCHMARK_DEPS + [
3584 ":indirection",
3585 ":packing",
3586 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003587)
3588
3589xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07003590 name = "f16_relu_bench",
3591 srcs = [
3592 "bench/f16-relu.cc",
3593 "src/xnnpack/AlignedAllocator.h",
3594 ] + MICROKERNEL_BENCHMARK_HDRS,
3595 deps = MICROKERNEL_BENCHMARK_DEPS,
3596)
3597
3598xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003599 name = "f32_conv_hwc_bench",
3600 srcs = [
3601 "bench/f32-conv-hwc.cc",
3602 "bench/dconv.h",
3603 "src/xnnpack/AlignedAllocator.h",
3604 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003605 deps = MICROKERNEL_BENCHMARK_DEPS + [
3606 ":packing",
3607 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003608)
3609
3610xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003611 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003612 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003613 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003614 "bench/dconv.h",
3615 "src/xnnpack/AlignedAllocator.h",
3616 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003617 deps = MICROKERNEL_BENCHMARK_DEPS + [
3618 ":packing",
3619 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003620)
3621
3622xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003623 name = "f16_dwconv_bench",
3624 srcs = [
3625 "bench/f16-dwconv.cc",
3626 "bench/dwconv.h",
3627 "bench/google/dwconv.h",
3628 "src/xnnpack/AlignedAllocator.h",
3629 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003630 deps = MICROKERNEL_BENCHMARK_DEPS + [
3631 ":indirection",
3632 ":packing",
3633 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003634)
3635
3636xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003637 name = "f32_dwconv_bench",
3638 srcs = [
3639 "bench/f32-dwconv.cc",
3640 "bench/dwconv.h",
3641 "src/xnnpack/AlignedAllocator.h",
3642 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003643 deps = MICROKERNEL_BENCHMARK_DEPS + [
3644 ":indirection",
3645 ":packing",
3646 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003647)
3648
3649xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003650 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003651 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003652 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003653 "bench/dwconv.h",
3654 "src/xnnpack/AlignedAllocator.h",
3655 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003656 deps = MICROKERNEL_BENCHMARK_DEPS + [
3657 ":indirection",
3658 ":packing",
3659 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003660)
3661
3662xnnpack_benchmark(
3663 name = "f32_gemm_bench",
3664 srcs = [
3665 "bench/f32-gemm.cc",
3666 "bench/gemm.h",
3667 "src/xnnpack/AlignedAllocator.h",
3668 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003669 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003670 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003671)
3672
3673xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003674 name = "f32_hswish_bench",
3675 srcs = [
3676 "bench/f32-hswish.cc",
3677 "src/xnnpack/AlignedAllocator.h",
3678 ] + MICROKERNEL_BENCHMARK_HDRS,
3679 deps = MICROKERNEL_BENCHMARK_DEPS,
3680)
3681
3682xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003683 name = "f32_raddexpminusmax_bench",
3684 srcs = [
3685 "bench/f32-raddexpminusmax.cc",
3686 "src/xnnpack/AlignedAllocator.h",
3687 ] + MICROKERNEL_BENCHMARK_HDRS,
3688 deps = MICROKERNEL_BENCHMARK_DEPS,
3689)
3690
3691xnnpack_benchmark(
3692 name = "f32_raddextexp_bench",
3693 srcs = [
3694 "bench/f32-raddextexp.cc",
3695 "src/xnnpack/AlignedAllocator.h",
3696 ] + MICROKERNEL_BENCHMARK_HDRS,
3697 deps = MICROKERNEL_BENCHMARK_DEPS,
3698)
3699
3700xnnpack_benchmark(
3701 name = "f32_raddstoreexpminusmax_bench",
3702 srcs = [
3703 "bench/f32-raddstoreexpminusmax.cc",
3704 "src/xnnpack/AlignedAllocator.h",
3705 ] + MICROKERNEL_BENCHMARK_HDRS,
3706 deps = MICROKERNEL_BENCHMARK_DEPS,
3707)
3708
3709xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07003710 name = "f32_relu_bench",
3711 srcs = [
3712 "bench/f32-relu.cc",
3713 "src/xnnpack/AlignedAllocator.h",
3714 ] + MICROKERNEL_BENCHMARK_HDRS,
3715 deps = MICROKERNEL_BENCHMARK_DEPS,
3716)
3717
3718xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003719 name = "f32_rmax_bench",
3720 srcs = [
3721 "bench/f32-rmax.cc",
3722 "src/xnnpack/AlignedAllocator.h",
3723 ] + MICROKERNEL_BENCHMARK_HDRS,
3724 deps = MICROKERNEL_BENCHMARK_DEPS,
3725)
3726
3727xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003728 name = "f32_sigmoid_bench",
3729 srcs = [
3730 "bench/f32-sigmoid.cc",
3731 "src/xnnpack/AlignedAllocator.h",
3732 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003733 deps = MICROKERNEL_BENCHMARK_DEPS,
3734)
3735
3736xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003737 name = "f32_spmm_bench",
3738 srcs = [
3739 "bench/f32-spmm.cc",
3740 "bench/gemm.h",
3741 "src/xnnpack/AlignedAllocator.h",
3742 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003743 deps = MICROKERNEL_BENCHMARK_DEPS,
3744)
3745
3746xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003747 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003748 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003749 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003750 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003751 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003752 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003753)
3754
3755xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003756 name = "f32_vscaleexpminusmax_bench",
3757 srcs = [
3758 "bench/f32-vscaleexpminusmax.cc",
3759 "src/xnnpack/AlignedAllocator.h",
3760 ] + MICROKERNEL_BENCHMARK_HDRS,
3761 deps = MICROKERNEL_BENCHMARK_DEPS,
3762)
3763
3764xnnpack_benchmark(
3765 name = "f32_vscaleextexp_bench",
3766 srcs = [
3767 "bench/f32-vscaleextexp.cc",
3768 "src/xnnpack/AlignedAllocator.h",
3769 ] + MICROKERNEL_BENCHMARK_HDRS,
3770 deps = MICROKERNEL_BENCHMARK_DEPS,
3771)
3772
3773xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003774 name = "f32_vsqrt_bench",
3775 srcs = [
3776 "bench/f32-vsqrt.cc",
3777 "src/xnnpack/AlignedAllocator.h",
3778 ] + MICROKERNEL_BENCHMARK_HDRS,
3779 deps = MICROKERNEL_BENCHMARK_DEPS,
3780)
3781
3782xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003783 name = "f32_im2col_gemm_bench",
3784 srcs = [
3785 "bench/f32-im2col-gemm.cc",
3786 "bench/conv.h",
3787 "src/xnnpack/AlignedAllocator.h",
3788 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003789 deps = MICROKERNEL_BENCHMARK_DEPS + [
3790 ":im2col",
3791 ":packing",
3792 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003793)
3794
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003795xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003796 name = "rounding_bench",
3797 srcs = [
3798 "bench/rounding.cc",
3799 "src/xnnpack/math-stubs.h",
3800 "src/xnnpack/AlignedAllocator.h",
3801 ] + MICROKERNEL_BENCHMARK_HDRS,
3802 deps = MICROKERNEL_BENCHMARK_DEPS,
3803)
3804
Marat Dukhan08c4a432019-10-03 09:29:21 -07003805########################### Benchmarks for operators ###########################
3806
3807xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003808 name = "average_pooling_bench",
3809 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003810 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003811 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003812 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003813)
3814
3815xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003816 name = "bankers_rounding_bench",
3817 srcs = ["bench/bankers-rounding.cc"],
3818 copts = xnnpack_optional_tflite_copts(),
3819 tags = ["nowin32"],
3820 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3821)
3822
3823xnnpack_benchmark(
3824 name = "ceiling_bench",
3825 srcs = ["bench/ceiling.cc"],
3826 copts = xnnpack_optional_tflite_copts(),
3827 tags = ["nowin32"],
3828 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3829)
3830
3831xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003832 name = "channel_shuffle_bench",
3833 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003834 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003835)
3836
3837xnnpack_benchmark(
3838 name = "convolution_bench",
3839 srcs = ["bench/convolution.cc"],
3840 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003841 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003842 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003843)
3844
3845xnnpack_benchmark(
3846 name = "deconvolution_bench",
3847 srcs = ["bench/deconvolution.cc"],
3848 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003849 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003850 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003851)
3852
3853xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003854 name = "floor_bench",
3855 srcs = ["bench/floor.cc"],
3856 copts = xnnpack_optional_tflite_copts(),
3857 tags = ["nowin32"],
3858 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3859)
3860
