blob: 5a1a535b307457bf2531f7c268bdaf12695a6679 [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 Dukhancc8f34c2020-08-05 16:36:38 -0700851 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
852 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
853 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhan27203da2020-08-05 15:19:03 -0700854 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
855 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
856 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
857 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
858 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
859 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
860 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
861 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
862 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
863 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
864 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
865 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
866 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
867 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
868 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700869 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -0700870 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700871 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -0700872 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700873 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700874 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700875 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700876 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700877 "src/x32-zip/x2-wasmsimd.c",
878 "src/x32-zip/x3-wasmsimd.c",
879 "src/x32-zip/x4-wasmsimd.c",
880 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700881 "src/math/roundne-wasmsimd-addsub.c",
882 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700883 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700884 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700885 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700886 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700887 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -0700888 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
889 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700890]
891
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800892PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -0700893 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700894 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700895 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700896 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700897 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700898 "src/f32-spmm/gen/4x1-minmax-psimd.c",
899 "src/f32-spmm/gen/8x1-minmax-psimd.c",
900 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700901]
902
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800903PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700904 "src/qs8-requantization/precise-psimd.c",
905 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700906 "src/qu8-requantization/precise-psimd.c",
907 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800908]
909
Marat Dukhan08c4a432019-10-03 09:29:21 -0700910# ISA-specific micro-kernels
911NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -0700912 "src/f32-argmaxpool/4x-neon-c4.c",
913 "src/f32-argmaxpool/9p8x-neon-c4.c",
914 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700915 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
916 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700917 "src/f32-clamp/gen/neon-x4.c",
918 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700919 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
920 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700921 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
922 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700923 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
924 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700925 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
926 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700927 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
928 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
929 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
930 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700931 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
932 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
933 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
934 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700935 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
936 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
937 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
938 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700939 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700940 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
941 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700942 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
943 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
944 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
945 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
946 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
947 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
948 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
949 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
950 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
951 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
952 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
953 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
954 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
955 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
956 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
957 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
958 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
959 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
960 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
961 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
962 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
963 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
964 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
965 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
966 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
967 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
968 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
969 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
970 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
971 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
972 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700973 "src/f32-hswish/gen/hswish-neon-x4.c",
974 "src/f32-hswish/gen/hswish-neon-x8.c",
975 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700976 "src/f32-ibilinear/gen/neon-c4.c",
977 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700978 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
979 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
980 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
981 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
982 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
983 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
984 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
985 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
986 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
987 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
988 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
989 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
990 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
991 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
992 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
993 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700994 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
995 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
996 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700997 "src/f32-ppmm/gen/4x8-minmax-neon.c",
998 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700999 "src/f32-prelu/gen/neon-1x4.c",
1000 "src/f32-prelu/gen/neon-1x8.c",
1001 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001002 "src/f32-prelu/gen/neon-2x4.c",
1003 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001004 "src/f32-prelu/gen/neon-2x16.c",
1005 "src/f32-prelu/gen/neon-4x4.c",
1006 "src/f32-prelu/gen/neon-4x8.c",
1007 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001008 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1009 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1010 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1011 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1012 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1013 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1014 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1015 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1016 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1017 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1018 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1019 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1020 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1021 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1022 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1023 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1024 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1025 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1026 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1027 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1028 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1029 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1030 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1031 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001032 "src/f32-relu/gen/neon-x4.c",
1033 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001034 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001035 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001036 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1037 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1038 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1039 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1040 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1041 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1042 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1043 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1044 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1045 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1046 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1047 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1048 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1049 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1050 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1051 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1052 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1053 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001054 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1055 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1056 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1057 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001058 "src/f32-vbinary/gen/vmax-neon-x4.c",
1059 "src/f32-vbinary/gen/vmax-neon-x8.c",
1060 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1061 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1062 "src/f32-vbinary/gen/vmin-neon-x4.c",
1063 "src/f32-vbinary/gen/vmin-neon-x8.c",
1064 "src/f32-vbinary/gen/vminc-neon-x4.c",
1065 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001066 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1067 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1068 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1069 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1070 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1071 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001072 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1073 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1074 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1075 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001076 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1077 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1078 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1079 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001080 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1081 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001082 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1083 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001084 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1085 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1086 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1087 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1088 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1089 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1090 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1091 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001092 "src/f32-vunary/gen/vabs-neon-x4.c",
1093 "src/f32-vunary/gen/vabs-neon-x8.c",
1094 "src/f32-vunary/gen/vneg-neon-x4.c",
1095 "src/f32-vunary/gen/vneg-neon-x8.c",
1096 "src/f32-vunary/gen/vsqr-neon-x4.c",
1097 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001098 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
1099 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
1100 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1101 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1102 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
1103 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1104 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1105 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001106 "src/qs8-requantization/precise-neon.c",
1107 "src/qs8-requantization/fp32-neon.c",
1108 "src/qs8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001109 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1110 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1111 "src/qu8-dwconv/up8x9-minmax-neon.c",
1112 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1113 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1114 "src/qu8-gemm/4x8-minmax-neon.c",
1115 "src/qu8-gemm/8x8-minmax-neon.c",
1116 "src/qu8-igemm/4x8-minmax-neon.c",
1117 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001118 "src/qu8-requantization/precise-neon.c",
1119 "src/qu8-requantization/fp32-neon.c",
1120 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001121 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001122 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001123 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001124 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001125 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001126 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001127 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001128 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001129 "src/x32-zip/x2-neon.c",
1130 "src/x32-zip/x3-neon.c",
1131 "src/x32-zip/x4-neon.c",
1132 "src/x32-zip/xm-neon.c",
1133 "src/x8-zip/x2-neon.c",
1134 "src/x8-zip/x3-neon.c",
1135 "src/x8-zip/x4-neon.c",
1136 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001137 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001138 "src/math/roundd-neon-addsub.c",
1139 "src/math/roundd-neon-cvt.c",
1140 "src/math/roundu-neon-addsub.c",
1141 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001142 "src/math/roundz-neon-addsub.c",
1143 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001144 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001145 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1146 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1147 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1148 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1149 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1150 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001151 "src/math/sqrt-neon-nr1rsqrts.c",
1152 "src/math/sqrt-neon-nr2rsqrts.c",
1153 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001154]
1155
1156NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001157 "src/f32-ibilinear/gen/neonfma-c4.c",
1158 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001159 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1160 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1161 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1162 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1163 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1164 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1165 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1166 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1167 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001168 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1169 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1170 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1171 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001172 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1173 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1174 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1175 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001176 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1177 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1178 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1179 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001180 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1181 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1182 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1183 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1184 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1185 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1186 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1187 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1188 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1189 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1190 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1191 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1192 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1193 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1194 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1195 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1196 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1197 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001198 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1199 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001200 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1201 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1202 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1203 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1204 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1205 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1206 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1207 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1208 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1209 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1210 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1211 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1212 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1213 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1214 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1215 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1216 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1217 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1218 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1219 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1220 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1221 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1222 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1223 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001224 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1225 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1226 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1227 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1228 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1229 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1230 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1231 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1232 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1233 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1234 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1235 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1236 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1237 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1238 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1239 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1240 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1241 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1242 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1243 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1244 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1245 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1246 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1247 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1248 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1249 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1250 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1251 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1252 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1253 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1254 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1255 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1256 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1257 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1258 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1259 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1260 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1261 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1262 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1263 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1264 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1265 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1266 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1267 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1268 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1269 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1270 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1271 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1272 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1273 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1274 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1275 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1276 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1277 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001278 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1279 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001280 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1281 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1282 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1283 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1284 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1285 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1286 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1287 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1288 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1289 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1290 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1291 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1292 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1293 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1294 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1295 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1296 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1297 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1298 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1299 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001300 "src/math/exp-neonfma-lut64-p2.c",
1301 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001302 "src/math/expminus-neonfma-lut2048-p1.c",
1303 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001304 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001305 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1306 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1307 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1308 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1309 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1310 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1311 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1312 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1313 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1314 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1315 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1316 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1317 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1318 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1319 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1320 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1321 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1322 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001323 "src/math/sqrt-neonfma-nr1fma.c",
1324 "src/math/sqrt-neonfma-nr2fma.c",
1325 "src/math/sqrt-neonfma-nr3fma.c",
1326 "src/math/sqrt-neonfma-nr2fma1adj.c",
1327 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001328]
1329
1330AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001331 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1332 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1333 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1334 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1335 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1336 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001337 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1338 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1339 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1340 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1341 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1342 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1343 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1344 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1345 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1346 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1347 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1348 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1349 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1350 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1351 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1352 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1353 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1354 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1355 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1356 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001357 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1358 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001359 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1360 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001361 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1362 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001363 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1364 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001365 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1366 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1367 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1368 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1369 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001370 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1371 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1372 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1373 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1374 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1375 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1376 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1377 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1378 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1379 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1380 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1381 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1382 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1383 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1384 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1385 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1386 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1387 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001388 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1389 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1390 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1391 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1392 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1393 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1394 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1395 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1396 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1397 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1398 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1399 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1400 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1401 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1402 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1403 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1404 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1405 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001406 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1407 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001408 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1409 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1410 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1411 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1412 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1413 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001414]
1415
Marat Dukhan8853b822020-05-07 12:19:01 -07001416NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001417 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1418 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1419 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1420 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1421 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1422 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1423 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1424 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001425 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001426 "src/math/roundd-neonv8.c",
1427 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001428 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001429]
1430
Marat Dukhan08c4a432019-10-03 09:29:21 -07001431AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001432 "src/f16-clamp/gen/neonfp16arith-x16.c",
1433 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001434 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1435 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1436 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1437 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1438 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1439 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1440 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1441 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1442 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1443 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1444 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1445 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001446 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1447 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001448 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1449 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001450 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001451 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001452 "src/f16-relu/gen/neonfp16arith-x16.c",
1453 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001454 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1455 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1456 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1457 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1458 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1459 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1460 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1461 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1462 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1463 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1464 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1465 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1466 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1467 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1468 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1469 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1470 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1471 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1472 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1473 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1474 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1475 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1476 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1477 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1478 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1479 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1480 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1481 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001482 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1483 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001484 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1485 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1486 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1487 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1488 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1489 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1490 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1491 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001492 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1493 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1494 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1495 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001496 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1497 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1498 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1499 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1500 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1501 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1502 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1503 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1504 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1505 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1506 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1507 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001508 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1509 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1510 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1511 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1512 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1513 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1514 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1515 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001516]
1517
1518SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001519 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1520 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001521 "src/f32-clamp/gen/sse-x4.c",
1522 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001523 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1524 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001525 "src/f32-dwconv-chw/3x3p1-sse.c",
1526 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001527 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1528 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1529 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1530 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1531 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1532 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1533 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1534 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1535 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1536 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1537 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1538 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001539 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001540 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1541 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001542 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1543 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1544 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1545 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1546 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1547 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1548 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1549 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1550 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1551 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1552 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1553 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1554 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001555 "src/f32-hswish/gen/hswish-sse-x4.c",
1556 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001557 "src/f32-ibilinear/gen/sse-c4.c",
1558 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001559 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1560 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1561 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1562 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1563 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1564 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1565 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001566 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1567 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1568 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001569 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001570 "src/f32-prelu/gen/sse-2x4.c",
1571 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001572 "src/f32-relu/gen/sse-x4.c",
1573 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001574 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001575 "src/f32-spmm/gen/4x1-minmax-sse.c",
1576 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001577 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001578 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1579 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1580 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1581 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1582 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1583 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1584 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1585 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001586 "src/f32-vbinary/gen/vmax-sse-x4.c",
1587 "src/f32-vbinary/gen/vmax-sse-x8.c",
1588 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1589 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1590 "src/f32-vbinary/gen/vmin-sse-x4.c",
1591 "src/f32-vbinary/gen/vmin-sse-x8.c",
1592 "src/f32-vbinary/gen/vminc-sse-x4.c",
1593 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001594 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1595 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1596 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1597 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1598 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1599 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1600 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1601 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001602 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1603 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1604 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1605 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001606 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1607 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1608 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1609 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001610 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1611 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001612 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1613 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001614 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1615 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001616 "src/f32-vunary/gen/vabs-sse-x4.c",
1617 "src/f32-vunary/gen/vabs-sse-x8.c",
1618 "src/f32-vunary/gen/vneg-sse-x4.c",
1619 "src/f32-vunary/gen/vneg-sse-x8.c",
1620 "src/f32-vunary/gen/vsqr-sse-x4.c",
1621 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001622 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001623 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001624 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001625 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001626 "src/math/roundd-sse-addsub.c",
1627 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001628 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001629 "src/math/sqrt-sse-hh1mac.c",
1630 "src/math/sqrt-sse-nr1mac.c",
1631 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001632]
1633
1634SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001635 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1636 "src/f32-argmaxpool/4x-sse2-c4.c",
1637 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001638 "src/f32-prelu/gen/sse2-2x4.c",