3861xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003862 name = "global_average_pooling_bench",
3863 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003864 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003865)
3866
3867xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07003868 name = "hardswish_bench",
3869 srcs = ["bench/hardswish.cc"],
3870 copts = xnnpack_optional_tflite_copts(),
3871 tags = ["nowin32"],
3872 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3873)
3874
3875xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003876 name = "max_pooling_bench",
3877 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003878 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003879)
3880
3881xnnpack_benchmark(
3882 name = "sigmoid_bench",
3883 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003884 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003885 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003886 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003887)
3888
3889xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003890 name = "prelu_bench",
3891 srcs = ["bench/prelu.cc"],
3892 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003893 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003894 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003895)
3896
3897xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003898 name = "softmax_bench",
3899 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003900 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003901 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003902 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003903)
3904
Marat Dukhan87727142020-06-24 15:24:10 -07003905xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07003906 name = "square_root_bench",
3907 srcs = ["bench/square-root.cc"],
3908 copts = xnnpack_optional_tflite_copts(),
3909 tags = ["nowin32"],
3910 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3911)
3912
3913xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003914 name = "truncation_bench",
3915 srcs = ["bench/truncation.cc"],
3916 deps = OPERATOR_BENCHMARK_DEPS,
3917)
3918
Marat Dukhanc068bb62019-10-04 13:24:39 -07003919############################# End-to-end benchmarks ############################
3920
3921cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003922 name = "fp32_mobilenet_v1",
3923 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003924 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003925 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003926 linkstatic = True,
3927 deps = [
3928 ":XNNPACK",
3929 "@pthreadpool",
3930 ],
3931)
3932
3933cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003934 name = "fp16_mobilenet_v1",
3935 srcs = ["models/fp16-mobilenet-v1.cc"],
3936 hdrs = ["models/models.h"],
3937 copts = xnnpack_std_cxxopts(),
3938 linkstatic = True,
3939 deps = [
3940 ":XNNPACK",
3941 "@FP16",
3942 "@pthreadpool",
3943 ],
3944)
3945
3946cc_library(
3947 name = "fp32_mobilenet_v2",
3948 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003949 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003950 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003951 linkstatic = True,
3952 deps = [
3953 ":XNNPACK",
3954 "@pthreadpool",
3955 ],
3956)
3957
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003958cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003959 name = "fp16_mobilenet_v2",
3960 srcs = ["models/fp16-mobilenet-v2.cc"],
3961 hdrs = ["models/models.h"],
3962 copts = xnnpack_std_cxxopts(),
3963 linkstatic = True,
3964 deps = [
3965 ":XNNPACK",
3966 "@FP16",
3967 "@pthreadpool",
3968 ],
3969)
3970
3971cc_library(
3972 name = "fp32_mobilenet_v3_large",
3973 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003974 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003975 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003976 linkstatic = True,
3977 deps = [
3978 ":XNNPACK",
3979 "@pthreadpool",
3980 ],
3981)
3982
3983cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003984 name = "fp32_mobilenet_v3_small",
3985 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003986 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003987 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003988 linkstatic = True,
3989 deps = [
3990 ":XNNPACK",
3991 "@pthreadpool",
3992 ],
3993)
3994
Marat Dukhanc068bb62019-10-04 13:24:39 -07003995xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07003996 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003997 srcs = [
3998 "bench/f32-dwconv-e2e.cc",
3999 "bench/end2end.h",
4000 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004001 deps = MICROKERNEL_BENCHMARK_DEPS + [
4002 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004003 ":fp32_mobilenet_v1",
4004 ":fp32_mobilenet_v2",
4005 ":fp32_mobilenet_v3_large",
4006 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004007 ],
4008)
4009
4010xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004011 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004012 srcs = [
4013 "bench/f32-gemm-e2e.cc",
4014 "bench/end2end.h",
4015 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004016 deps = MICROKERNEL_BENCHMARK_DEPS + [
4017 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004018 ":fp32_mobilenet_v1",
4019 ":fp32_mobilenet_v2",
4020 ":fp32_mobilenet_v3_large",
4021 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004022 ],
4023)
4024
4025xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004026 name = "end2end_bench",
4027 srcs = ["bench/end2end.cc"],
4028 deps = [
4029 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004030 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004031 ":fp16_mobilenet_v1",
4032 ":fp16_mobilenet_v2",
4033 ":fp32_mobilenet_v1",
4034 ":fp32_mobilenet_v2",
4035 ":fp32_mobilenet_v3_large",
4036 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004037 "@pthreadpool",
4038 ],
4039)
4040
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004041#################### Accuracy evaluation for math functions ####################
4042
4043xnnpack_benchmark(
4044 name = "f32_exp_eval",
4045 srcs = [
4046 "eval/f32-exp.cc",
4047 "src/xnnpack/AlignedAllocator.h",
4048 ] + ACCURACY_EVAL_HDRS,
4049 deps = ACCURACY_EVAL_DEPS,
4050)
4051
Marat Dukhan515c9772019-10-17 18:07:57 -07004052xnnpack_benchmark(
4053 name = "f32_expminus_eval",
4054 srcs = [
4055 "eval/f32-expminus.cc",
4056 "src/xnnpack/AlignedAllocator.h",
4057 ] + ACCURACY_EVAL_HDRS,
4058 deps = ACCURACY_EVAL_DEPS,
4059)
4060
Marat Dukhan98ba4412019-10-23 02:14:28 -07004061xnnpack_benchmark(
4062 name = "f32_extexp_eval",
4063 srcs = [
4064 "eval/f32-extexp.cc",
4065 "src/xnnpack/AlignedAllocator.h",
4066 ] + ACCURACY_EVAL_HDRS,
4067 deps = ACCURACY_EVAL_DEPS,
4068)
4069
Marat Dukhan8853b822020-05-07 12:19:01 -07004070xnnpack_unit_test(
4071 name = "f32_roundne_eval",
4072 srcs = [
4073 "eval/f32-roundne.cc",
4074 "src/xnnpack/AlignedAllocator.h",
4075 "src/xnnpack/math-stubs.h",
4076 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004077 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004078 deps = MICROKERNEL_TEST_DEPS,
4079)
4080
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004081xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004082 name = "f32_roundd_eval",
4083 srcs = [
4084 "eval/f32-roundd.cc",
4085 "src/xnnpack/AlignedAllocator.h",
4086 "src/xnnpack/math-stubs.h",
4087 ] + MICROKERNEL_TEST_HDRS,
4088 automatic = False,
4089 deps = MICROKERNEL_TEST_DEPS,
4090)
4091
4092xnnpack_unit_test(
4093 name = "f32_roundu_eval",
4094 srcs = [
4095 "eval/f32-roundu.cc",
4096 "src/xnnpack/AlignedAllocator.h",
4097 "src/xnnpack/math-stubs.h",
4098 ] + MICROKERNEL_TEST_HDRS,
4099 automatic = False,
4100 deps = MICROKERNEL_TEST_DEPS,
4101)
4102
4103xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004104 name = "f32_roundz_eval",
4105 srcs = [
4106 "eval/f32-roundz.cc",
4107 "src/xnnpack/AlignedAllocator.h",
4108 "src/xnnpack/math-stubs.h",
4109 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004110 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004111 deps = MICROKERNEL_TEST_DEPS,
4112)
4113
Marat Dukhan346a9e52019-11-15 09:06:30 -08004114xnnpack_benchmark(
4115 name = "f32_sigmoid_eval",
4116 srcs = [
4117 "eval/f32-sigmoid.cc",
4118 "src/xnnpack/AlignedAllocator.h",
4119 ] + ACCURACY_EVAL_HDRS,
4120 deps = ACCURACY_EVAL_DEPS,
4121)
4122
Marat Dukhan84000762020-06-29 18:38:43 -07004123xnnpack_benchmark(
4124 name = "f32_sqrt_eval",
4125 srcs = [
4126 "eval/f32-sqrt.cc",
4127 "src/xnnpack/AlignedAllocator.h",
4128 ] + ACCURACY_EVAL_HDRS,
4129 deps = ACCURACY_EVAL_DEPS + [
4130 ":bench_utils",
4131 ],
4132)
4133
Marat Dukhan08c4a432019-10-03 09:29:21 -07004134######################### Unit tests for micro-kernels #########################
4135
4136xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004137 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004138 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004139 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004140 "test/gemm-microkernel-tester.h",
4141 "src/xnnpack/AlignedAllocator.h",