1639 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001640 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1641 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1642 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1643 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1644 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1645 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1646 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1647 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1648 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1649 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1650 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1651 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001652 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001653 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001654 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001655 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001656 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1657 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001658 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1659 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001660 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1661 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1662 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1663 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1664 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1665 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1666 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1667 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001668 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1669 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1670 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001671 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1672 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001673 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1674 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001675 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1676 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001677 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1678 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001679 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001680 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1681 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001682 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001683 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1684 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1685 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001686 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1687 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001688 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1689 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001690 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1691 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001692 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001693 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1694 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001695 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001696 "src/qs8-requantization/precise-sse2.c",
1697 "src/qs8-requantization/fp32-sse2.c",
1698 "src/qs8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001699 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1700 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1701 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1702 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1703 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1704 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1705 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1706 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001707 "src/qu8-requantization/precise-sse2.c",
1708 "src/qu8-requantization/fp32-sse2.c",
1709 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001710 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001711 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001712 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001713 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001714 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001715 "src/x32-zip/x2-sse2.c",
1716 "src/x32-zip/x3-sse2.c",
1717 "src/x32-zip/x4-sse2.c",
1718 "src/x32-zip/xm-sse2.c",
1719 "src/x8-zip/x2-sse2.c",
1720 "src/x8-zip/x3-sse2.c",
1721 "src/x8-zip/x4-sse2.c",
1722 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001723 "src/math/exp-sse2-p5.c",
1724 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001725 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001726 "src/math/roundd-sse2-cvt.c",
1727 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001728 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001729 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001730]
1731
1732SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001733 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
1734 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
1735 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001736 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1737 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001738 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1739 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001740 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
1741 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001742 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
1743 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001744 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001745 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
1746 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001747 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001748 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
1749 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
1750 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001751 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
1752 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001753 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
1754 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001755 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
1756 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001757 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001758 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
1759 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001760 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001761 "src/qs8-requantization/precise-ssse3.c",
1762 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001763 "src/qu8-requantization/precise-ssse3.c",
1764 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001765]
1766
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001767SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001768 "src/f32-prelu/gen/sse41-2x4.c",
1769 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001770 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001771 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001772 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001773 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001774 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1775 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001776 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1777 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001778 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1779 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1780 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1781 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1782 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1783 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1784 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1785 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001786 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
1787 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
1788 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001789 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
1790 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001791 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
1792 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001793 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
1794 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001795 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
1796 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001797 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001798 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
1799 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001800 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001801 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
1802 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
1803 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001804 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
1805 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001806 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
1807 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001808 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
1809 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001810 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001811 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
1812 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001813 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001814 "src/qs8-requantization/fp32-sse4.c",
1815 "src/qs8-requantization/precise-sse4.c",
1816 "src/qs8-requantization/q31-sse4.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001817 "src/qu8-requantization/precise-sse4.c",
1818 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001819 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001820 "src/math/roundd-sse41.c",
1821 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001822 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001823]
1824
Marat Dukhan08c4a432019-10-03 09:29:21 -07001825AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001826 "src/f32-clamp/gen/avx-x8.c",
1827 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001828 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1829 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1830 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1831 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1832 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1833 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1834 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1835 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1836 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1837 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1838 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1839 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1840 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1841 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1842 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1843 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1844 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1845 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1846 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1847 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1848 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1849 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1850 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1851 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1852 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1853 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1854 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1855 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1856 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1857 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001858 "src/f32-hswish/gen/hswish-avx-x8.c",
1859 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001860 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1861 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1862 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1863 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1864 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1865 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1866 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1867 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1868 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001869 "src/f32-prelu/gen/avx-2x8.c",
1870 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001871 "src/f32-relu/gen/avx-x16.c",
1872 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001873 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001874 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1875 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1876 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1877 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1878 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1879 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1880 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1881 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001882 "src/f32-vbinary/gen/vmax-avx-x8.c",
1883 "src/f32-vbinary/gen/vmax-avx-x16.c",
1884 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1885 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1886 "src/f32-vbinary/gen/vmin-avx-x8.c",
1887 "src/f32-vbinary/gen/vmin-avx-x16.c",
1888 "src/f32-vbinary/gen/vminc-avx-x8.c",
1889 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001890 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1891 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1892 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1893 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1894 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1895 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1896 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1897 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001898 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1899 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1900 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1901 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001902 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1903 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1904 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1905 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001906 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1907 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001908 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001909 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1910 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1911 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1912 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1913 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1914 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1915 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1916 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001917 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1918 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001919 "src/f32-vunary/gen/vabs-avx-x8.c",
1920 "src/f32-vunary/gen/vabs-avx-x16.c",
1921 "src/f32-vunary/gen/vneg-avx-x8.c",
1922 "src/f32-vunary/gen/vneg-avx-x16.c",
1923 "src/f32-vunary/gen/vsqr-avx-x8.c",
1924 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001925]
1926
Marat Dukhan1566fee2020-08-02 21:55:41 -07001927XOP_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001928 "src/qs8-dwconv/gen/up8x9-minmax-xop-mul16.c",
1929 "src/qs8-dwconv/gen/up16x9-minmax-xop-mul16.c",
1930 "src/qs8-dwconv/gen/up24x9-minmax-xop-mul16.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001931 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
1932 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
1933 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
1934 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001935 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
1936 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001937 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
1938 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001939 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001940 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
1941 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001942 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001943 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
1944 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
1945 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001946 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
1947 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
1948 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
1949 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
1950 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
1951 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001952 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001953 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
1954 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001955 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001956]
1957
Marat Dukhanfda12b82019-11-21 12:27:59 -08001958FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001959 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1960 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1961 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1962 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1963 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1964 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1965 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1966 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1967 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1968 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1969 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1970 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1971 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1972 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1973 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1974 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1975 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1976 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1977 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1978 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1979 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1980 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1981 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1982 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1983 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1984 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1985 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1986 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1987 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1988 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1989 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1990 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1991 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1992 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1993 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1994 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1995 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1996 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1997 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1998 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001999 "src/f32-hswish/gen/hswish-fma3-x8.c",
2000 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002001 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
2002 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2003 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2004 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2005 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2006 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2007 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2008 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2009 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2010 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2011 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2012 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2013 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2014 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002015 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2016 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2017 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2018 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2019 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2020 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2021 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2022 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002023 "src/math/sqrt-fma3-nr1fma.c",
2024 "src/math/sqrt-fma3-nr2fma.c",
2025 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002026]
2027