4142 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004143 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004144)
4145
4146xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004147 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004148 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004149 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004150 "test/spmm-microkernel-tester.h",
4151 "src/xnnpack/AlignedAllocator.h",
4152 ] + MICROKERNEL_TEST_HDRS,
4153 deps = MICROKERNEL_TEST_DEPS,
4154)
4155
4156xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004157 name = "f32_argmaxpool_test",
4158 srcs = [
4159 "test/f32-argmaxpool.cc",
4160 "test/argmaxpool-microkernel-tester.h",
4161 "src/xnnpack/AlignedAllocator.h",
4162 ] + MICROKERNEL_TEST_HDRS,
4163 deps = MICROKERNEL_TEST_DEPS,
4164)
4165
4166xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004167 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004168 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004169 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004170 "test/avgpool-microkernel-tester.h",
4171 "src/xnnpack/AlignedAllocator.h",
4172 ] + MICROKERNEL_TEST_HDRS,
4173 deps = MICROKERNEL_TEST_DEPS,
4174)
4175
4176xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004177 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004178 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004179 "test/f32-ibilinear.cc",
4180 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004181 "src/xnnpack/AlignedAllocator.h",
4182 ] + MICROKERNEL_TEST_HDRS,
4183 deps = MICROKERNEL_TEST_DEPS,
4184)
4185
4186xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004187 name = "f16_clamp_test",
4188 srcs = [
4189 "test/f16-clamp.cc",
4190 "test/clamp-microkernel-tester.h",
4191 ] + MICROKERNEL_TEST_HDRS,
4192 deps = MICROKERNEL_TEST_DEPS,
4193)
4194
4195xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004196 name = "f32_clamp_test",
4197 srcs = [
4198 "test/f32-clamp.cc",
4199 "test/clamp-microkernel-tester.h",
4200 ] + MICROKERNEL_TEST_HDRS,
4201 deps = MICROKERNEL_TEST_DEPS,
4202)
4203
4204xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004205 name = "f32_igemm_test",
4206 srcs = [
4207 "test/f32-igemm.cc",
4208 "test/gemm-microkernel-tester.h",
4209 "src/xnnpack/AlignedAllocator.h",
4210 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004211 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004212)
4213
4214xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004215 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004216 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004217 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004218 "test/gemm-microkernel-tester.h",
4219 "src/xnnpack/AlignedAllocator.h",
4220 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004221 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004222)
4223
4224xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004225 name = "f32_igemm_minmax_test",
4226 srcs = [
4227 "test/f32-igemm-minmax.cc",
4228 "test/gemm-microkernel-tester.h",
4229 "src/xnnpack/AlignedAllocator.h",
4230 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004231 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004232)
4233
4234xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004235 name = "f16_igemm_minmax_test",
4236 srcs = [
4237 "test/f16-igemm-minmax.cc",
4238 "test/gemm-microkernel-tester.h",
4239 "src/xnnpack/AlignedAllocator.h",
4240 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004241 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004242)
4243
4244xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004245 name = "f32_conv_hwc_test",
4246 srcs = [
4247 "test/f32-conv-hwc.cc",
4248 "test/conv-hwc-microkernel-tester.h",
4249 "src/xnnpack/AlignedAllocator.h",
4250 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004251 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004252)
4253
4254xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004255 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004256 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004257 "test/f32-conv-hwc2chw.cc",
4258 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004259 "src/xnnpack/AlignedAllocator.h",
4260 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004261 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004262)
4263
4264xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004265 name = "f32_dwconv_test",
4266 srcs = [
4267 "test/f32-dwconv.cc",
4268 "test/dwconv-microkernel-tester.h",
4269 "src/xnnpack/AlignedAllocator.h",
4270 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004271 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004272)
4273
4274xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004275 name = "f16_dwconv_minmax_test",
4276 srcs = [
4277 "test/f16-dwconv-minmax.cc",
4278 "test/dwconv-microkernel-tester.h",
4279 "src/xnnpack/AlignedAllocator.h",
4280 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004281 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004282)
4283
4284xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004285 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004286 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004287 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004288 "test/dwconv-microkernel-tester.h",
4289 "src/xnnpack/AlignedAllocator.h",
4290 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004291 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004292)
4293
4294xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004295 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004296 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004297 "test/f32-dwconv-chw.cc",
4298 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004299 "src/xnnpack/AlignedAllocator.h",
4300 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004301 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004302)
4303
4304xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004305 name = "f16_gavgpool_minmax_test",
4306 srcs = [
4307 "test/f16-gavgpool-minmax.cc",
4308 "test/gavgpool-microkernel-tester.h",
4309 "src/xnnpack/AlignedAllocator.h",
4310 ] + MICROKERNEL_TEST_HDRS,
4311 deps = MICROKERNEL_TEST_DEPS,
4312)
4313
4314xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004315 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004316 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004317 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004318 "test/gavgpool-microkernel-tester.h",
4319 "src/xnnpack/AlignedAllocator.h",
4320 ] + MICROKERNEL_TEST_HDRS,
4321 deps = MICROKERNEL_TEST_DEPS,
4322)
4323
4324xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004325 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004326 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004327 "test/f32-gavgpool-cw.cc",
4328 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004329 "src/xnnpack/AlignedAllocator.h",
4330 ] + MICROKERNEL_TEST_HDRS,
4331 deps = MICROKERNEL_TEST_DEPS,
4332)
4333
4334xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004335 name = "f32_gemm_test",
4336 srcs = [
4337 "test/f32-gemm.cc",
4338 "test/gemm-microkernel-tester.h",
4339 "src/xnnpack/AlignedAllocator.h",
4340 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004341 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004342)
4343
4344xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004345 name = "f32_gemm_relu_test",
4346 srcs = [
4347 "test/f32-gemm-relu.cc",
4348 "test/gemm-microkernel-tester.h",
4349 "src/xnnpack/AlignedAllocator.h",
4350 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004351 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004352)
4353
4354xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004355 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004356 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004357 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004358 "test/gemm-microkernel-tester.h",
4359 "src/xnnpack/AlignedAllocator.h",
4360 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004361 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004362)
4363
4364xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004365 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004366 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004367 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004368 "test/gemm-microkernel-tester.h",
4369 "src/xnnpack/AlignedAllocator.h",
4370 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004371 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004372)
4373
4374xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004375 name = "f16_hswish_test",
4376 srcs = [
4377 "test/f16-hswish.cc",
4378 "test/hswish-microkernel-tester.h",
4379 ] + MICROKERNEL_TEST_HDRS,
4380 deps = MICROKERNEL_TEST_DEPS,
4381)
4382
4383xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004384 name = "f32_hswish_test",
4385 srcs = [
4386 "test/f32-hswish.cc",
4387 "test/hswish-microkernel-tester.h",
4388 ] + MICROKERNEL_TEST_HDRS,
4389 deps = MICROKERNEL_TEST_DEPS,
4390)