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002028AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002029 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2030 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2031 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2032 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2033 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2034 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2035 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2036 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2037 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2038 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2039 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2040 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2041 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2042 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2043 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2044 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2045 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2046 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2047 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2048 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2049 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2050 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2051 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2052 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2053 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2054 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2055 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2056 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2057 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2058 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2059 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2060 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2061 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2062 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2063 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2064 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002065 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2066 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2067 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2068 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2069 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2070 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2071 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2072 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2073 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2074 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2075 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2076 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2077 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2078 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2079 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2080 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2081 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2082 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2083 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2084 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2085 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2086 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2087 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2088 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2089 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2090 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2091 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2092 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2093 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2094 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002095 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2096 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2097 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2098 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2099 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2100 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2101 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2102 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2103 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2104 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2105 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2106 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2107 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2108 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2109 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2110 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2111 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2112 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2113 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2114 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2115 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2116 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2117 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2118 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002119 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2120 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2121 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2122 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002123 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2124 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2125 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002126 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2127 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2128 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002129 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2130 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2131 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002132 "src/math/exp-avx2-p5.c",
2133 "src/math/exp-avx2-perm-p3.c",
2134 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002135 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002136 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002137 "src/math/sigmoid-avx2-rr2-p5-div.c",
2138 "src/math/sigmoid-avx2-rr1-p5-div.c",
2139 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2140 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2141 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2142 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002143]
2144
Marat Dukhan08c4a432019-10-03 09:29:21 -07002145AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002146 "src/f32-clamp/gen/avx512f-x16.c",
2147 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002148 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2149 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2150 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2151 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2152 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2153 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2154 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2155 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2156 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2157 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2158 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2159 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2160 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2161 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2162 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2163 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2164 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2165 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2166 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2167 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2168 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2169 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2170 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2171 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002172 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2173 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002174 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2175 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2176 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2177 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2178 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2179 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002180 "src/f32-prelu/gen/avx512f-2x16.c",
2181 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002182 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2183 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2184 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2185 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2186 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2187 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2188 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2189 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2190 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2191 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2192 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2193 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2194 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2195 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2196 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2197 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2198 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2199 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2200 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2201 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2202 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2203 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2204 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2205 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2206 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2207 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2208 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2209 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2210 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2211 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2212 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2213 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2214 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2215 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2216 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2217 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002218 "src/f32-relu/gen/avx512f-x16.c",
2219 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002220 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002221 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2222 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2223 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2224 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2225 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2226 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2227 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2228 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002229 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2230 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2231 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2232 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2233 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2234 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2235 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2236 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002237 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2238 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2239 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2240 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2241 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2242 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2243 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2244 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002245 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2246 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2247 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2248 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002249 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2250 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2251 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2252 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002253 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2254 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002255 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002256 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2257 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2258 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2259 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2260 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2261 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2262 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2263 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2264 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2265 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2266 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2267 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2268 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2269 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2270 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2271 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2272 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2273 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2274 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2275 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2276 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2277 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2278 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2279 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002280 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2281 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2282 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2283 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2284 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2285 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2286 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2287 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002288 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2289 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2290 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2291 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2292 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2293 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2294 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2295 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002296 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2297 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2298 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2299 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2300 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2301 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002302 "src/math/exp-avx512f-p5-scalef.c",
2303 "src/math/exp-avx512f-p5.c",
2304 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002305 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002306 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002307 "src/math/sqrt-avx512f-nr1fma.c",
2308 "src/math/sqrt-avx512f-nr2fma.c",
2309 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002310]
2311
2312AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002313 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002314 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002315 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2316 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002317 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002318 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002319 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002320 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002321 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2322 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002323 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2324 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2325 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2326 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002327]
2328
2329AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002330 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002331 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2332 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002333 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002334 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2335 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002336 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2337 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2338 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002339 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002340 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2341 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2342 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002343 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002344 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2345 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002346 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002347 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2348 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2349 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2350 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2351 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2352 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2353 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2354 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2355 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2356 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2357 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2358 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2359 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2360 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2361 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2362 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2363 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2364 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002365 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2366 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002367 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002368 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2369 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2370 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2371 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2372 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2373 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2374 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2375 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2376 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2377 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2378 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2379 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2380 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2381 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2382 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2383 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2384 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2385 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002386 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2387 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2388 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2389 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2390 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2391 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2392 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2393 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2394 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2395 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2396 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2397 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2398 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2399 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2400 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2401 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2402 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2403 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002404]
2405
Marat Dukhan1b354632020-03-23 12:50:22 -07002406INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002407 "src/xnnpack/argmaxpool.h",
2408 "src/xnnpack/avgpool.h",
2409 "src/xnnpack/clamp.h",
2410 "src/xnnpack/common.h",
2411 "src/xnnpack/conv.h",
2412 "src/xnnpack/dwconv.h",
2413 "src/xnnpack/gavgpool.h",
2414 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002415 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002416 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002417 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002418 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002419 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002420 "src/xnnpack/lut.h",
2421 "src/xnnpack/math.h",
2422 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002423 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002424 "src/xnnpack/packx.h",
2425 "src/xnnpack/pad.h",
2426 "src/xnnpack/params.h",
2427 "src/xnnpack/pavgpool.h",
2428 "src/xnnpack/ppmm.h",
2429 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002430 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002431 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002432 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002433 "src/xnnpack/rmax.h",
2434 "src/xnnpack/scalar-utils.h",
2435 "src/xnnpack/spmm.h",
2436 "src/xnnpack/unpool.h",
2437 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002438 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002439 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002440 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002441 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002442 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002443 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002444 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002445]
2446
2447INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002448 "include/xnnpack.h",
2449 "src/xnnpack/allocator.h",
2450 "src/xnnpack/compute.h",
2451 "src/xnnpack/im2col.h",
2452 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002453 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002454 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002455 "src/xnnpack/operator.h",
2456 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002457 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002458 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002459 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002460 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002461]
2462
Marat Dukhan1b354632020-03-23 12:50:22 -07002463ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002464 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002465]
2466
Marat Dukhan1b354632020-03-23 12:50:22 -07002467MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002468 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002469 "include/xnnpack.h",
2470]
2471
Marat Dukhan1b354632020-03-23 12:50:22 -07002472MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002473 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002474 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002475 "src/xnnpack/requantization.h",
2476 "include/xnnpack.h",
2477]
2478
2479OPERATOR_TEST_PARAMS_HDRS = [
2480 "src/xnnpack/params.h",
2481 "src/xnnpack/common.h",
2482]
2483
2484WEIGHTS_PACK_HDRS = [
2485 "src/xnnpack/pack.h",
2486 "src/xnnpack/operator.h",
2487 "src/xnnpack/compute.h",
2488]
2489
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002490LOGGING_COPTS = select({
2491 # No logging in optimized mode
2492 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2493 # Full logging in debug mode
2494 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2495 # Error-only logging in default (fastbuild) mode
2496 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2497})
2498
Marat Dukhan3b59de22020-06-03 20:15:19 -07002499LOGGING_SRCS = select({
2500 # No logging in optimized mode
2501 ":optimized_build": [],
2502 "//conditions:default": [
2503 "src/operator-strings.c",
2504 "src/subgraph-strings.c",
2505 ],
2506})
2507
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002508LOGGING_HDRS = [
2509 "src/xnnpack/log.h",
2510]
2511
Marat Dukhan08c4a432019-10-03 09:29:21 -07002512xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002513 name = "tables",
2514 srcs = TABLE_SRCS,
2515 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002516 gcc_copts = xnnpack_gcc_std_copts(),
2517 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002518)
2519
2520xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002521 name = "scalar_ukernels",
2522 srcs = SCALAR_UKERNELS,
2523 hdrs = INTERNAL_HDRS,
2524 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002525 gcc_copts = xnnpack_gcc_std_copts(),
2526 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002527 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002528 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002529 "@FP16",
2530 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002531 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002532 ],
2533)
2534
2535xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002536 name = "scalar_ukernels_test_mode",
2537 srcs = SCALAR_UKERNELS,
2538 hdrs = INTERNAL_HDRS,
2539 aarch32_copts = ["-marm"],
2540 copts = [
2541 "-UNDEBUG",
2542 "-DXNN_TEST_MODE=1",
2543 ],
2544 gcc_copts = xnnpack_gcc_std_copts(),
2545 msvc_copts = xnnpack_msvc_std_copts(),
2546 deps = [
2547 ":tables",
2548 "@FP16",
2549 "@FXdiv",
2550 "@pthreadpool",
2551 ],
2552)
2553
2554xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002555 name = "wasm_ukernels",
2556 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002557 gcc_copts = xnnpack_gcc_std_copts(),
2558 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002559 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002560 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002561 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002562 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002563 "@FP16",
2564 "@FXdiv",
2565 "@pthreadpool",
2566 ],
2567)
2568
2569xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002570 name = "wasm_ukernels_test_mode",
2571 hdrs = INTERNAL_HDRS,
2572 copts = [
2573 "-UNDEBUG",
2574 "-DXNN_TEST_MODE=1",
2575 ],
2576 gcc_copts = xnnpack_gcc_std_copts(),
2577 msvc_copts = xnnpack_msvc_std_copts(),
2578 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002579 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002580 deps = [
2581 ":tables",
2582 "@FP16",
2583 "@FXdiv",
2584 "@pthreadpool",
2585 ],
2586)
2587
2588xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002589 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002590 hdrs = INTERNAL_HDRS,
2591 aarch32_copts = [
2592 "-marm",
2593 "-mfpu=neon",
2594 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002595 gcc_copts = xnnpack_gcc_std_copts(),
2596 gcc_x86_copts = ["-msse2"],
2597 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002598 optimized_copts = [
2599 "-O3",
2600 "-ffast-math",
2601 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002602 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002603 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002604 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002605 "@FP16",
2606 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002607 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002608 ],
2609)
2610
2611xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002612 name = "psimd_fastmath_ukernels_test_mode",
2613 hdrs = INTERNAL_HDRS,
2614 aarch32_copts = [
2615 "-marm",
2616 "-mfpu=neon",
2617 ],
2618 copts = [
2619 "-UNDEBUG",
2620 "-DXNN_TEST_MODE=1",
2621 ],
2622 gcc_copts = xnnpack_gcc_std_copts(),
2623 gcc_x86_copts = ["-msse2"],
2624 msvc_copts = xnnpack_msvc_std_copts(),
2625 optimized_copts = [
2626 "-O3",
2627 "-ffast-math",
2628 ],
2629 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2630 deps = [
2631 ":tables",
2632 "@FP16",
2633 "@psimd",
2634 "@pthreadpool",
2635 ],
2636)
2637
2638xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002639 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002640 hdrs = INTERNAL_HDRS,
2641 aarch32_copts = [
2642 "-marm",
2643 "-mfpu=neon",
2644 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002645 gcc_copts = xnnpack_gcc_std_copts(),
2646 gcc_x86_copts = ["-msse2"],
2647 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002648 optimized_copts = [
2649 "-O3",
2650 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002651 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002652 deps = [
2653 ":tables",
2654 "@FP16",
2655 "@psimd",
2656 "@pthreadpool",
2657 ],
2658)
2659
2660xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002661 name = "psimd_accmath_ukernels_test_mode",
2662 hdrs = INTERNAL_HDRS,
2663 aarch32_copts = [
2664 "-marm",
2665 "-mfpu=neon",
2666 ],
2667 copts = [
2668 "-UNDEBUG",
2669 "-DXNN_TEST_MODE=1",
2670 ],
2671 gcc_copts = xnnpack_gcc_std_copts(),
2672 gcc_x86_copts = ["-msse2"],
2673 msvc_copts = xnnpack_msvc_std_copts(),
2674 optimized_copts = [
2675 "-O3",
2676 ],
2677 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2678 deps = [
2679 ":tables",
2680 "@FP16",
2681 "@psimd",
2682 "@pthreadpool",
2683 ],
2684)
2685
2686xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002687 name = "neon_ukernels",
2688 hdrs = INTERNAL_HDRS,
2689 aarch32_copts = [
2690 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002691 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002692 "-mfpu=neon",
2693 ],
2694 aarch32_srcs = NEON_UKERNELS,
2695 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002696 gcc_copts = xnnpack_gcc_std_copts(),
2697 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002698 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002699 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002700 "@FP16",
2701 "@pthreadpool",
2702 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002703)
2704
2705xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002706 name = "neon_ukernels_test_mode",
2707 hdrs = INTERNAL_HDRS,
2708 aarch32_copts = [
2709 "-marm",
2710 "-march=armv7-a",
2711 "-mfpu=neon",
2712 ],
2713 aarch32_srcs = NEON_UKERNELS,
2714 aarch64_srcs = NEON_UKERNELS,
2715 copts = [
2716 "-UNDEBUG",
2717 "-DXNN_TEST_MODE=1",
2718 ],
2719 gcc_copts = xnnpack_gcc_std_copts(),
2720 msvc_copts = xnnpack_msvc_std_copts(),
2721 deps = [
2722 ":tables",
2723 "@FP16",
2724 "@pthreadpool",
2725 ],
2726)
2727
2728xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002729 name = "neonfma_ukernels",
2730 hdrs = INTERNAL_HDRS,
2731 aarch32_copts = [
2732 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002733 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002734 "-mfpu=neon-vfpv4",
2735 ],
2736 aarch32_srcs = NEONFMA_UKERNELS,
2737 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002738 apple_aarch32_copts = [
2739 "-mcpu=swift",
2740 "-mtune=generic",
2741 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002742 gcc_copts = xnnpack_gcc_std_copts(),
2743 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002744 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002745 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002746 "@FP16",
2747 "@pthreadpool",
2748 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002749)
2750
2751xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002752 name = "neonfma_ukernels_test_mode",
2753 hdrs = INTERNAL_HDRS,
2754 aarch32_copts = [
2755 "-marm",
2756 "-march=armv7-a",
2757 "-mfpu=neon-vfpv4",
2758 ],
2759 aarch32_srcs = NEONFMA_UKERNELS,
2760 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002761 apple_aarch32_copts = [
2762 "-mcpu=swift",
2763 "-mtune=generic",
2764 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002765 copts = [
2766 "-UNDEBUG",
2767 "-DXNN_TEST_MODE=1",
2768 ],
2769 gcc_copts = xnnpack_gcc_std_copts(),
2770 msvc_copts = xnnpack_msvc_std_copts(),
2771 deps = [
2772 ":tables",
2773 "@FP16",
2774 "@pthreadpool",
2775 ],
2776)
2777
2778xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002779 name = "neonv8_ukernels",
2780 hdrs = INTERNAL_HDRS,
2781 aarch32_copts = [
2782 "-marm",
2783 "-march=armv8-a",
2784 "-mfpu=neon-fp-armv8",
2785 ],
2786 aarch32_srcs = NEONV8_UKERNELS,
2787 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002788 apple_aarch32_copts = [
2789 "-mcpu=cyclone",
2790 "-mtune=generic",
2791 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002792 gcc_copts = xnnpack_gcc_std_copts(),
2793 msvc_copts = xnnpack_msvc_std_copts(),
2794 deps = [
2795 ":tables",
2796 "@FP16",
2797 "@pthreadpool",
2798 ],
2799)
2800
2801xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002802 name = "neonv8_ukernels_test_mode",
2803 hdrs = INTERNAL_HDRS,
2804 aarch32_copts = [
2805 "-marm",
2806 "-march=armv8-a",
2807 "-mfpu=neon-fp-armv8",
2808 ],
2809 aarch32_srcs = NEONV8_UKERNELS,
2810 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002811 apple_aarch32_copts = [
2812 "-mcpu=cyclone",
2813 "-mtune=generic",
2814 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002815 copts = [
2816 "-UNDEBUG",
2817 "-DXNN_TEST_MODE=1",
2818 ],
2819 gcc_copts = xnnpack_gcc_std_copts(),
2820 msvc_copts = xnnpack_msvc_std_copts(),
2821 deps = [
2822 ":tables",
2823 "@FP16",
2824 "@pthreadpool",
2825 ],
2826)
2827
2828xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002829 name = "neonfp16arith_ukernels",
2830 hdrs = INTERNAL_HDRS,
2831 aarch64_copts = ["-march=armv8.2-a+fp16"],
2832 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002833 gcc_copts = xnnpack_gcc_std_copts(),
2834 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002835 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002836 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002837 "@FP16",
2838 "@pthreadpool",
2839 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002840)
2841
2842xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002843 name = "neonfp16arith_ukernels_test_mode",
2844 hdrs = INTERNAL_HDRS,
2845 aarch64_copts = ["-march=armv8.2-a+fp16"],
2846 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2847 copts = [
2848 "-UNDEBUG",
2849 "-DXNN_TEST_MODE=1",
2850 ],
2851 gcc_copts = xnnpack_gcc_std_copts(),
2852 msvc_copts = xnnpack_msvc_std_copts(),
2853 deps = [
2854 ":tables",
2855 "@FP16",
2856 "@pthreadpool",
2857 ],
2858)
2859
2860xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002861 name = "sse2_ukernels",
2862 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002863 gcc_copts = xnnpack_gcc_std_copts(),
2864 gcc_x86_copts = ["-msse2"],
2865 msvc_copts = xnnpack_msvc_std_copts(),
2866 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002867 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002868 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002869 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002870 "@FP16",
2871 "@pthreadpool",
2872 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002873)
2874
2875xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002876 name = "sse2_ukernels_test_mode",
2877 hdrs = INTERNAL_HDRS,
2878 copts = [
2879 "-UNDEBUG",
2880 "-DXNN_TEST_MODE=1",
2881 ],
2882 gcc_copts = xnnpack_gcc_std_copts(),
2883 gcc_x86_copts = ["-msse2"],
2884 msvc_copts = xnnpack_msvc_std_copts(),
2885 msvc_x86_32_copts = ["/arch:SSE2"],
2886 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2887 deps = [
2888 ":tables",
2889 "@FP16",
2890 "@pthreadpool",
2891 ],
2892)
2893
2894xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002895 name = "ssse3_ukernels",
2896 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002897 gcc_copts = xnnpack_gcc_std_copts(),
2898 gcc_x86_copts = ["-mssse3"],
2899 msvc_copts = xnnpack_msvc_std_copts(),
2900 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002901 x86_srcs = SSSE3_UKERNELS,
2902 deps = [
2903 ":tables",
2904 "@FP16",
2905 "@pthreadpool",
2906 ],
2907)
2908
2909xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002910 name = "ssse3_ukernels_test_mode",
2911 hdrs = INTERNAL_HDRS,
2912 copts = [
2913 "-UNDEBUG",
2914 "-DXNN_TEST_MODE=1",
2915 ],
2916 gcc_copts = xnnpack_gcc_std_copts(),
2917 gcc_x86_copts = ["-mssse3"],
2918 msvc_copts = xnnpack_msvc_std_copts(),
2919 msvc_x86_32_copts = ["/arch:SSE2"],
2920 x86_srcs = SSSE3_UKERNELS,
2921 deps = [
2922 ":tables",
2923 "@FP16",
2924 "@pthreadpool",
2925 ],
2926)
2927
2928xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002929 name = "sse41_ukernels",
2930 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002931 gcc_copts = xnnpack_gcc_std_copts(),
2932 gcc_x86_copts = ["-msse4.1"],
2933 msvc_copts = xnnpack_msvc_std_copts(),
2934 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002935 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002936 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002937 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002938 "@FP16",
2939 "@pthreadpool",
2940 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002941)
2942
2943xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002944 name = "sse41_ukernels_test_mode",
2945 hdrs = INTERNAL_HDRS,
2946 copts = [
2947 "-UNDEBUG",
2948 "-DXNN_TEST_MODE=1",
2949 ],
2950 gcc_copts = xnnpack_gcc_std_copts(),
2951 gcc_x86_copts = ["-msse4.1"],
2952 msvc_copts = xnnpack_msvc_std_copts(),
2953 msvc_x86_32_copts = ["/arch:SSE2"],
2954 x86_srcs = SSE41_UKERNELS,
2955 deps = [
2956 ":tables",
2957 "@FP16",
2958 "@pthreadpool",
2959 ],
2960)
2961
2962xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002963 name = "avx_ukernels",
2964 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002965 gcc_copts = xnnpack_gcc_std_copts(),
2966 gcc_x86_copts = ["-mavx"],
2967 msvc_copts = xnnpack_msvc_std_copts(),
2968 msvc_x86_32_copts = ["/arch:AVX"],
2969 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002970 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002971 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002972 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002973 "@FP16",
2974 "@pthreadpool",
2975 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002976)
2977
2978xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002979 name = "avx_ukernels_test_mode",
2980 hdrs = INTERNAL_HDRS,
2981 copts = [
2982 "-UNDEBUG",
2983 "-DXNN_TEST_MODE=1",
2984 ],
2985 gcc_copts = xnnpack_gcc_std_copts(),
2986 gcc_x86_copts = ["-mavx"],
2987 msvc_copts = xnnpack_msvc_std_copts(),
2988 msvc_x86_32_copts = ["/arch:AVX"],
2989 msvc_x86_64_copts = ["/arch:AVX"],
2990 x86_srcs = AVX_UKERNELS,
2991 deps = [
2992 ":tables",
2993 "@FP16",
2994 "@pthreadpool",
2995 ],
2996)
2997
2998xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07002999 name = "xop_ukernels",
3000 hdrs = INTERNAL_HDRS,
3001 gcc_copts = xnnpack_gcc_std_copts(),
3002 gcc_x86_copts = ["-mxop"],
3003 msvc_copts = xnnpack_msvc_std_copts(),
3004 msvc_x86_32_copts = ["/arch:AVX"],
3005 msvc_x86_64_copts = ["/arch:AVX"],
3006 x86_srcs = XOP_UKERNELS,
3007 deps = [
3008 ":tables",
3009 "@FP16",
3010 "@pthreadpool",
3011 ],
3012)
3013
3014xnnpack_cc_library(
3015 name = "xop_ukernels_test_mode",
3016 hdrs = INTERNAL_HDRS,
3017 copts = [
3018 "-UNDEBUG",
3019 "-DXNN_TEST_MODE=1",
3020 ],
3021 gcc_copts = xnnpack_gcc_std_copts(),
3022 gcc_x86_copts = ["-mxop"],
3023 msvc_copts = xnnpack_msvc_std_copts(),
3024 msvc_x86_32_copts = ["/arch:AVX"],
3025 msvc_x86_64_copts = ["/arch:AVX"],
3026 x86_srcs = XOP_UKERNELS,
3027 deps = [
3028 ":tables",
3029 "@FP16",
3030 "@pthreadpool",
3031 ],
3032)
3033
3034xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003035 name = "fma3_ukernels",
3036 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003037 gcc_copts = xnnpack_gcc_std_copts(),
3038 gcc_x86_copts = ["-mfma"],
3039 msvc_copts = xnnpack_msvc_std_copts(),
3040 msvc_x86_32_copts = ["/arch:AVX"],
3041 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003042 x86_srcs = FMA3_UKERNELS,
3043 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003044 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003045 "@FP16",
3046 "@pthreadpool",
3047 ],
3048)
3049
3050xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003051 name = "fma3_ukernels_test_mode",
3052 hdrs = INTERNAL_HDRS,
3053 copts = [
3054 "-UNDEBUG",
3055 "-DXNN_TEST_MODE=1",
3056 ],
3057 gcc_copts = xnnpack_gcc_std_copts(),
3058 gcc_x86_copts = ["-mfma"],
3059 msvc_copts = xnnpack_msvc_std_copts(),
3060 msvc_x86_32_copts = ["/arch:AVX"],
3061 msvc_x86_64_copts = ["/arch:AVX"],
3062 x86_srcs = FMA3_UKERNELS,
3063 deps = [
3064 ":tables",
3065 "@FP16",
3066 "@pthreadpool",
3067 ],
3068)
3069
3070xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003071 name = "avx2_ukernels",
3072 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003073 gcc_copts = xnnpack_gcc_std_copts(),
3074 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003075 "-mfma",
3076 "-mavx2",
3077 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003078 msvc_copts = xnnpack_msvc_std_copts(),
3079 msvc_x86_32_copts = ["/arch:AVX2"],
3080 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003081 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003082 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003083 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003084 "@FP16",
3085 "@pthreadpool",
3086 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003087)
3088
3089xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003090 name = "avx2_ukernels_test_mode",
3091 hdrs = INTERNAL_HDRS,
3092 copts = [
3093 "-UNDEBUG",
3094 "-DXNN_TEST_MODE=1",
3095 ],
3096 gcc_copts = xnnpack_gcc_std_copts(),
3097 gcc_x86_copts = [
3098 "-mfma",
3099 "-mavx2",
3100 ],
3101 msvc_copts = xnnpack_msvc_std_copts(),
3102 msvc_x86_32_copts = ["/arch:AVX2"],
3103 msvc_x86_64_copts = ["/arch:AVX2"],
3104 x86_srcs = AVX2_UKERNELS,
3105 deps = [
3106 ":tables",
3107 "@FP16",
3108 "@pthreadpool",
3109 ],
3110)
3111
3112xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003113 name = "avx512f_ukernels",
3114 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003115 gcc_copts = xnnpack_gcc_std_copts(),
3116 gcc_x86_copts = ["-mavx512f"],
3117 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3118 msvc_copts = xnnpack_msvc_std_copts(),
3119 msvc_x86_32_copts = ["/arch:AVX512"],
3120 msvc_x86_64_copts = ["/arch:AVX512"],
3121 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003122 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003123 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003124 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003125 "@FP16",
3126 "@pthreadpool",
3127 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003128)
3129
3130xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003131 name = "avx512f_ukernels_test_mode",
3132 hdrs = INTERNAL_HDRS,
3133 copts = [
3134 "-UNDEBUG",
3135 "-DXNN_TEST_MODE=1",
3136 ],
3137 gcc_copts = xnnpack_gcc_std_copts(),
3138 gcc_x86_copts = ["-mavx512f"],
3139 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3140 msvc_copts = xnnpack_msvc_std_copts(),
3141 msvc_x86_32_copts = ["/arch:AVX512"],
3142 msvc_x86_64_copts = ["/arch:AVX512"],
3143 msys_copts = ["-fno-asynchronous-unwind-tables"],
3144 x86_srcs = AVX512F_UKERNELS,
3145 deps = [
3146 ":tables",
3147 "@FP16",
3148 "@pthreadpool",
3149 ],
3150)
3151
3152xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003153 name = "asm_ukernels",
3154 hdrs = ["src/xnnpack/assembly.h"],
3155 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003156 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003157 aarch64_srcs = AARCH64_ASM_UKERNELS,
3158)
3159
Marat Dukhan3b59de22020-06-03 20:15:19 -07003160xnnpack_cc_library(
3161 name = "logging_utils",
3162 srcs = LOGGING_SRCS,
3163 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3164 copts = LOGGING_COPTS + [
3165 "-Isrc",
3166 "-Iinclude",
3167 ] + select({
3168 ":debug_build": [],
3169 "//conditions:default": xnnpack_min_size_copts(),
3170 }),
3171 gcc_copts = xnnpack_gcc_std_copts(),
3172 msvc_copts = xnnpack_msvc_std_copts(),
3173 visibility = xnnpack_visibility(),
3174 deps = [
3175 "@FP16",
3176 "@clog",
3177 "@pthreadpool",
3178 ],
3179)
3180
Marat Dukhan08c4a432019-10-03 09:29:21 -07003181xnnpack_aggregate_library(
3182 name = "ukernels",
3183 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003184 ":neon_ukernels",
3185 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003186 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003187 ":asm_ukernels",
3188 ],
3189 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003190 ":neon_ukernels",
3191 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003192 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003193 ":neonfp16arith_ukernels",
3194 ":asm_ukernels",
3195 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003196 generic_deps = [
3197 ":scalar_ukernels",
3198 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003199 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003200 ":psimd_fastmath_ukernels",
3201 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003202 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003203 wasm_deps = [
3204 ":wasm_ukernels",
3205 ],
3206 wasmsimd_deps = [
3207 ":wasm_ukernels",
3208 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003209 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003210 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003211 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003212 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003213 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003214 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003215 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003216 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003217 ":avx512f_ukernels",
3218 ],
3219)
3220
Marat Dukhan33fcf782020-05-24 14:27:15 -07003221xnnpack_aggregate_library(
3222 name = "ukernels_test_mode",
3223 aarch32_deps = [
3224 ":neon_ukernels_test_mode",
3225 ":neonfma_ukernels_test_mode",
3226 ":neonv8_ukernels_test_mode",
3227 ":asm_ukernels",
3228 ],
3229 aarch64_deps = [
3230 ":neon_ukernels_test_mode",
3231 ":neonfma_ukernels_test_mode",
3232 ":neonv8_ukernels_test_mode",
3233 ":neonfp16arith_ukernels_test_mode",
3234 ":asm_ukernels",
3235 ],
3236 generic_deps = [
3237 ":scalar_ukernels_test_mode",
3238 ],
3239 psimd_deps = [
3240 ":psimd_fastmath_ukernels_test_mode",
3241 ":psimd_accmath_ukernels_test_mode",
3242 ],
3243 wasm_deps = [
3244 ":wasm_ukernels_test_mode",
3245 ],
3246 wasmsimd_deps = [
3247 ":wasm_ukernels_test_mode",
3248 ],
3249 x86_deps = [
3250 ":sse2_ukernels_test_mode",
3251 ":ssse3_ukernels_test_mode",
3252 ":sse41_ukernels_test_mode",