4391
4392xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004393 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004394 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004395 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004396 "test/maxpool-microkernel-tester.h",
4397 ] + MICROKERNEL_TEST_HDRS,
4398 deps = MICROKERNEL_TEST_DEPS,
4399)
4400
4401xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004402 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004403 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004404 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004405 "test/avgpool-microkernel-tester.h",
4406 "src/xnnpack/AlignedAllocator.h",
4407 ] + MICROKERNEL_TEST_HDRS,
4408 deps = MICROKERNEL_TEST_DEPS,
4409)
4410
4411xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004412 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004413 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004414 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004415 "test/gemm-microkernel-tester.h",
4416 "src/xnnpack/AlignedAllocator.h",
4417 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004418 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004419)
4420
4421xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004422 name = "f16_prelu_test",
4423 srcs = [
4424 "test/f16-prelu.cc",
4425 "test/prelu-microkernel-tester.h",
4426 "src/xnnpack/AlignedAllocator.h",
4427 ] + MICROKERNEL_TEST_HDRS,
4428 deps = MICROKERNEL_TEST_DEPS,
4429)
4430
4431xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004432 name = "f32_prelu_test",
4433 srcs = [
4434 "test/f32-prelu.cc",
4435 "test/prelu-microkernel-tester.h",
4436 "src/xnnpack/AlignedAllocator.h",
4437 ] + MICROKERNEL_TEST_HDRS,
4438 deps = MICROKERNEL_TEST_DEPS,
4439)
4440
4441xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004442 name = "f32_raddexpminusmax_test",
4443 srcs = [
4444 "test/f32-raddexpminusmax.cc",
4445 "test/raddexpminusmax-microkernel-tester.h",
4446 ] + MICROKERNEL_TEST_HDRS,
4447 deps = MICROKERNEL_TEST_DEPS,
4448)
4449
4450xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004451 name = "f32_raddextexp_test",
4452 srcs = [
4453 "test/f32-raddextexp.cc",
4454 "test/raddextexp-microkernel-tester.h",
4455 ] + MICROKERNEL_TEST_HDRS,
4456 deps = MICROKERNEL_TEST_DEPS,
4457)
4458
4459xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004460 name = "f32_raddstoreexpminusmax_test",
4461 srcs = [
4462 "test/f32-raddstoreexpminusmax.cc",
4463 "test/raddstoreexpminusmax-microkernel-tester.h",
4464 ] + MICROKERNEL_TEST_HDRS,
4465 deps = MICROKERNEL_TEST_DEPS,
4466)
4467
4468xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004469 name = "f32_relu_test",
4470 srcs = [
4471 "test/f32-relu.cc",
4472 "test/vunary-microkernel-tester.h",
4473 ] + MICROKERNEL_TEST_HDRS,
4474 deps = MICROKERNEL_TEST_DEPS,
4475)
4476
4477xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004478 name = "f32_rmax_test",
4479 srcs = [
4480 "test/f32-rmax.cc",
4481 "test/rmax-microkernel-tester.h",
4482 ] + MICROKERNEL_TEST_HDRS,
4483 deps = MICROKERNEL_TEST_DEPS,
4484)
4485
4486xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004487 name = "f32_sigmoid_test",
4488 srcs = [
4489 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004490 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004491 ] + MICROKERNEL_TEST_HDRS,
4492 deps = MICROKERNEL_TEST_DEPS,
4493)
4494
4495xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004496 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004497 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004498 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004499 "test/spmm-microkernel-tester.h",
4500 "src/xnnpack/AlignedAllocator.h",
4501 ] + MICROKERNEL_TEST_HDRS,
4502 deps = MICROKERNEL_TEST_DEPS,
4503)
4504
4505xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004506 name = "f16_vadd_minmax_test",
4507 srcs = [
4508 "test/f16-vadd-minmax.cc",
4509 "test/vbinary-microkernel-tester.h",
4510 ] + MICROKERNEL_TEST_HDRS,
4511 deps = MICROKERNEL_TEST_DEPS,
4512)
4513
4514xnnpack_unit_test(
4515 name = "f16_vaddc_minmax_test",
4516 srcs = [
4517 "test/f16-vaddc-minmax.cc",
4518 "test/vbinaryc-microkernel-tester.h",
4519 ] + MICROKERNEL_TEST_HDRS,
4520 deps = MICROKERNEL_TEST_DEPS,
4521)
4522
4523xnnpack_unit_test(
4524 name = "f16_vdiv_minmax_test",
4525 srcs = [
4526 "test/f16-vdiv-minmax.cc",
4527 "test/vbinary-microkernel-tester.h",
4528 ] + MICROKERNEL_TEST_HDRS,
4529 deps = MICROKERNEL_TEST_DEPS,
4530)
4531
4532xnnpack_unit_test(
4533 name = "f16_vdivc_minmax_test",
4534 srcs = [
4535 "test/f16-vdivc-minmax.cc",
4536 "test/vbinaryc-microkernel-tester.h",
4537 ] + MICROKERNEL_TEST_HDRS,
4538 deps = MICROKERNEL_TEST_DEPS,
4539)
4540
4541xnnpack_unit_test(
4542 name = "f16_vrdivc_minmax_test",
4543 srcs = [
4544 "test/f16-vrdivc-minmax.cc",
4545 "test/vbinaryc-microkernel-tester.h",
4546 ] + MICROKERNEL_TEST_HDRS,
4547 deps = MICROKERNEL_TEST_DEPS,
4548)
4549
4550xnnpack_unit_test(
4551 name = "f16_vmax_test",
4552 srcs = [
4553 "test/f16-vmax.cc",
4554 "test/vbinary-microkernel-tester.h",
4555 ] + MICROKERNEL_TEST_HDRS,
4556 deps = MICROKERNEL_TEST_DEPS,
4557)
4558
4559xnnpack_unit_test(
4560 name = "f16_vmaxc_test",
4561 srcs = [
4562 "test/f16-vmaxc.cc",
4563 "test/vbinaryc-microkernel-tester.h",
4564 ] + MICROKERNEL_TEST_HDRS,
4565 deps = MICROKERNEL_TEST_DEPS,
4566)
4567
4568xnnpack_unit_test(
4569 name = "f16_vmin_test",
4570 srcs = [
4571 "test/f16-vmin.cc",
4572 "test/vbinary-microkernel-tester.h",
4573 ] + MICROKERNEL_TEST_HDRS,
4574 deps = MICROKERNEL_TEST_DEPS,
4575)
4576
4577xnnpack_unit_test(
4578 name = "f16_vminc_test",
4579 srcs = [
4580 "test/f16-vminc.cc",
4581 "test/vbinaryc-microkernel-tester.h",
4582 ] + MICROKERNEL_TEST_HDRS,
4583 deps = MICROKERNEL_TEST_DEPS,
4584)
4585
4586xnnpack_unit_test(
4587 name = "f16_vmul_minmax_test",
4588 srcs = [
4589 "test/f16-vmul-minmax.cc",
4590 "test/vbinary-microkernel-tester.h",
4591 ] + MICROKERNEL_TEST_HDRS,
4592 deps = MICROKERNEL_TEST_DEPS,
4593)
4594
4595xnnpack_unit_test(
4596 name = "f16_vmulc_minmax_test",
4597 srcs = [
4598 "test/f16-vmulc-minmax.cc",
4599 "test/vbinaryc-microkernel-tester.h",
4600 ] + MICROKERNEL_TEST_HDRS,
4601 deps = MICROKERNEL_TEST_DEPS,
4602)
4603
4604xnnpack_unit_test(
4605 name = "f16_vsub_minmax_test",
4606 srcs = [
4607 "test/f16-vsub-minmax.cc",
4608 "test/vbinary-microkernel-tester.h",
4609 ] + MICROKERNEL_TEST_HDRS,
4610 deps = MICROKERNEL_TEST_DEPS,
4611)
4612
4613xnnpack_unit_test(
4614 name = "f16_vsubc_minmax_test",
4615 srcs = [
4616 "test/f16-vsubc-minmax.cc",
4617 "test/vbinaryc-microkernel-tester.h",
4618 ] + MICROKERNEL_TEST_HDRS,
4619 deps = MICROKERNEL_TEST_DEPS,
4620)
4621
4622xnnpack_unit_test(
4623 name = "f16_vrsubc_minmax_test",
4624 srcs = [
4625 "test/f16-vrsubc-minmax.cc",
4626 "test/vbinaryc-microkernel-tester.h",
4627 ] + MICROKERNEL_TEST_HDRS,
4628 deps = MICROKERNEL_TEST_DEPS,
4629)
4630
4631xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004632 name = "f32_vabs_test",
4633 srcs = [
4634 "test/f32-vabs.cc",
4635 "test/vunary-microkernel-tester.h",
4636 ] + MICROKERNEL_TEST_HDRS,
4637 deps = MICROKERNEL_TEST_DEPS,
4638)
4639
4640xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004641 name = "f32_vadd_test",
4642 srcs = [
4643 "test/f32-vadd.cc",
4644 "test/vbinary-microkernel-tester.h",
4645 ] + MICROKERNEL_TEST_HDRS,
4646 deps = MICROKERNEL_TEST_DEPS,
4647)
4648
4649xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004650 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004651 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004652 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004653 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004654 ] + MICROKERNEL_TEST_HDRS,
4655 deps = MICROKERNEL_TEST_DEPS,
4656)
4657
4658xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004659 name = "f32_vaddc_test",
4660 srcs = [
4661 "test/f32-vaddc.cc",
4662 "test/vbinaryc-microkernel-tester.h",
4663 ] + MICROKERNEL_TEST_HDRS,
4664 deps = MICROKERNEL_TEST_DEPS,
4665)
4666
4667xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004668 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004669 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004670 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004671 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004672 ] + MICROKERNEL_TEST_HDRS,
4673 deps = MICROKERNEL_TEST_DEPS,
4674)
4675
4676xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004677 name = "f32_vdiv_test",
4678 srcs = [
4679 "test/f32-vdiv.cc",
4680 "test/vbinary-microkernel-tester.h",
4681 ] + MICROKERNEL_TEST_HDRS,
4682 deps = MICROKERNEL_TEST_DEPS,
4683)
4684
4685xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004686 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004687 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004688 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004689 "test/vbinary-microkernel-tester.h",
4690 ] + MICROKERNEL_TEST_HDRS,
4691 deps = MICROKERNEL_TEST_DEPS,
4692)
4693
4694xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004695 name = "f32_vdivc_test",
4696 srcs = [
4697 "test/f32-vdivc.cc",
4698 "test/vbinaryc-microkernel-tester.h",
4699 ] + MICROKERNEL_TEST_HDRS,
4700 deps = MICROKERNEL_TEST_DEPS,
4701)
4702
4703xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004704 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004705 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004706 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004707 "test/vbinaryc-microkernel-tester.h",
4708 ] + MICROKERNEL_TEST_HDRS,
4709 deps = MICROKERNEL_TEST_DEPS,
4710)
4711
4712xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004713 name = "f32_vrdivc_test",
4714 srcs = [
4715 "test/f32-vrdivc.cc",
4716 "test/vbinaryc-microkernel-tester.h",
4717 ] + MICROKERNEL_TEST_HDRS,
4718 deps = MICROKERNEL_TEST_DEPS,
4719)
4720
4721xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004722 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004723 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004724 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004725 "test/vbinaryc-microkernel-tester.h",
4726 ] + MICROKERNEL_TEST_HDRS,
4727 deps = MICROKERNEL_TEST_DEPS,
4728)
4729
4730xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004731 name = "f32_vmax_test",
4732 srcs = [
4733 "test/f32-vmax.cc",
4734 "test/vbinary-microkernel-tester.h",
4735 ] + MICROKERNEL_TEST_HDRS,
4736 deps = MICROKERNEL_TEST_DEPS,
4737)
4738
4739xnnpack_unit_test(
4740 name = "f32_vmaxc_test",
4741 srcs = [
4742 "test/f32-vmaxc.cc",
4743 "test/vbinaryc-microkernel-tester.h",
4744 ] + MICROKERNEL_TEST_HDRS,
4745 deps = MICROKERNEL_TEST_DEPS,
4746)
4747
4748xnnpack_unit_test(
4749 name = "f32_vmin_test",
4750 srcs = [
4751 "test/f32-vmin.cc",
4752 "test/vbinary-microkernel-tester.h",
4753 ] + MICROKERNEL_TEST_HDRS,
4754 deps = MICROKERNEL_TEST_DEPS,
4755)
4756
4757xnnpack_unit_test(
4758 name = "f32_vminc_test",
4759 srcs = [
4760 "test/f32-vminc.cc",
4761 "test/vbinaryc-microkernel-tester.h",
4762 ] + MICROKERNEL_TEST_HDRS,
4763 deps = MICROKERNEL_TEST_DEPS,
4764)
4765
4766xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004767 name = "f32_vmul_test",
4768 srcs = [
4769 "test/f32-vmul.cc",
4770 "test/vbinary-microkernel-tester.h",
4771 ] + MICROKERNEL_TEST_HDRS,
4772 deps = MICROKERNEL_TEST_DEPS,
4773)
4774
4775xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004776 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004777 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004778 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004779 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004780 ] + MICROKERNEL_TEST_HDRS,
4781 deps = MICROKERNEL_TEST_DEPS,
4782)
4783
4784xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004785 name = "f32_vmulc_test",
4786 srcs = [
4787 "test/f32-vmulc.cc",
4788 "test/vbinaryc-microkernel-tester.h",
4789 ] + MICROKERNEL_TEST_HDRS,
4790 deps = MICROKERNEL_TEST_DEPS,
4791)
4792
4793xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004794 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004795 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004796 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004797 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004798 ] + MICROKERNEL_TEST_HDRS,
4799 deps = MICROKERNEL_TEST_DEPS,
4800)
4801
4802xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004803 name = "f16_vmulcaddc_minmax_test",
4804 srcs = [
4805 "test/f16-vmulcaddc-minmax.cc",
4806 "test/vmulcaddc-microkernel-tester.h",
4807 "src/xnnpack/AlignedAllocator.h",
4808 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004809 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004810)
4811
4812xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004813 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004814 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004815 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004816 "test/vmulcaddc-microkernel-tester.h",
4817 "src/xnnpack/AlignedAllocator.h",
4818 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004819 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004820)
4821
4822xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004823 name = "f32_vlrelu_test",
4824 srcs = [
4825 "test/f32-vlrelu.cc",
4826 "test/vunary-microkernel-tester.h",
4827 ] + MICROKERNEL_TEST_HDRS,
4828 deps = MICROKERNEL_TEST_DEPS,
4829)
4830
4831xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004832 name = "f32_vneg_test",
4833 srcs = [
4834 "test/f32-vneg.cc",
4835 "test/vunary-microkernel-tester.h",
4836 ] + MICROKERNEL_TEST_HDRS,
4837 deps = MICROKERNEL_TEST_DEPS,
4838)
4839
4840xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004841 name = "f32_vrndne_test",
4842 srcs = [
4843 "test/f32-vrndne.cc",
4844 "test/vunary-microkernel-tester.h",
4845 ] + MICROKERNEL_TEST_HDRS,
4846 deps = MICROKERNEL_TEST_DEPS,
4847)
4848
4849xnnpack_unit_test(
4850 name = "f32_vrndz_test",
4851 srcs = [
4852 "test/f32-vrndz.cc",
4853 "test/vunary-microkernel-tester.h",
4854 ] + MICROKERNEL_TEST_HDRS,
4855 deps = MICROKERNEL_TEST_DEPS,
4856)
4857
4858xnnpack_unit_test(
4859 name = "f32_vrndu_test",
4860 srcs = [
4861 "test/f32-vrndu.cc",
4862 "test/vunary-microkernel-tester.h",
4863 ] + MICROKERNEL_TEST_HDRS,
4864 deps = MICROKERNEL_TEST_DEPS,
4865)
4866
4867xnnpack_unit_test(
4868 name = "f32_vrndd_test",
4869 srcs = [
4870 "test/f32-vrndd.cc",
4871 "test/vunary-microkernel-tester.h",
4872 ] + MICROKERNEL_TEST_HDRS,
4873 deps = MICROKERNEL_TEST_DEPS,
4874)
4875
4876xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004877 name = "f32_vscale_test",
4878 srcs = [
4879 "test/f32-vscale.cc",
4880 "test/vscale-microkernel-tester.h",
4881 ] + MICROKERNEL_TEST_HDRS,
4882 deps = MICROKERNEL_TEST_DEPS,
4883)
4884
4885xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004886 name = "f32_vscaleexpminusmax_test",
4887 srcs = [
4888 "test/f32-vscaleexpminusmax.cc",
4889 "test/vscaleexpminusmax-microkernel-tester.h",
4890 ] + MICROKERNEL_TEST_HDRS,
4891 deps = MICROKERNEL_TEST_DEPS,
4892)
4893
4894xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004895 name = "f32_vscaleextexp_test",
4896 srcs = [
4897 "test/f32-vscaleextexp.cc",
4898 "test/vscaleextexp-microkernel-tester.h",
4899 ] + MICROKERNEL_TEST_HDRS,
4900 deps = MICROKERNEL_TEST_DEPS,
4901)
4902
4903xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004904 name = "f32_vsqr_test",
4905 srcs = [
4906 "test/f32-vsqr.cc",
4907 "test/vunary-microkernel-tester.h",
4908 ] + MICROKERNEL_TEST_HDRS,
4909 deps = MICROKERNEL_TEST_DEPS,
4910)
4911
4912xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004913 name = "f32_vsqrdiff_test",
4914 srcs = [
4915 "test/f32-vsqrdiff.cc",
4916 "test/vbinary-microkernel-tester.h",
4917 ] + MICROKERNEL_TEST_HDRS,
4918 deps = MICROKERNEL_TEST_DEPS,
4919)
4920
4921xnnpack_unit_test(
4922 name = "f32_vsqrdiffc_test",
4923 srcs = [
4924 "test/f32-vsqrdiffc.cc",
4925 "test/vbinaryc-microkernel-tester.h",
4926 ] + MICROKERNEL_TEST_HDRS,
4927 deps = MICROKERNEL_TEST_DEPS,
4928)
4929
4930xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004931 name = "f32_vsqrt_test",
4932 srcs = [
4933 "test/f32-vsqrt.cc",
4934 "test/vunary-microkernel-tester.h",
4935 ] + MICROKERNEL_TEST_HDRS,
4936 deps = MICROKERNEL_TEST_DEPS,
4937)
4938
4939xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004940 name = "f32_vsub_test",
4941 srcs = [
4942 "test/f32-vsub.cc",
4943 "test/vbinary-microkernel-tester.h",
4944 ] + MICROKERNEL_TEST_HDRS,
4945 deps = MICROKERNEL_TEST_DEPS,
4946)
4947
4948xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004949 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07004950 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004951 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004952 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004953 ] + MICROKERNEL_TEST_HDRS,
4954 deps = MICROKERNEL_TEST_DEPS,
4955)
4956
4957xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004958 name = "f32_vsubc_test",
4959 srcs = [
4960 "test/f32-vsubc.cc",
4961 "test/vbinaryc-microkernel-tester.h",
4962 ] + MICROKERNEL_TEST_HDRS,
4963 deps = MICROKERNEL_TEST_DEPS,
4964)
4965
4966xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004967 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004968 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004969 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004970 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004971 ] + MICROKERNEL_TEST_HDRS,
4972 deps = MICROKERNEL_TEST_DEPS,
4973)
4974
4975xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004976 name = "f32_vrsubc_test",
4977 srcs = [
4978 "test/f32-vrsubc.cc",
4979 "test/vbinaryc-microkernel-tester.h",
4980 ] + MICROKERNEL_TEST_HDRS,
4981 deps = MICROKERNEL_TEST_DEPS,
4982)
4983
4984xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004985 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004986 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004987 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004988 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07004989 ] + MICROKERNEL_TEST_HDRS,
4990 deps = MICROKERNEL_TEST_DEPS,