3253 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003254 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003255 ":fma3_ukernels_test_mode",
3256 ":avx2_ukernels_test_mode",
3257 ":avx512f_ukernels_test_mode",
3258 ],
3259)
3260
Marat Dukhan86df3842020-06-17 14:03:31 -07003261# This target must be compatible with Android NDK r17 and compile with Clang 6.
3262# It is similar to :ukernels target, but excludes microkernels which require
3263# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003264xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003265 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003266 aarch32_deps = [
3267 ":neon_ukernels",
3268 ":neonfma_ukernels",
3269 ":neonv8_ukernels",
3270 ":asm_ukernels",
3271 ],
3272 aarch64_deps = [
3273 ":neon_ukernels",
3274 ":neonfma_ukernels",
3275 ":neonv8_ukernels",
3276 ":asm_ukernels",
3277 ],
3278 generic_deps = [
3279 ":scalar_ukernels",
3280 ],
3281 psimd_deps = [
3282 ":psimd_fastmath_ukernels",
3283 ":psimd_accmath_ukernels",
3284 ],
3285 wasm_deps = [
3286 ":wasm_ukernels",
3287 ],
3288 wasmsimd_deps = [
3289 ":wasm_ukernels",
3290 ],
3291 x86_deps = [
3292 ":sse2_ukernels",
3293 ":ssse3_ukernels",
3294 ":sse41_ukernels",
3295 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003296 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003297 ":fma3_ukernels",
3298 ":avx2_ukernels",
3299 ":avx512f_ukernels",
3300 ],
3301)
3302
Marat Dukhan08c4a432019-10-03 09:29:21 -07003303xnnpack_cc_library(
3304 name = "im2col",
3305 srcs = ["src/im2col.c"],
3306 hdrs = [
3307 "src/xnnpack/common.h",
3308 "src/xnnpack/im2col.h",
3309 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003310 gcc_copts = xnnpack_gcc_std_copts(),
3311 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003312)
3313
3314xnnpack_cc_library(
3315 name = "indirection",
3316 srcs = ["src/indirection.c"],
3317 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003318 gcc_copts = xnnpack_gcc_std_copts(),
3319 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003320 deps = [
3321 "@FP16",
3322 "@FXdiv",
3323 "@pthreadpool",
3324 ],
3325)
3326
3327xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003328 name = "indirection_test_mode",
3329 srcs = ["src/indirection.c"],
3330 hdrs = INTERNAL_HDRS,
3331 copts = [
3332 "-UNDEBUG",
3333 "-DXNN_TEST_MODE=1",
3334 ],
3335 gcc_copts = xnnpack_gcc_std_copts(),
3336 msvc_copts = xnnpack_msvc_std_copts(),
3337 deps = [
3338 "@FP16",
3339 "@FXdiv",
3340 "@pthreadpool",
3341 ],
3342)
3343
3344xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003345 name = "packing",
3346 srcs = ["src/packing.c"],
3347 hdrs = INTERNAL_HDRS,
3348 gcc_copts = xnnpack_gcc_std_copts(),
3349 msvc_copts = xnnpack_msvc_std_copts(),
3350 deps = [
3351 "@FP16",
3352 "@FXdiv",
3353 "@pthreadpool",
3354 ],
3355)
3356
3357xnnpack_cc_library(
3358 name = "packing_test_mode",
3359 srcs = ["src/packing.c"],
3360 hdrs = INTERNAL_HDRS,
3361 copts = [
3362 "-UNDEBUG",
3363 "-DXNN_TEST_MODE=1",
3364 ],
3365 gcc_copts = xnnpack_gcc_std_copts(),
3366 msvc_copts = xnnpack_msvc_std_copts(),
3367 deps = [
3368 "@FP16",
3369 "@FXdiv",
3370 "@pthreadpool",
3371 ],
3372)
3373
3374xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003375 name = "operator_run",
3376 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003377 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003378 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003379 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3380 "//conditions:default": [],
3381 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003382 gcc_copts = xnnpack_gcc_std_copts(),
3383 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003384 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003385 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003386 "@FP16",
3387 "@FXdiv",
3388 "@clog",
3389 "@pthreadpool",
3390 ],
3391)
3392
Chao Mei6ddfc602020-05-13 22:29:36 -07003393xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003394 name = "operator_run_test_mode",
3395 srcs = ["src/operator-run.c"],
3396 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3397 copts = LOGGING_COPTS + [
3398 "-UNDEBUG",
3399 "-DXNN_TEST_MODE=1",
3400 ] + select({
3401 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3402 "//conditions:default": [],
3403 }),
3404 gcc_copts = xnnpack_gcc_std_copts(),
3405 msvc_copts = xnnpack_msvc_std_copts(),
3406 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003407 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003408 "@FP16",
3409 "@FXdiv",
3410 "@clog",
3411 "@pthreadpool",
3412 ],
3413)
3414
3415xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003416 name = "memory_planner",
3417 srcs = ["src/memory-planner.c"],
3418 hdrs = INTERNAL_HDRS,
3419 defines = select({
3420 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3421 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3422 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3423 }),
3424 gcc_copts = xnnpack_gcc_std_copts(),
3425 msvc_copts = xnnpack_msvc_std_copts(),
3426 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003427 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003428 "@pthreadpool",
3429 ],
3430)
3431
Marat Dukhan33fcf782020-05-24 14:27:15 -07003432xnnpack_cc_library(
3433 name = "memory_planner_test_mode",
3434 srcs = ["src/memory-planner.c"],
3435 hdrs = INTERNAL_HDRS,
3436 copts = [
3437 "-UNDEBUG",
3438 "-DXNN_TEST_MODE=1",
3439 ],
3440 defines = select({
3441 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3442 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3443 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3444 }),
3445 gcc_copts = xnnpack_gcc_std_copts(),
3446 msvc_copts = xnnpack_msvc_std_copts(),
3447 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003448 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003449 "@pthreadpool",
3450 ],
3451)
3452
Marat Dukhan08c4a432019-10-03 09:29:21 -07003453cc_library(
3454 name = "enable_assembly",
3455 defines = select({
3456 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3457 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003458 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003459 }),
3460)
3461
Marat Dukhan9de90e02020-06-18 16:04:12 -07003462cc_library(
3463 name = "enable_sparse",
3464 defines = select({
3465 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3466 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003467 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003468 }),
3469)
3470
Marat Dukhancf056b22019-10-07 10:26:29 -07003471xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003472 name = "operators",
3473 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003474 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003475 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003476 ],
3477 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003478 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003479 "-Isrc",
3480 "-Iinclude",
3481 ] + select({
3482 ":debug_build": [],
3483 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003484 }) + select({
3485 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3486 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003487 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003488 gcc_copts = xnnpack_gcc_std_copts(),
3489 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003490 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003491 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003492 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003493 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003494 "@FP16",
3495 "@FXdiv",
3496 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003497 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003498 ],
3499)
3500
Marat Dukhan10a38082020-04-17 03:58:35 -07003501xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003502 name = "operators_test_mode",
3503 srcs = OPERATOR_SRCS + [
3504 "src/memory.c",
3505 "src/operator-delete.c",
3506 ],
3507 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3508 copts = LOGGING_COPTS + [
3509 "-Isrc",
3510 "-Iinclude",
3511 "-UNDEBUG",
3512 "-DXNN_TEST_MODE=1",
3513 ] + select({
3514 ":debug_build": [],
3515 "//conditions:default": xnnpack_min_size_copts(),
3516 }) + select({
3517 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3518 "//conditions:default": [],
3519 }),
3520 gcc_copts = xnnpack_gcc_std_copts(),
3521 msvc_copts = xnnpack_msvc_std_copts(),
3522 deps = [
3523 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003524 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003525 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003526 "@FP16",
3527 "@FXdiv",
3528 "@clog",
3529 "@pthreadpool",
3530 ],
3531)
3532
3533xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003534 name = "XNNPACK",
3535 srcs = [
3536 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003537 "src/runtime.c",
3538 "src/subgraph.c",
3539 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003540 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003541 hdrs = ["include/xnnpack.h"],
3542 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003543 "-Isrc",
3544 "-Iinclude",
3545 ] + select({
3546 ":debug_build": [],
3547 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003548 }) + select({
3549 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3550 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003551 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003552 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003553 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003554 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003555 visibility = xnnpack_visibility(),
3556 deps = [
3557 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003558 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003559 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003560 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003561 ":operator_run",
3562 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003563 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003564 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003565 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003566 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003567 ] + select({
3568 ":emscripten": [],
3569 "//conditions:default": ["@cpuinfo"],
3570 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003571)
3572
Marat Dukhan10a38082020-04-17 03:58:35 -07003573xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003574 name = "XNNPACK_test_mode",
3575 srcs = [
3576 "src/init.c",
3577 "src/runtime.c",
3578 "src/subgraph.c",
3579 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003580 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003581 hdrs = ["include/xnnpack.h"],
3582 copts = LOGGING_COPTS + [
3583 "-Isrc",
3584 "-Iinclude",
3585 "-UNDEBUG",
3586 "-DXNN_TEST_MODE=1",
3587 ] + select({
3588 ":debug_build": [],
3589 "//conditions:default": xnnpack_min_size_copts(),
3590 }) + select({
3591 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3592 "//conditions:default": [],
3593 }),
3594 gcc_copts = xnnpack_gcc_std_copts(),
3595 includes = ["include"],
3596 msvc_copts = xnnpack_msvc_std_copts(),
3597 visibility = xnnpack_visibility(),
3598 deps = [
3599 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003600 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003601 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003602 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003603 ":operator_run_test_mode",
3604 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003605 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003606 "@clog",
3607 "@FP16",
3608 "@pthreadpool",
3609 ] + select({
3610 ":emscripten": [],
3611 "//conditions:default": ["@cpuinfo"],
3612 }),
3613)
3614
3615xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003616 name = "xnnpack_f32",
3617 srcs = [
3618 "src/init.c",
3619 "src/runtime.c",
3620 "src/subgraph.c",
3621 "src/tensor.c",
3622 ] + SUBGRAPH_SRCS,
3623 hdrs = ["include/xnnpack.h"],
3624 copts = LOGGING_COPTS + [
3625 "-Isrc",
3626 "-Iinclude",
3627 ] + select({
3628 ":debug_build": [],
3629 "//conditions:default": xnnpack_min_size_copts(),
3630 }) + select({
3631 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3632 "//conditions:default": [],
3633 }),
3634 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07003635 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003636 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003637 "XNN_NO_U8_OPERATORS",
3638 "XNN_NO_X8_OPERATORS",
3639 "XNN_NO_F16_OPERATORS",
3640 "XNN_NO_X16_OPERATORS",
3641 ],
3642 gcc_copts = xnnpack_gcc_std_copts(),
3643 includes = ["include"],
3644 msvc_copts = xnnpack_msvc_std_copts(),
3645 visibility = xnnpack_visibility(),
3646 deps = [
3647 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003648 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003649 ":logging_utils",
3650 ":memory_planner",
3651 ":operator_run",
3652 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003653 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003654 "@clog",
3655 "@FP16",
3656 "@pthreadpool",
3657 ] + select({
3658 ":emscripten": [],
3659 "//conditions:default": ["@cpuinfo"],
3660 }),
3661)
3662
3663xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003664 name = "xnnpack_operators_nhwc_f32",
3665 srcs = [
3666 "src/init.c",
3667 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003668 hdrs = ["include/xnnpack.h"],
3669 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003670 "-Isrc",
3671 "-Iinclude",
3672 ] + select({
3673 ":debug_build": [],
3674 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003675 }) + select({
3676 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3677 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003678 }),
3679 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07003680 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003681 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003682 "XNN_NO_U8_OPERATORS",
3683 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003684 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003685 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003686 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003687 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003688 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003689 visibility = xnnpack_visibility(),
3690 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003691 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003692 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003693 ":operator_run",
3694 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003695 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003696 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003697 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003698 ] + select({
3699 ":emscripten": [],
3700 "//conditions:default": ["@cpuinfo"],
3701 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003702)
3703
Marat Dukhancf056b22019-10-07 10:26:29 -07003704xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003705 name = "bench_utils",
3706 srcs = ["bench/utils.cc"],
3707 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003708 deps = [
3709 "@com_google_benchmark//:benchmark",
3710 "@cpuinfo",
3711 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003712)
3713
Frank Barchard7e955972019-10-11 10:34:25 -07003714######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003715
3716xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07003717 name = "qs8_gemm_bench",
3718 srcs = [
3719 "bench/gemm.h",
3720 "bench/qs8-gemm.cc",
3721 "src/xnnpack/AlignedAllocator.h",
3722 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3723 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
3724)
3725
3726xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07003727 name = "qs8_requantization_bench",
3728 srcs = [
3729 "bench/qs8-requantization.cc",
3730 "src/xnnpack/requantization-stubs.h",
3731 "src/xnnpack/AlignedAllocator.h",
3732 ] + MICROKERNEL_BENCHMARK_HDRS,
3733 deps = MICROKERNEL_BENCHMARK_DEPS,
3734)
3735
3736xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003737 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003738 srcs = [
3739 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003740 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003741 "src/xnnpack/AlignedAllocator.h",
3742 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003743 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003744 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003745)
3746
3747xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07003748 name = "qu8_requantization_bench",
3749 srcs = [
3750 "bench/qu8-requantization.cc",
3751 "src/xnnpack/requantization-stubs.h",
3752 "src/xnnpack/AlignedAllocator.h",
3753 ] + MICROKERNEL_BENCHMARK_HDRS,
3754 deps = MICROKERNEL_BENCHMARK_DEPS,
3755)
3756
3757xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003758 name = "f16_igemm_bench",
3759 srcs = [
3760 "bench/f16-igemm.cc",
3761 "bench/conv.h",
3762 "bench/google/conv.h",
3763 "src/xnnpack/AlignedAllocator.h",
3764 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003765 deps = MICROKERNEL_BENCHMARK_DEPS + [
3766 ":indirection",
3767 ":packing",
3768 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003769)
3770
3771xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003772 name = "f16_gemm_bench",
3773 srcs = [
3774 "bench/f16-gemm.cc",
3775 "bench/gemm.h",
3776 "src/xnnpack/AlignedAllocator.h",
3777 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003778 deps = MICROKERNEL_BENCHMARK_DEPS + [
3779 ":packing",
3780 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003781)
3782
3783xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003784 name = "f16_spmm_bench",
3785 srcs = [
3786 "bench/f16-spmm.cc",
3787 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003788 "src/xnnpack/AlignedAllocator.h",
3789 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003790 deps = MICROKERNEL_BENCHMARK_DEPS,
3791)
3792
3793xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003794 name = "f32_igemm_bench",
3795 srcs = [
3796 "bench/f32-igemm.cc",
3797 "bench/conv.h",
3798 "src/xnnpack/AlignedAllocator.h",
3799 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003800 deps = MICROKERNEL_BENCHMARK_DEPS + [
3801 ":indirection",
3802 ":packing",
3803 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003804)
3805
3806xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07003807 name = "f16_relu_bench",
3808 srcs = [
3809 "bench/f16-relu.cc",
3810 "src/xnnpack/AlignedAllocator.h",
3811 ] + MICROKERNEL_BENCHMARK_HDRS,
3812 deps = MICROKERNEL_BENCHMARK_DEPS,
3813)
3814
3815xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003816 name = "f32_conv_hwc_bench",
3817 srcs = [
3818 "bench/f32-conv-hwc.cc",
3819 "bench/dconv.h",
3820 "src/xnnpack/AlignedAllocator.h",
3821 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003822 deps = MICROKERNEL_BENCHMARK_DEPS + [
3823 ":packing",
3824 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003825)
3826
3827xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003828 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003829 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003830 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003831 "bench/dconv.h",
3832 "src/xnnpack/AlignedAllocator.h",
3833 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003834 deps = MICROKERNEL_BENCHMARK_DEPS + [
3835 ":packing",
3836 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003837)
3838
3839xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003840 name = "f16_dwconv_bench",
3841 srcs = [
3842 "bench/f16-dwconv.cc",
3843 "bench/dwconv.h",
3844 "bench/google/dwconv.h",
3845 "src/xnnpack/AlignedAllocator.h",
3846 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003847 deps = MICROKERNEL_BENCHMARK_DEPS + [
3848 ":indirection",
3849 ":packing",
3850 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003851)
3852
3853xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003854 name = "f32_dwconv_bench",
3855 srcs = [
3856 "bench/f32-dwconv.cc",
3857 "bench/dwconv.h",
3858 "src/xnnpack/AlignedAllocator.h",
3859 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003860 deps = MICROKERNEL_BENCHMARK_DEPS + [
3861 ":indirection",
3862 ":packing",
3863 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003864)
3865
3866xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003867 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003868 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003869 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003870 "bench/dwconv.h",
3871 "src/xnnpack/AlignedAllocator.h",
3872 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003873 deps = MICROKERNEL_BENCHMARK_DEPS + [
3874 ":indirection",
3875 ":packing",
3876 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003877)
3878
3879xnnpack_benchmark(
3880 name = "f32_gemm_bench",
3881 srcs = [
3882 "bench/f32-gemm.cc",
3883 "bench/gemm.h",
3884 "src/xnnpack/AlignedAllocator.h",
3885 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003886 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003887 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003888)
3889
3890xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003891 name = "f32_hswish_bench",
3892 srcs = [
3893 "bench/f32-hswish.cc",
3894 "src/xnnpack/AlignedAllocator.h",
3895 ] + MICROKERNEL_BENCHMARK_HDRS,
3896 deps = MICROKERNEL_BENCHMARK_DEPS,
3897)
3898
3899xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003900 name = "f32_raddexpminusmax_bench",
3901 srcs = [
3902 "bench/f32-raddexpminusmax.cc",
3903 "src/xnnpack/AlignedAllocator.h",
3904 ] + MICROKERNEL_BENCHMARK_HDRS,
3905 deps = MICROKERNEL_BENCHMARK_DEPS,
3906)
3907
3908xnnpack_benchmark(
3909 name = "f32_raddextexp_bench",
3910 srcs = [
3911 "bench/f32-raddextexp.cc",
3912 "src/xnnpack/AlignedAllocator.h",
3913 ] + MICROKERNEL_BENCHMARK_HDRS,
3914 deps = MICROKERNEL_BENCHMARK_DEPS,
3915)
3916
3917xnnpack_benchmark(
3918 name = "f32_raddstoreexpminusmax_bench",
3919 srcs = [
3920 "bench/f32-raddstoreexpminusmax.cc",
3921 "src/xnnpack/AlignedAllocator.h",
3922 ] + MICROKERNEL_BENCHMARK_HDRS,
3923 deps = MICROKERNEL_BENCHMARK_DEPS,
3924)
3925
3926xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07003927 name = "f32_relu_bench",
3928 srcs = [
3929 "bench/f32-relu.cc",
3930 "src/xnnpack/AlignedAllocator.h",
3931 ] + MICROKERNEL_BENCHMARK_HDRS,
3932 deps = MICROKERNEL_BENCHMARK_DEPS,
3933)
3934
3935xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003936 name = "f32_rmax_bench",