4991)
4992
4993xnnpack_unit_test(
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07004994 name = "qs8_requantization_test",
4995 srcs = [
4996 "src/xnnpack/requantization-stubs.h",
4997 "test/qs8-requantization.cc",
4998 "test/requantization-tester.h",
4999 ] + MICROKERNEL_TEST_HDRS,
5000 deps = MICROKERNEL_TEST_DEPS,
5001)
5002
5003xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005004 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005005 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005006 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005007 "test/avgpool-microkernel-tester.h",
5008 "src/xnnpack/AlignedAllocator.h",
5009 ] + MICROKERNEL_TEST_HDRS,
5010 deps = MICROKERNEL_TEST_DEPS,
5011)
5012
5013xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005014 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005015 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005016 "test/qu8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005017 "test/gemm-microkernel-tester.h",
5018 "src/xnnpack/AlignedAllocator.h",
5019 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005020 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005021)
5022
5023xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005024 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005025 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005026 "test/qu8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005027 "test/dwconv-microkernel-tester.h",
5028 "src/xnnpack/AlignedAllocator.h",
5029 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005030 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005031)
5032
5033xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005034 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005035 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005036 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005037 "test/gavgpool-microkernel-tester.h",
5038 "src/xnnpack/AlignedAllocator.h",
5039 ] + MICROKERNEL_TEST_HDRS,
5040 deps = MICROKERNEL_TEST_DEPS,
5041)
5042
5043xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005044 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005045 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005046 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005047 "test/gemm-microkernel-tester.h",
5048 "src/xnnpack/AlignedAllocator.h",
5049 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005050 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005051)
5052
5053xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005054 name = "qu8_requantization_test",
5055 srcs = [
5056 "src/xnnpack/requantization-stubs.h",
5057 "test/qu8-requantization.cc",
5058 "test/requantization-tester.h",
5059 ] + MICROKERNEL_TEST_HDRS,
5060 deps = MICROKERNEL_TEST_DEPS,
5061)
5062
5063xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005064 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005065 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005066 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005067 "test/vadd-microkernel-tester.h",
5068 ] + MICROKERNEL_TEST_HDRS,
5069 deps = MICROKERNEL_TEST_DEPS,
5070)
5071
5072xnnpack_unit_test(
5073 name = "u8_clamp_test",
5074 srcs = [
5075 "test/u8-clamp.cc",
5076 "test/clamp-microkernel-tester.h",
5077 ] + MICROKERNEL_TEST_HDRS,
5078 deps = MICROKERNEL_TEST_DEPS,
5079)
5080
5081xnnpack_unit_test(
5082 name = "u8_lut32norm_test",
5083 srcs = [
5084 "test/u8-lut32norm.cc",
5085 "test/lut-norm-microkernel-tester.h",
5086 ] + MICROKERNEL_TEST_HDRS,
5087 deps = MICROKERNEL_TEST_DEPS,
5088)
5089
5090xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005091 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005092 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005093 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005094 "test/maxpool-microkernel-tester.h",
5095 ] + MICROKERNEL_TEST_HDRS,
5096 deps = MICROKERNEL_TEST_DEPS,
5097)
5098
5099xnnpack_unit_test(
5100 name = "u8_rmax_test",
5101 srcs = [
5102 "test/u8-rmax.cc",
5103 "test/rmax-microkernel-tester.h",
5104 ] + MICROKERNEL_TEST_HDRS,
5105 deps = MICROKERNEL_TEST_DEPS,
5106)
5107
5108xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005109 name = "x32_fill_test",
5110 srcs = [
5111 "test/x32-fill.cc",
5112 "test/fill-microkernel-tester.h",
5113 ] + MICROKERNEL_TEST_HDRS,
5114 deps = MICROKERNEL_TEST_DEPS,
5115)
5116
5117xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005118 name = "x32_packx_test",
5119 srcs = [
5120 "test/x32-packx.cc",
5121 "test/pack-microkernel-tester.h",
5122 "src/xnnpack/AlignedAllocator.h",
5123 ] + MICROKERNEL_TEST_HDRS,
5124 deps = MICROKERNEL_TEST_DEPS,
5125)
5126
5127xnnpack_unit_test(
5128 name = "x32_pad_test",
5129 srcs = [
5130 "test/x32-pad.cc",
5131 "test/pad-microkernel-tester.h",
5132 ] + MICROKERNEL_TEST_HDRS,
5133 deps = MICROKERNEL_TEST_DEPS,
5134)
5135
5136xnnpack_unit_test(
5137 name = "x32_unpool_test",
5138 srcs = [
5139 "test/x32-unpool.cc",
5140 "test/unpool-microkernel-tester.h",
5141 ] + MICROKERNEL_TEST_HDRS,
5142 deps = MICROKERNEL_TEST_DEPS,
5143)
5144
5145xnnpack_unit_test(
5146 name = "x32_zip_test",
5147 srcs = [
5148 "test/x32-zip.cc",
5149 "test/zip-microkernel-tester.h",
5150 ] + MICROKERNEL_TEST_HDRS,
5151 deps = MICROKERNEL_TEST_DEPS,
5152)
5153
5154xnnpack_unit_test(
5155 name = "x8_lut_test",
5156 srcs = [
5157 "test/x8-lut.cc",
5158 "test/lut-microkernel-tester.h",
5159 ] + MICROKERNEL_TEST_HDRS,
5160 deps = MICROKERNEL_TEST_DEPS,
5161)
5162
5163xnnpack_unit_test(
5164 name = "x8_zip_test",
5165 srcs = [
5166 "test/x8-zip.cc",
5167 "test/zip-microkernel-tester.h",
5168 ] + MICROKERNEL_TEST_HDRS,
5169 deps = MICROKERNEL_TEST_DEPS,
5170)
5171
Marat Dukhan20c3b922020-03-10 03:45:06 -07005172########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005173
5174xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005175 name = "operator_size_test",
5176 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005177 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005178)
5179
Marat Dukhan20c3b922020-03-10 03:45:06 -07005180xnnpack_binary(
5181 name = "subgraph_size_test",
5182 srcs = ["test/subgraph-size.c"],
5183 deps = [":XNNPACK"],
5184)
5185
5186########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005187
5188xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005189 name = "abs_nc_test",
5190 srcs = [
5191 "test/abs-nc.cc",
5192 "test/abs-operator-tester.h",
5193 ],
5194 deps = OPERATOR_TEST_DEPS,
5195)
5196
5197xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005198 name = "add_nd_test",
5199 srcs = [
5200 "test/add-nd.cc",
5201 "test/binary-elementwise-operator-tester.h",
5202 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005203 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005204)
5205
5206xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005207 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005208 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005209 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005210 "test/argmax-pooling-operator-tester.h",
5211 ] + OPERATOR_TEST_PARAMS_HDRS,
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 = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005217 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005218 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005219 "test/average-pooling-operator-tester.h",
5220 ] + OPERATOR_TEST_PARAMS_HDRS,
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 Dukhan64e52512020-06-09 13:41:16 -07005225 name = "bankers_rounding_nc_test",
5226 srcs = [
5227 "test/bankers-rounding-nc.cc",
5228 "test/bankers-rounding-operator-tester.h",
5229 ],
5230 deps = OPERATOR_TEST_DEPS,
5231)
5232
5233xnnpack_unit_test(
5234 name = "ceiling_nc_test",
5235 srcs = [
5236 "test/ceiling-nc.cc",
5237 "test/ceiling-operator-tester.h",
5238 ],
5239 deps = OPERATOR_TEST_DEPS,
5240)
5241
5242xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005243 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005244 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005245 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005246 "test/channel-shuffle-operator-tester.h",
5247 ],
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 Dukhanefc47b82019-11-18 09:25:38 -08005252 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005253 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005254 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005255 "test/clamp-operator-tester.h",
5256 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005257 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005258)
5259
5260xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005261 name = "constant_pad_nd_test",
5262 srcs = [
5263 "test/constant-pad-nd.cc",
5264 "test/constant-pad-operator-tester.h",
5265 ],
5266 deps = OPERATOR_TEST_DEPS,
5267)
5268
5269xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005270 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005271 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005272 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005273 "test/convolution-operator-tester.h",
5274 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005275 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005276)
5277
5278xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005279 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005280 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005281 "test/convolution-nchw.cc",
5282 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005283 ],
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 Dukhan4e21b272020-06-04 18:45:01 -07005288 name = "copy_nc_test",
5289 srcs = [
5290 "test/copy-nc.cc",
5291 "test/copy-operator-tester.h",
5292 ],
5293 deps = OPERATOR_TEST_DEPS,
5294)
5295
5296xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005297 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005298 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005299 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005300 "test/deconvolution-operator-tester.h",
5301 ] + 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 Dukhan69180502019-12-06 15:00:31 -08005306 name = "divide_nd_test",
5307 srcs = [
5308 "test/binary-elementwise-operator-tester.h",
5309 "test/divide-nd.cc",
5310 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005311 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005312)
5313
5314xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005315 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005316 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005317 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005318 "test/fully-connected-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 Dukhan64e52512020-06-09 13:41:16 -07005324 name = "floor_nc_test",
5325 srcs = [
5326 "test/floor-nc.cc",
5327 "test/floor-operator-tester.h",
5328 ],
5329 deps = OPERATOR_TEST_DEPS,
5330)
5331
5332xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005333 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005334 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005335 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005336 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005337 ] + 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 Dukhanefc47b82019-11-18 09:25:38 -08005342 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005343 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005344 "test/global-average-pooling-ncw.cc",
5345 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005346 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005347 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005348)
5349
5350xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005351 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005352 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005353 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005354 "test/hardswish-operator-tester.h",
5355 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005356 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005357)
5358
5359xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005360 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005361 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005362 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005363 "test/leaky-relu-operator-tester.h",
5364 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005365 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005366)
5367
5368xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005369 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005370 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005371 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005372 "test/max-pooling-operator-tester.h",
5373 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005374 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005375)
5376
5377xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005378 name = "maximum_nd_test",
5379 srcs = [
5380 "test/binary-elementwise-operator-tester.h",
5381 "test/maximum-nd.cc",
5382 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005383 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005384)
5385
5386xnnpack_unit_test(
5387 name = "minimum_nd_test",
5388 srcs = [
5389 "test/binary-elementwise-operator-tester.h",
5390 "test/minimum-nd.cc",
5391 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005392 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005393)
5394
5395xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005396 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005397 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005398 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005399 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005400 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005401 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005402)
5403
5404xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005405 name = "negate_nc_test",
5406 srcs = [
5407 "test/negate-nc.cc",
5408 "test/negate-operator-tester.h",
5409 ],
5410 deps = OPERATOR_TEST_DEPS,
5411)
5412
5413xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005414 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005415 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005416 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005417 "test/prelu-operator-tester.h",
5418 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005419 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005420)
5421
5422xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005423 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005424 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005425 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005426 "test/resize-bilinear-operator-tester.h",
5427 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005428 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005429)
5430
5431xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005432 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005433 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005434 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005435 "test/sigmoid-operator-tester.h",
5436 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005437 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005438)
5439
5440xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005441 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005442 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005443 "test/softmax-nc.cc",
5444 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005445 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005446 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005447)
5448
5449xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005450 name = "square_nc_test",
5451 srcs = [
5452 "test/square-nc.cc",
5453 "test/square-operator-tester.h",
5454 ],
5455 deps = OPERATOR_TEST_DEPS,
5456)
5457
5458xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005459 name = "square_root_nc_test",
5460 srcs = [
5461 "test/square-root-nc.cc",
5462 "test/square-root-operator-tester.h",
5463 ],
5464 deps = OPERATOR_TEST_DEPS,
5465)
5466
5467xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005468 name = "squared_difference_nd_test",
5469 srcs = [
5470 "test/binary-elementwise-operator-tester.h",
5471 "test/squared-difference-nd.cc",
5472 ],
5473 deps = OPERATOR_TEST_DEPS,
5474)
5475
5476xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005477 name = "subtract_nd_test",
5478 srcs = [
5479 "test/binary-elementwise-operator-tester.h",
5480 "test/subtract-nd.cc",
5481 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005482 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005483)
5484
5485xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005486 name = "truncation_nc_test",
5487 srcs = [
5488 "test/truncation-nc.cc",
5489 "test/truncation-operator-tester.h",
5490 ],
5491 deps = OPERATOR_TEST_DEPS,
5492)
5493
5494xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005495 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005496 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005497 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005498 "test/unpooling-operator-tester.h",
5499 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005500 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005501)
5502
Chao Mei6ddfc602020-05-13 22:29:36 -07005503############################### Misc unit tests ###############################
5504
5505xnnpack_unit_test(
5506 name = "memory_planner_test",
5507 srcs = [
5508 "test/memory-planner-test.cc",
5509 ],
5510 deps = [
5511 ":XNNPACK",
5512 ":memory_planner",
5513 ],
5514)
5515
Marat Dukhan08c4a432019-10-03 09:29:21 -07005516############################# Build configurations #############################
5517
Marat Dukhanb8642352019-10-30 15:43:02 -07005518# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005519config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005520 name = "xnn_enable_assembly_explicit_true",
5521 define_values = {"xnn_enable_assembly": "true"},
5522)
5523
5524# Disables usage of assembly kernels.