3937 srcs = [
3938 "bench/f32-rmax.cc",
3939 "src/xnnpack/AlignedAllocator.h",
3940 ] + MICROKERNEL_BENCHMARK_HDRS,
3941 deps = MICROKERNEL_BENCHMARK_DEPS,
3942)
3943
3944xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003945 name = "f32_sigmoid_bench",
3946 srcs = [
3947 "bench/f32-sigmoid.cc",
3948 "src/xnnpack/AlignedAllocator.h",
3949 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003950 deps = MICROKERNEL_BENCHMARK_DEPS,
3951)
3952
3953xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003954 name = "f32_spmm_bench",
3955 srcs = [
3956 "bench/f32-spmm.cc",
3957 "bench/gemm.h",
3958 "src/xnnpack/AlignedAllocator.h",
3959 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003960 deps = MICROKERNEL_BENCHMARK_DEPS,
3961)
3962
3963xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003964 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003965 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003966 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003967 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003968 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003969 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003970)
3971
3972xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003973 name = "f32_vscaleexpminusmax_bench",
3974 srcs = [
3975 "bench/f32-vscaleexpminusmax.cc",
3976 "src/xnnpack/AlignedAllocator.h",
3977 ] + MICROKERNEL_BENCHMARK_HDRS,
3978 deps = MICROKERNEL_BENCHMARK_DEPS,
3979)
3980
3981xnnpack_benchmark(
3982 name = "f32_vscaleextexp_bench",
3983 srcs = [
3984 "bench/f32-vscaleextexp.cc",
3985 "src/xnnpack/AlignedAllocator.h",
3986 ] + MICROKERNEL_BENCHMARK_HDRS,
3987 deps = MICROKERNEL_BENCHMARK_DEPS,
3988)
3989
3990xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003991 name = "f32_vsqrt_bench",
3992 srcs = [
3993 "bench/f32-vsqrt.cc",
3994 "src/xnnpack/AlignedAllocator.h",
3995 ] + MICROKERNEL_BENCHMARK_HDRS,
3996 deps = MICROKERNEL_BENCHMARK_DEPS,
3997)
3998
3999xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004000 name = "f32_im2col_gemm_bench",
4001 srcs = [
4002 "bench/f32-im2col-gemm.cc",
4003 "bench/conv.h",
4004 "src/xnnpack/AlignedAllocator.h",
4005 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004006 deps = MICROKERNEL_BENCHMARK_DEPS + [
4007 ":im2col",
4008 ":packing",
4009 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004010)
4011
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004012xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004013 name = "rounding_bench",
4014 srcs = [
4015 "bench/rounding.cc",
4016 "src/xnnpack/math-stubs.h",
4017 "src/xnnpack/AlignedAllocator.h",
4018 ] + MICROKERNEL_BENCHMARK_HDRS,
4019 deps = MICROKERNEL_BENCHMARK_DEPS,
4020)
4021
Marat Dukhan08c4a432019-10-03 09:29:21 -07004022########################### Benchmarks for operators ###########################
4023
4024xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004025 name = "average_pooling_bench",
4026 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004027 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004028 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004029 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004030)
4031
4032xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004033 name = "bankers_rounding_bench",
4034 srcs = ["bench/bankers-rounding.cc"],
4035 copts = xnnpack_optional_tflite_copts(),
4036 tags = ["nowin32"],
4037 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4038)
4039
4040xnnpack_benchmark(
4041 name = "ceiling_bench",
4042 srcs = ["bench/ceiling.cc"],
4043 copts = xnnpack_optional_tflite_copts(),
4044 tags = ["nowin32"],
4045 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4046)
4047
4048xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004049 name = "channel_shuffle_bench",
4050 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004051 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004052)
4053
4054xnnpack_benchmark(
4055 name = "convolution_bench",
4056 srcs = ["bench/convolution.cc"],
4057 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004058 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004059 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004060)
4061
4062xnnpack_benchmark(
4063 name = "deconvolution_bench",
4064 srcs = ["bench/deconvolution.cc"],
4065 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004066 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004067 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004068)
4069
4070xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004071 name = "floor_bench",
4072 srcs = ["bench/floor.cc"],
4073 copts = xnnpack_optional_tflite_copts(),
4074 tags = ["nowin32"],
4075 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4076)
4077
4078xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004079 name = "global_average_pooling_bench",
4080 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004081 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004082)
4083
4084xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004085 name = "hardswish_bench",
4086 srcs = ["bench/hardswish.cc"],
4087 copts = xnnpack_optional_tflite_copts(),
4088 tags = ["nowin32"],
4089 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4090)
4091
4092xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004093 name = "max_pooling_bench",
4094 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004095 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004096)
4097
4098xnnpack_benchmark(
4099 name = "sigmoid_bench",
4100 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004101 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004102 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004103 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004104)
4105
4106xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004107 name = "prelu_bench",
4108 srcs = ["bench/prelu.cc"],
4109 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004110 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004111 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004112)
4113
4114xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004115 name = "softmax_bench",
4116 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004117 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004118 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004119 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004120)
4121
Marat Dukhan87727142020-06-24 15:24:10 -07004122xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004123 name = "square_root_bench",
4124 srcs = ["bench/square-root.cc"],
4125 copts = xnnpack_optional_tflite_copts(),
4126 tags = ["nowin32"],
4127 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4128)
4129
4130xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004131 name = "truncation_bench",
4132 srcs = ["bench/truncation.cc"],
4133 deps = OPERATOR_BENCHMARK_DEPS,
4134)
4135
Marat Dukhanc068bb62019-10-04 13:24:39 -07004136############################# End-to-end benchmarks ############################
4137
4138cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004139 name = "fp32_mobilenet_v1",
4140 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004141 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004142 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004143 linkstatic = True,
4144 deps = [
4145 ":XNNPACK",
4146 "@pthreadpool",
4147 ],
4148)
4149
4150cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004151 name = "fp16_mobilenet_v1",
4152 srcs = ["models/fp16-mobilenet-v1.cc"],
4153 hdrs = ["models/models.h"],
4154 copts = xnnpack_std_cxxopts(),
4155 linkstatic = True,
4156 deps = [
4157 ":XNNPACK",
4158 "@FP16",
4159 "@pthreadpool",
4160 ],
4161)
4162
4163cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004164 name = "qs8_mobilenet_v1",
4165 srcs = ["models/qs8-mobilenet-v1.cc"],
4166 hdrs = ["models/models.h"],
4167 copts = xnnpack_std_cxxopts(),
4168 linkstatic = True,
4169 deps = [
4170 ":XNNPACK",
4171 "@pthreadpool",
4172 ],
4173)
4174
4175cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004176 name = "fp32_mobilenet_v2",
4177 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004178 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004179 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004180 linkstatic = True,
4181 deps = [
4182 ":XNNPACK",
4183 "@pthreadpool",
4184 ],
4185)
4186
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004187cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004188 name = "fp16_mobilenet_v2",
4189 srcs = ["models/fp16-mobilenet-v2.cc"],
4190 hdrs = ["models/models.h"],
4191 copts = xnnpack_std_cxxopts(),
4192 linkstatic = True,
4193 deps = [
4194 ":XNNPACK",
4195 "@FP16",
4196 "@pthreadpool",
4197 ],
4198)
4199
4200cc_library(
4201 name = "fp32_mobilenet_v3_large",
4202 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004203 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004204 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004205 linkstatic = True,
4206 deps = [
4207 ":XNNPACK",
4208 "@pthreadpool",
4209 ],
4210)
4211
4212cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004213 name = "fp32_mobilenet_v3_small",
4214 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004215 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004216 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004217 linkstatic = True,
4218 deps = [
4219 ":XNNPACK",
4220 "@pthreadpool",
4221 ],
4222)
4223
Marat Dukhanc068bb62019-10-04 13:24:39 -07004224xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004225 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004226 srcs = [
4227 "bench/f32-dwconv-e2e.cc",
4228 "bench/end2end.h",
4229 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004230 deps = MICROKERNEL_BENCHMARK_DEPS + [
4231 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004232 ":fp32_mobilenet_v1",
4233 ":fp32_mobilenet_v2",
4234 ":fp32_mobilenet_v3_large",
4235 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004236 ],
4237)
4238
4239xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004240 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004241 srcs = [
4242 "bench/f32-gemm-e2e.cc",
4243 "bench/end2end.h",
4244 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004245 deps = MICROKERNEL_BENCHMARK_DEPS + [
4246 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004247 ":fp32_mobilenet_v1",
4248 ":fp32_mobilenet_v2",
4249 ":fp32_mobilenet_v3_large",
4250 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004251 ],
4252)
4253
4254xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004255 name = "end2end_bench",
4256 srcs = ["bench/end2end.cc"],
4257 deps = [
4258 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004259 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004260 ":fp16_mobilenet_v1",
4261 ":fp16_mobilenet_v2",
4262 ":fp32_mobilenet_v1",
4263 ":fp32_mobilenet_v2",
4264 ":fp32_mobilenet_v3_large",
4265 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004266 ":qs8_mobilenet_v1",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004267 "@pthreadpool",
4268 ],
4269)
4270
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004271#################### Accuracy evaluation for math functions ####################
4272
4273xnnpack_benchmark(
4274 name = "f32_exp_eval",
4275 srcs = [
4276 "eval/f32-exp.cc",
4277 "src/xnnpack/AlignedAllocator.h",
4278 ] + ACCURACY_EVAL_HDRS,
4279 deps = ACCURACY_EVAL_DEPS,
4280)
4281
Marat Dukhan515c9772019-10-17 18:07:57 -07004282xnnpack_benchmark(
4283 name = "f32_expminus_eval",
4284 srcs = [
4285 "eval/f32-expminus.cc",
4286 "src/xnnpack/AlignedAllocator.h",
4287 ] + ACCURACY_EVAL_HDRS,
4288 deps = ACCURACY_EVAL_DEPS,
4289)
4290
Marat Dukhan98ba4412019-10-23 02:14:28 -07004291xnnpack_benchmark(
4292 name = "f32_extexp_eval",
4293 srcs = [
4294 "eval/f32-extexp.cc",
4295 "src/xnnpack/AlignedAllocator.h",
4296 ] + ACCURACY_EVAL_HDRS,
4297 deps = ACCURACY_EVAL_DEPS,
4298)
4299
Marat Dukhan8853b822020-05-07 12:19:01 -07004300xnnpack_unit_test(
4301 name = "f32_roundne_eval",
4302 srcs = [
4303 "eval/f32-roundne.cc",
4304 "src/xnnpack/AlignedAllocator.h",
4305 "src/xnnpack/math-stubs.h",
4306 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004307 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004308 deps = MICROKERNEL_TEST_DEPS,
4309)
4310
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004311xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004312 name = "f32_roundd_eval",
4313 srcs = [
4314 "eval/f32-roundd.cc",
4315 "src/xnnpack/AlignedAllocator.h",
4316 "src/xnnpack/math-stubs.h",
4317 ] + MICROKERNEL_TEST_HDRS,
4318 automatic = False,
4319 deps = MICROKERNEL_TEST_DEPS,
4320)
4321
4322xnnpack_unit_test(
4323 name = "f32_roundu_eval",
4324 srcs = [
4325 "eval/f32-roundu.cc",
4326 "src/xnnpack/AlignedAllocator.h",
4327 "src/xnnpack/math-stubs.h",
4328 ] + MICROKERNEL_TEST_HDRS,
4329 automatic = False,
4330 deps = MICROKERNEL_TEST_DEPS,
4331)
4332
4333xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004334 name = "f32_roundz_eval",
4335 srcs = [
4336 "eval/f32-roundz.cc",
4337 "src/xnnpack/AlignedAllocator.h",
4338 "src/xnnpack/math-stubs.h",
4339 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004340 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004341 deps = MICROKERNEL_TEST_DEPS,
4342)
4343
Marat Dukhan346a9e52019-11-15 09:06:30 -08004344xnnpack_benchmark(
4345 name = "f32_sigmoid_eval",
4346 srcs = [
4347 "eval/f32-sigmoid.cc",
4348 "src/xnnpack/AlignedAllocator.h",
4349 ] + ACCURACY_EVAL_HDRS,
4350 deps = ACCURACY_EVAL_DEPS,
4351)
4352
Marat Dukhan84000762020-06-29 18:38:43 -07004353xnnpack_benchmark(
4354 name = "f32_sqrt_eval",
4355 srcs = [
4356 "eval/f32-sqrt.cc",
4357 "src/xnnpack/AlignedAllocator.h",
4358 ] + ACCURACY_EVAL_HDRS,
4359 deps = ACCURACY_EVAL_DEPS + [
4360 ":bench_utils",
4361 ],
4362)
4363
Marat Dukhan08c4a432019-10-03 09:29:21 -07004364######################### Unit tests for micro-kernels #########################
4365
4366xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004367 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004368 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004369 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004370 "test/gemm-microkernel-tester.h",
4371 "src/xnnpack/AlignedAllocator.h",
4372 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004373 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004374)
4375
4376xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004377 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004378 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004379 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004380 "test/spmm-microkernel-tester.h",
4381 "src/xnnpack/AlignedAllocator.h",
4382 ] + MICROKERNEL_TEST_HDRS,
4383 deps = MICROKERNEL_TEST_DEPS,
4384)
4385
4386xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004387 name = "f32_argmaxpool_test",
4388 srcs = [
4389 "test/f32-argmaxpool.cc",
4390 "test/argmaxpool-microkernel-tester.h",
4391 "src/xnnpack/AlignedAllocator.h",
4392 ] + MICROKERNEL_TEST_HDRS,
4393 deps = MICROKERNEL_TEST_DEPS,
4394)
4395
4396xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004397 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004398 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004399 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004400 "test/avgpool-microkernel-tester.h",
4401 "src/xnnpack/AlignedAllocator.h",
4402 ] + MICROKERNEL_TEST_HDRS,
4403 deps = MICROKERNEL_TEST_DEPS,
4404)
4405
4406xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004407 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004408 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004409 "test/f32-ibilinear.cc",
4410 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004411 "src/xnnpack/AlignedAllocator.h",
4412 ] + MICROKERNEL_TEST_HDRS,
4413 deps = MICROKERNEL_TEST_DEPS,
4414)
4415
4416xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004417 name = "f16_clamp_test",
4418 srcs = [
4419 "test/f16-clamp.cc",
4420 "test/clamp-microkernel-tester.h",
4421 ] + MICROKERNEL_TEST_HDRS,
4422 deps = MICROKERNEL_TEST_DEPS,
4423)
4424
4425xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004426 name = "f32_clamp_test",
4427 srcs = [
4428 "test/f32-clamp.cc",
4429 "test/clamp-microkernel-tester.h",
4430 ] + MICROKERNEL_TEST_HDRS,
4431 deps = MICROKERNEL_TEST_DEPS,
4432)
4433
4434xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004435 name = "f32_igemm_test",
4436 srcs = [
4437 "test/f32-igemm.cc",
4438 "test/gemm-microkernel-tester.h",
4439 "src/xnnpack/AlignedAllocator.h",
4440 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004441 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004442)
4443
4444xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004445 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004446 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004447 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004448 "test/gemm-microkernel-tester.h",
4449 "src/xnnpack/AlignedAllocator.h",
4450 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004451 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004452)
4453
4454xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004455 name = "f32_igemm_minmax_test",
4456 srcs = [
4457 "test/f32-igemm-minmax.cc",
4458 "test/gemm-microkernel-tester.h",
4459 "src/xnnpack/AlignedAllocator.h",
4460 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004461 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004462)
4463
4464xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004465 name = "f16_igemm_minmax_test",
4466 srcs = [
4467 "test/f16-igemm-minmax.cc",
4468 "test/gemm-microkernel-tester.h",
4469 "src/xnnpack/AlignedAllocator.h",
4470 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004471 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004472)
4473
4474xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004475 name = "f32_conv_hwc_test",
4476 srcs = [
4477 "test/f32-conv-hwc.cc",
4478 "test/conv-hwc-microkernel-tester.h",
4479 "src/xnnpack/AlignedAllocator.h",
4480 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004481 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004482)
4483
4484xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004485 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004486 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004487 "test/f32-conv-hwc2chw.cc",
4488 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004489 "src/xnnpack/AlignedAllocator.h",
4490 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004491 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004492)
4493
4494xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004495 name = "f32_dwconv_test",
4496 srcs = [
4497 "test/f32-dwconv.cc",
4498 "test/dwconv-microkernel-tester.h",
4499 "src/xnnpack/AlignedAllocator.h",
4500 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004501 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004502)
4503
4504xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004505 name = "f16_dwconv_minmax_test",
4506 srcs = [
4507 "test/f16-dwconv-minmax.cc",
4508 "test/dwconv-microkernel-tester.h",
4509 "src/xnnpack/AlignedAllocator.h",
4510 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004511 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004512)
4513
4514xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004515 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004516 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004517 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004518 "test/dwconv-microkernel-tester.h",
4519 "src/xnnpack/AlignedAllocator.h",
4520 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004521 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004522)
4523
4524xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004525 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004526 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004527 "test/f32-dwconv-chw.cc",
4528 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004529 "src/xnnpack/AlignedAllocator.h",
4530 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004531 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004532)
4533
4534xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004535 name = "f16_gavgpool_minmax_test",
4536 srcs = [
4537 "test/f16-gavgpool-minmax.cc",
4538 "test/gavgpool-microkernel-tester.h",
4539 "src/xnnpack/AlignedAllocator.h",
4540 ] + MICROKERNEL_TEST_HDRS,
4541 deps = MICROKERNEL_TEST_DEPS,
4542)
4543
4544xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004545 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004546 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004547 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004548 "test/gavgpool-microkernel-tester.h",
4549 "src/xnnpack/AlignedAllocator.h",
4550 ] + MICROKERNEL_TEST_HDRS,
4551 deps = MICROKERNEL_TEST_DEPS,
4552)
4553
4554xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004555 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004556 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004557 "test/f32-gavgpool-cw.cc",
4558 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004559 "src/xnnpack/AlignedAllocator.h",
4560 ] + MICROKERNEL_TEST_HDRS,
4561 deps = MICROKERNEL_TEST_DEPS,
4562)
4563
4564xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004565 name = "f32_gemm_test",
4566 srcs = [
4567 "test/f32-gemm.cc",
4568 "test/gemm-microkernel-tester.h",
4569 "src/xnnpack/AlignedAllocator.h",
4570 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004571 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004572)
4573
4574xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004575 name = "f32_gemm_relu_test",
4576 srcs = [
4577 "test/f32-gemm-relu.cc",
4578 "test/gemm-microkernel-tester.h",
4579 "src/xnnpack/AlignedAllocator.h",
4580 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004581 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004582)
4583
4584xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004585 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004586 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004587 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004588 "test/gemm-microkernel-tester.h",
4589 "src/xnnpack/AlignedAllocator.h",
4590 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004591 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004592)
4593
4594xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004595 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004596 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004597 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004598 "test/gemm-microkernel-tester.h",