5525config_setting(
5526 name = "xnn_enable_assembly_explicit_false",
5527 define_values = {"xnn_enable_assembly": "false"},
5528)
5529
Marat Dukhan9de90e02020-06-18 16:04:12 -07005530# Enables usage of sparse inference.
5531config_setting(
5532 name = "xnn_enable_sparse_explicit_true",
5533 define_values = {"xnn_enable_sparse": "true"},
5534)
5535
5536# Disables usage of sparse inference.
5537config_setting(
5538 name = "xnn_enable_sparse_explicit_false",
5539 define_values = {"xnn_enable_sparse": "false"},
5540)
5541
Marat Dukhan05702cf2020-03-26 15:41:33 -07005542# Disables usage of HMP-aware optimizations.
5543config_setting(
5544 name = "xnn_enable_hmp_explicit_false",
5545 define_values = {"xnn_enable_hmp": "false"},
5546)
5547
Chao Mei6ddfc602020-05-13 22:29:36 -07005548# Enable usage of optimized memory allocation
5549config_setting(
5550 name = "xnn_enable_memopt_explicit_true",
5551 define_values = {"xnnpack_enable_memopt": "true"},
5552)
5553
5554# Disable usage of optimized memory allocation
5555config_setting(
5556 name = "xnn_enable_memopt_explicit_false",
5557 define_values = {"xnnpack_enable_memopt": "false"},
5558)
5559
Marat Dukhanb8642352019-10-30 15:43:02 -07005560# Builds with -c dbg
5561config_setting(
5562 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005563 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005564 "compilation_mode": "dbg",
5565 },
5566)
5567
5568# Builds with -c opt
5569config_setting(
5570 name = "optimized_build",
5571 values = {
5572 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005573 },
5574)
5575
5576config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005577 name = "linux_k8",
5578 values = {"cpu": "k8"},
5579)
5580
5581config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005582 name = "linux_arm",
5583 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005584)
5585
5586config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005587 name = "linux_armeabi",
5588 values = {"cpu": "armeabi"},
5589)
5590
5591config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005592 name = "linux_armhf",
5593 values = {"cpu": "armhf"},
5594)
5595
5596config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005597 name = "linux_armv7a",
5598 values = {"cpu": "armv7a"},
5599)
5600
5601config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005602 name = "linux_aarch64",
5603 values = {"cpu": "aarch64"},
5604)
5605
5606config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005607 name = "android",
5608 values = {"crosstool_top": "//external:android/crosstool"},
5609)
5610
5611config_setting(
5612 name = "android_armv7",
5613 values = {
5614 "crosstool_top": "//external:android/crosstool",
5615 "cpu": "armeabi-v7a",
5616 },
5617)
5618
5619config_setting(
5620 name = "android_arm64",
5621 values = {
5622 "crosstool_top": "//external:android/crosstool",
5623 "cpu": "arm64-v8a",
5624 },
5625)
5626
5627config_setting(
5628 name = "android_x86",
5629 values = {
5630 "crosstool_top": "//external:android/crosstool",
5631 "cpu": "x86",
5632 },
5633)
5634
5635config_setting(
5636 name = "android_x86_64",
5637 values = {
5638 "crosstool_top": "//external:android/crosstool",
5639 "cpu": "x86_64",
5640 },
5641)
5642
5643config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005644 name = "windows_x86_64",
5645 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005646)
5647
5648config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005649 name = "windows_x86_64_clang",
5650 values = {
5651 "compiler": "clang-cl",
5652 "cpu": "x64_windows",
5653 },
5654)
5655
5656config_setting(
5657 name = "windows_x86_64_mingw",
5658 values = {
5659 "compiler": "mingw-gcc",
5660 "cpu": "x64_windows",
5661 },
5662)
5663
5664config_setting(
5665 name = "windows_x86_64_msys",
5666 values = {
5667 "compiler": "msys-gcc",
5668 "cpu": "x64_windows",
5669 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005670)
5671
5672config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005673 name = "macos_x86_64",
5674 values = {
5675 "apple_platform_type": "macos",
5676 "cpu": "darwin",
5677 },
5678)
5679
5680config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005681 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005682 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005683)
5684
5685config_setting(
5686 name = "emscripten_wasm",
5687 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005688 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005689 "cpu": "wasm",
5690 },
5691)
5692
5693config_setting(
5694 name = "emscripten_wasmsimd",
5695 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005696 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005697 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005698 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005699 },
5700)
5701
5702config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005703 name = "ios_armv7",
5704 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005705 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005706 "cpu": "ios_armv7",
5707 },
5708)
5709
5710config_setting(
5711 name = "ios_arm64",
5712 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005713 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005714 "cpu": "ios_arm64",
5715 },
5716)
5717
5718config_setting(
5719 name = "ios_arm64e",
5720 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005721 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005722 "cpu": "ios_arm64e",
5723 },
5724)
5725
5726config_setting(
5727 name = "ios_x86",
5728 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005729 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005730 "cpu": "ios_i386",
5731 },
5732)
5733
5734config_setting(
5735 name = "ios_x86_64",
5736 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005737 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005738 "cpu": "ios_x86_64",
5739 },
5740)
5741
5742config_setting(
5743 name = "watchos_armv7k",
5744 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005745 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005746 "cpu": "watchos_armv7k",
5747 },
5748)
5749
5750config_setting(
5751 name = "watchos_arm64_32",
5752 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005753 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005754 "cpu": "watchos_arm64_32",
5755 },
5756)
5757
5758config_setting(
5759 name = "watchos_x86",
5760 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005761 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005762 "cpu": "watchos_i386",
5763 },
5764)
5765
5766config_setting(
5767 name = "watchos_x86_64",
5768 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005769 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005770 "cpu": "watchos_x86_64",
5771 },
5772)
5773
5774config_setting(
5775 name = "tvos_arm64",
5776 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005777 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005778 "cpu": "tvos_arm64",
5779 },
5780)
5781
5782config_setting(
5783 name = "tvos_x86_64",
5784 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005785 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005786 "cpu": "tvos_x86_64",
5787 },
5788)