4599 "src/xnnpack/AlignedAllocator.h",
4600 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004601 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004602)
4603
4604xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004605 name = "f16_hswish_test",
4606 srcs = [
4607 "test/f16-hswish.cc",
4608 "test/hswish-microkernel-tester.h",
4609 ] + MICROKERNEL_TEST_HDRS,
4610 deps = MICROKERNEL_TEST_DEPS,
4611)
4612
4613xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004614 name = "f32_hswish_test",
4615 srcs = [
4616 "test/f32-hswish.cc",
4617 "test/hswish-microkernel-tester.h",
4618 ] + MICROKERNEL_TEST_HDRS,
4619 deps = MICROKERNEL_TEST_DEPS,
4620)
4621
4622xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004623 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004624 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004625 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004626 "test/maxpool-microkernel-tester.h",
4627 ] + MICROKERNEL_TEST_HDRS,
4628 deps = MICROKERNEL_TEST_DEPS,
4629)
4630
4631xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004632 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004633 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004634 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004635 "test/avgpool-microkernel-tester.h",
4636 "src/xnnpack/AlignedAllocator.h",
4637 ] + MICROKERNEL_TEST_HDRS,
4638 deps = MICROKERNEL_TEST_DEPS,
4639)
4640
4641xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004642 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004643 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004644 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004645 "test/gemm-microkernel-tester.h",
4646 "src/xnnpack/AlignedAllocator.h",
4647 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004648 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004649)
4650
4651xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004652 name = "f16_prelu_test",
4653 srcs = [
4654 "test/f16-prelu.cc",
4655 "test/prelu-microkernel-tester.h",
4656 "src/xnnpack/AlignedAllocator.h",
4657 ] + MICROKERNEL_TEST_HDRS,
4658 deps = MICROKERNEL_TEST_DEPS,
4659)
4660
4661xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004662 name = "f32_prelu_test",
4663 srcs = [
4664 "test/f32-prelu.cc",
4665 "test/prelu-microkernel-tester.h",
4666 "src/xnnpack/AlignedAllocator.h",
4667 ] + MICROKERNEL_TEST_HDRS,
4668 deps = MICROKERNEL_TEST_DEPS,
4669)
4670
4671xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004672 name = "f32_raddexpminusmax_test",
4673 srcs = [
4674 "test/f32-raddexpminusmax.cc",
4675 "test/raddexpminusmax-microkernel-tester.h",
4676 ] + MICROKERNEL_TEST_HDRS,
4677 deps = MICROKERNEL_TEST_DEPS,
4678)
4679
4680xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004681 name = "f32_raddextexp_test",
4682 srcs = [
4683 "test/f32-raddextexp.cc",
4684 "test/raddextexp-microkernel-tester.h",
4685 ] + MICROKERNEL_TEST_HDRS,
4686 deps = MICROKERNEL_TEST_DEPS,
4687)
4688
4689xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004690 name = "f32_raddstoreexpminusmax_test",
4691 srcs = [
4692 "test/f32-raddstoreexpminusmax.cc",
4693 "test/raddstoreexpminusmax-microkernel-tester.h",
4694 ] + MICROKERNEL_TEST_HDRS,
4695 deps = MICROKERNEL_TEST_DEPS,
4696)
4697
4698xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004699 name = "f32_relu_test",
4700 srcs = [
4701 "test/f32-relu.cc",
4702 "test/vunary-microkernel-tester.h",
4703 ] + MICROKERNEL_TEST_HDRS,
4704 deps = MICROKERNEL_TEST_DEPS,
4705)
4706
4707xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004708 name = "f32_rmax_test",
4709 srcs = [
4710 "test/f32-rmax.cc",
4711 "test/rmax-microkernel-tester.h",
4712 ] + MICROKERNEL_TEST_HDRS,
4713 deps = MICROKERNEL_TEST_DEPS,
4714)
4715
4716xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004717 name = "f32_sigmoid_test",
4718 srcs = [
4719 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004720 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004721 ] + MICROKERNEL_TEST_HDRS,
4722 deps = MICROKERNEL_TEST_DEPS,
4723)
4724
4725xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004726 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004727 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004728 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004729 "test/spmm-microkernel-tester.h",
4730 "src/xnnpack/AlignedAllocator.h",
4731 ] + MICROKERNEL_TEST_HDRS,
4732 deps = MICROKERNEL_TEST_DEPS,
4733)
4734
4735xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004736 name = "f16_vadd_minmax_test",
4737 srcs = [
4738 "test/f16-vadd-minmax.cc",
4739 "test/vbinary-microkernel-tester.h",
4740 ] + MICROKERNEL_TEST_HDRS,
4741 deps = MICROKERNEL_TEST_DEPS,
4742)
4743
4744xnnpack_unit_test(
4745 name = "f16_vaddc_minmax_test",
4746 srcs = [
4747 "test/f16-vaddc-minmax.cc",
4748 "test/vbinaryc-microkernel-tester.h",
4749 ] + MICROKERNEL_TEST_HDRS,
4750 deps = MICROKERNEL_TEST_DEPS,
4751)
4752
4753xnnpack_unit_test(
4754 name = "f16_vdiv_minmax_test",
4755 srcs = [
4756 "test/f16-vdiv-minmax.cc",
4757 "test/vbinary-microkernel-tester.h",
4758 ] + MICROKERNEL_TEST_HDRS,
4759 deps = MICROKERNEL_TEST_DEPS,
4760)
4761
4762xnnpack_unit_test(
4763 name = "f16_vdivc_minmax_test",
4764 srcs = [
4765 "test/f16-vdivc-minmax.cc",
4766 "test/vbinaryc-microkernel-tester.h",
4767 ] + MICROKERNEL_TEST_HDRS,
4768 deps = MICROKERNEL_TEST_DEPS,
4769)
4770
4771xnnpack_unit_test(
4772 name = "f16_vrdivc_minmax_test",
4773 srcs = [
4774 "test/f16-vrdivc-minmax.cc",
4775 "test/vbinaryc-microkernel-tester.h",
4776 ] + MICROKERNEL_TEST_HDRS,
4777 deps = MICROKERNEL_TEST_DEPS,
4778)
4779
4780xnnpack_unit_test(
4781 name = "f16_vmax_test",
4782 srcs = [
4783 "test/f16-vmax.cc",
4784 "test/vbinary-microkernel-tester.h",
4785 ] + MICROKERNEL_TEST_HDRS,
4786 deps = MICROKERNEL_TEST_DEPS,
4787)
4788
4789xnnpack_unit_test(
4790 name = "f16_vmaxc_test",
4791 srcs = [
4792 "test/f16-vmaxc.cc",
4793 "test/vbinaryc-microkernel-tester.h",
4794 ] + MICROKERNEL_TEST_HDRS,
4795 deps = MICROKERNEL_TEST_DEPS,
4796)
4797
4798xnnpack_unit_test(
4799 name = "f16_vmin_test",
4800 srcs = [
4801 "test/f16-vmin.cc",
4802 "test/vbinary-microkernel-tester.h",
4803 ] + MICROKERNEL_TEST_HDRS,
4804 deps = MICROKERNEL_TEST_DEPS,
4805)
4806
4807xnnpack_unit_test(
4808 name = "f16_vminc_test",
4809 srcs = [
4810 "test/f16-vminc.cc",
4811 "test/vbinaryc-microkernel-tester.h",
4812 ] + MICROKERNEL_TEST_HDRS,
4813 deps = MICROKERNEL_TEST_DEPS,
4814)
4815
4816xnnpack_unit_test(
4817 name = "f16_vmul_minmax_test",
4818 srcs = [
4819 "test/f16-vmul-minmax.cc",
4820 "test/vbinary-microkernel-tester.h",
4821 ] + MICROKERNEL_TEST_HDRS,
4822 deps = MICROKERNEL_TEST_DEPS,
4823)
4824
4825xnnpack_unit_test(
4826 name = "f16_vmulc_minmax_test",
4827 srcs = [
4828 "test/f16-vmulc-minmax.cc",
4829 "test/vbinaryc-microkernel-tester.h",
4830 ] + MICROKERNEL_TEST_HDRS,
4831 deps = MICROKERNEL_TEST_DEPS,
4832)
4833
4834xnnpack_unit_test(
4835 name = "f16_vsub_minmax_test",
4836 srcs = [
4837 "test/f16-vsub-minmax.cc",
4838 "test/vbinary-microkernel-tester.h",
4839 ] + MICROKERNEL_TEST_HDRS,
4840 deps = MICROKERNEL_TEST_DEPS,
4841)
4842
4843xnnpack_unit_test(
4844 name = "f16_vsubc_minmax_test",
4845 srcs = [
4846 "test/f16-vsubc-minmax.cc",
4847 "test/vbinaryc-microkernel-tester.h",
4848 ] + MICROKERNEL_TEST_HDRS,
4849 deps = MICROKERNEL_TEST_DEPS,
4850)
4851
4852xnnpack_unit_test(
4853 name = "f16_vrsubc_minmax_test",
4854 srcs = [
4855 "test/f16-vrsubc-minmax.cc",
4856 "test/vbinaryc-microkernel-tester.h",
4857 ] + MICROKERNEL_TEST_HDRS,
4858 deps = MICROKERNEL_TEST_DEPS,
4859)
4860
4861xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004862 name = "f32_vabs_test",
4863 srcs = [
4864 "test/f32-vabs.cc",
4865 "test/vunary-microkernel-tester.h",
4866 ] + MICROKERNEL_TEST_HDRS,
4867 deps = MICROKERNEL_TEST_DEPS,
4868)
4869
4870xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004871 name = "f32_vadd_test",
4872 srcs = [
4873 "test/f32-vadd.cc",
4874 "test/vbinary-microkernel-tester.h",
4875 ] + MICROKERNEL_TEST_HDRS,
4876 deps = MICROKERNEL_TEST_DEPS,
4877)
4878
4879xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004880 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004881 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004882 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004883 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004884 ] + MICROKERNEL_TEST_HDRS,
4885 deps = MICROKERNEL_TEST_DEPS,
4886)
4887
4888xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004889 name = "f32_vaddc_test",
4890 srcs = [
4891 "test/f32-vaddc.cc",
4892 "test/vbinaryc-microkernel-tester.h",
4893 ] + MICROKERNEL_TEST_HDRS,
4894 deps = MICROKERNEL_TEST_DEPS,
4895)
4896
4897xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004898 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004899 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004900 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004901 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004902 ] + MICROKERNEL_TEST_HDRS,
4903 deps = MICROKERNEL_TEST_DEPS,
4904)
4905
4906xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004907 name = "f32_vdiv_test",
4908 srcs = [
4909 "test/f32-vdiv.cc",
4910 "test/vbinary-microkernel-tester.h",
4911 ] + MICROKERNEL_TEST_HDRS,
4912 deps = MICROKERNEL_TEST_DEPS,
4913)
4914
4915xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004916 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004917 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004918 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004919 "test/vbinary-microkernel-tester.h",
4920 ] + MICROKERNEL_TEST_HDRS,
4921 deps = MICROKERNEL_TEST_DEPS,
4922)
4923
4924xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004925 name = "f32_vdivc_test",
4926 srcs = [
4927 "test/f32-vdivc.cc",
4928 "test/vbinaryc-microkernel-tester.h",
4929 ] + MICROKERNEL_TEST_HDRS,
4930 deps = MICROKERNEL_TEST_DEPS,
4931)
4932
4933xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004934 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004935 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004936 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004937 "test/vbinaryc-microkernel-tester.h",
4938 ] + MICROKERNEL_TEST_HDRS,
4939 deps = MICROKERNEL_TEST_DEPS,
4940)
4941
4942xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004943 name = "f32_vrdivc_test",
4944 srcs = [
4945 "test/f32-vrdivc.cc",
4946 "test/vbinaryc-microkernel-tester.h",
4947 ] + MICROKERNEL_TEST_HDRS,
4948 deps = MICROKERNEL_TEST_DEPS,
4949)
4950
4951xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004952 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004953 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004954 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004955 "test/vbinaryc-microkernel-tester.h",
4956 ] + MICROKERNEL_TEST_HDRS,
4957 deps = MICROKERNEL_TEST_DEPS,
4958)
4959
4960xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004961 name = "f32_vmax_test",
4962 srcs = [
4963 "test/f32-vmax.cc",
4964 "test/vbinary-microkernel-tester.h",
4965 ] + MICROKERNEL_TEST_HDRS,
4966 deps = MICROKERNEL_TEST_DEPS,
4967)
4968
4969xnnpack_unit_test(
4970 name = "f32_vmaxc_test",
4971 srcs = [
4972 "test/f32-vmaxc.cc",
4973 "test/vbinaryc-microkernel-tester.h",
4974 ] + MICROKERNEL_TEST_HDRS,
4975 deps = MICROKERNEL_TEST_DEPS,
4976)
4977
4978xnnpack_unit_test(
4979 name = "f32_vmin_test",
4980 srcs = [
4981 "test/f32-vmin.cc",
4982 "test/vbinary-microkernel-tester.h",
4983 ] + MICROKERNEL_TEST_HDRS,
4984 deps = MICROKERNEL_TEST_DEPS,
4985)
4986
4987xnnpack_unit_test(
4988 name = "f32_vminc_test",
4989 srcs = [
4990 "test/f32-vminc.cc",
4991 "test/vbinaryc-microkernel-tester.h",
4992 ] + MICROKERNEL_TEST_HDRS,
4993 deps = MICROKERNEL_TEST_DEPS,
4994)
4995
4996xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004997 name = "f32_vmul_test",
4998 srcs = [
4999 "test/f32-vmul.cc",
5000 "test/vbinary-microkernel-tester.h",
5001 ] + MICROKERNEL_TEST_HDRS,
5002 deps = MICROKERNEL_TEST_DEPS,
5003)
5004
5005xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005006 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005007 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005008 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005009 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005010 ] + MICROKERNEL_TEST_HDRS,
5011 deps = MICROKERNEL_TEST_DEPS,
5012)
5013
5014xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005015 name = "f32_vmulc_test",
5016 srcs = [
5017 "test/f32-vmulc.cc",
5018 "test/vbinaryc-microkernel-tester.h",
5019 ] + MICROKERNEL_TEST_HDRS,
5020 deps = MICROKERNEL_TEST_DEPS,
5021)
5022
5023xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005024 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005025 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005026 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005027 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005028 ] + MICROKERNEL_TEST_HDRS,
5029 deps = MICROKERNEL_TEST_DEPS,
5030)
5031
5032xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005033 name = "f16_vmulcaddc_minmax_test",
5034 srcs = [
5035 "test/f16-vmulcaddc-minmax.cc",
5036 "test/vmulcaddc-microkernel-tester.h",
5037 "src/xnnpack/AlignedAllocator.h",
5038 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005039 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005040)
5041
5042xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005043 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005044 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005045 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005046 "test/vmulcaddc-microkernel-tester.h",
5047 "src/xnnpack/AlignedAllocator.h",
5048 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005049 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005050)
5051
5052xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005053 name = "f32_vlrelu_test",
5054 srcs = [
5055 "test/f32-vlrelu.cc",
5056 "test/vunary-microkernel-tester.h",
5057 ] + MICROKERNEL_TEST_HDRS,
5058 deps = MICROKERNEL_TEST_DEPS,
5059)
5060
5061xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005062 name = "f32_vneg_test",
5063 srcs = [
5064 "test/f32-vneg.cc",
5065 "test/vunary-microkernel-tester.h",
5066 ] + MICROKERNEL_TEST_HDRS,
5067 deps = MICROKERNEL_TEST_DEPS,
5068)
5069
5070xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005071 name = "f32_vrndne_test",
5072 srcs = [
5073 "test/f32-vrndne.cc",
5074 "test/vunary-microkernel-tester.h",
5075 ] + MICROKERNEL_TEST_HDRS,
5076 deps = MICROKERNEL_TEST_DEPS,
5077)
5078
5079xnnpack_unit_test(
5080 name = "f32_vrndz_test",
5081 srcs = [
5082 "test/f32-vrndz.cc",
5083 "test/vunary-microkernel-tester.h",
5084 ] + MICROKERNEL_TEST_HDRS,
5085 deps = MICROKERNEL_TEST_DEPS,
5086)
5087
5088xnnpack_unit_test(
5089 name = "f32_vrndu_test",
5090 srcs = [
5091 "test/f32-vrndu.cc",
5092 "test/vunary-microkernel-tester.h",
5093 ] + MICROKERNEL_TEST_HDRS,
5094 deps = MICROKERNEL_TEST_DEPS,
5095)
5096
5097xnnpack_unit_test(
5098 name = "f32_vrndd_test",
5099 srcs = [
5100 "test/f32-vrndd.cc",
5101 "test/vunary-microkernel-tester.h",
5102 ] + MICROKERNEL_TEST_HDRS,
5103 deps = MICROKERNEL_TEST_DEPS,
5104)
5105
5106xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005107 name = "f32_vscale_test",
5108 srcs = [
5109 "test/f32-vscale.cc",
5110 "test/vscale-microkernel-tester.h",
5111 ] + MICROKERNEL_TEST_HDRS,
5112 deps = MICROKERNEL_TEST_DEPS,
5113)
5114
5115xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005116 name = "f32_vscaleexpminusmax_test",
5117 srcs = [
5118 "test/f32-vscaleexpminusmax.cc",
5119 "test/vscaleexpminusmax-microkernel-tester.h",
5120 ] + MICROKERNEL_TEST_HDRS,
5121 deps = MICROKERNEL_TEST_DEPS,
5122)
5123
5124xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005125 name = "f32_vscaleextexp_test",
5126 srcs = [
5127 "test/f32-vscaleextexp.cc",
5128 "test/vscaleextexp-microkernel-tester.h",
5129 ] + MICROKERNEL_TEST_HDRS,
5130 deps = MICROKERNEL_TEST_DEPS,
5131)
5132
5133xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005134 name = "f32_vsqr_test",
5135 srcs = [
5136 "test/f32-vsqr.cc",
5137 "test/vunary-microkernel-tester.h",
5138 ] + MICROKERNEL_TEST_HDRS,
5139 deps = MICROKERNEL_TEST_DEPS,
5140)
5141
5142xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005143 name = "f32_vsqrdiff_test",
5144 srcs = [
5145 "test/f32-vsqrdiff.cc",
5146 "test/vbinary-microkernel-tester.h",
5147 ] + MICROKERNEL_TEST_HDRS,
5148 deps = MICROKERNEL_TEST_DEPS,
5149)
5150
5151xnnpack_unit_test(
5152 name = "f32_vsqrdiffc_test",
5153 srcs = [
5154 "test/f32-vsqrdiffc.cc",
5155 "test/vbinaryc-microkernel-tester.h",
5156 ] + MICROKERNEL_TEST_HDRS,
5157 deps = MICROKERNEL_TEST_DEPS,
5158)
5159
5160xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005161 name = "f32_vsqrt_test",
5162 srcs = [
5163 "test/f32-vsqrt.cc",
5164 "test/vunary-microkernel-tester.h",
5165 ] + MICROKERNEL_TEST_HDRS,
5166 deps = MICROKERNEL_TEST_DEPS,
5167)
5168
5169xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005170 name = "f32_vsub_test",
5171 srcs = [
5172 "test/f32-vsub.cc",
5173 "test/vbinary-microkernel-tester.h",
5174 ] + MICROKERNEL_TEST_HDRS,
5175 deps = MICROKERNEL_TEST_DEPS,
5176)
5177
5178xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005179 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005180 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005181 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005182 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005183 ] + MICROKERNEL_TEST_HDRS,
5184 deps = MICROKERNEL_TEST_DEPS,
5185)
5186
5187xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005188 name = "f32_vsubc_test",
5189 srcs = [
5190 "test/f32-vsubc.cc",
5191 "test/vbinaryc-microkernel-tester.h",
5192 ] + MICROKERNEL_TEST_HDRS,
5193 deps = MICROKERNEL_TEST_DEPS,
5194)
5195
5196xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005197 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005198 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005199 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005200 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005201 ] + MICROKERNEL_TEST_HDRS,
5202 deps = MICROKERNEL_TEST_DEPS,
5203)
5204
5205xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005206 name = "f32_vrsubc_test",
5207 srcs = [
5208 "test/f32-vrsubc.cc",
5209 "test/vbinaryc-microkernel-tester.h",
5210 ] + MICROKERNEL_TEST_HDRS,
5211 deps = MICROKERNEL_TEST_DEPS,
5212)
5213
5214xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005215 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005216 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005217 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005218 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005219 ] + MICROKERNEL_TEST_HDRS,
5220 deps = MICROKERNEL_TEST_DEPS,
5221)
5222
5223xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005224 name = "qs8_dwconv_minmax_test",
5225 srcs = [
5226 "test/qs8-dwconv-minmax.cc",
5227 "test/dwconv-microkernel-tester.h",
5228 "src/xnnpack/AlignedAllocator.h",
5229 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5230 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5231)
5232
5233xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005234 name = "qs8_gemm_minmax_test",
5235 srcs = [
5236 "test/qs8-gemm-minmax.cc",
5237 "test/gemm-microkernel-tester.h",
5238 "src/xnnpack/AlignedAllocator.h",
5239 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5240 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5241)
5242
5243xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005244 name = "qs8_igemm_minmax_test",
5245 srcs = [
5246 "test/qs8-igemm-minmax.cc",
5247 "test/gemm-microkernel-tester.h",
5248 "src/xnnpack/AlignedAllocator.h",
5249 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5250 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5251)
5252
5253xnnpack_unit_test(
5254 name = "qs8_requantization_test",
5255 srcs = [
5256 "src/xnnpack/requantization-stubs.h",
5257 "test/qs8-requantization.cc",
5258 "test/requantization-tester.h",
5259 ] + MICROKERNEL_TEST_HDRS,
5260 deps = MICROKERNEL_TEST_DEPS,
5261)
5262
5263xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005264 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005265 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005266 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005267 "test/avgpool-microkernel-tester.h",
5268 "src/xnnpack/AlignedAllocator.h",
5269 ] + MICROKERNEL_TEST_HDRS,
5270 deps = MICROKERNEL_TEST_DEPS,
5271)
5272
5273xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005274 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005275 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005276 "test/qu8-dwconv-minmax.cc",
5277 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005278 "src/xnnpack/AlignedAllocator.h",
5279 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005280 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005281)
5282
5283xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005284 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005285 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005286 "test/qu8-igemm-minmax.cc",
5287 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005288 "src/xnnpack/AlignedAllocator.h",
5289 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005290 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005291)
5292
5293xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005294 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005295 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005296 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005297 "test/gavgpool-microkernel-tester.h",
5298 "src/xnnpack/AlignedAllocator.h",
5299 ] + MICROKERNEL_TEST_HDRS,
5300 deps = MICROKERNEL_TEST_DEPS,
5301)
5302
5303xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005304 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005305 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005306 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005307 "test/gemm-microkernel-tester.h",
5308 "src/xnnpack/AlignedAllocator.h",
5309 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005310 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005311)
5312
5313xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005314 name = "qu8_requantization_test",
5315 srcs = [
5316 "src/xnnpack/requantization-stubs.h",
5317 "test/qu8-requantization.cc",
5318 "test/requantization-tester.h",
5319 ] + MICROKERNEL_TEST_HDRS,
5320 deps = MICROKERNEL_TEST_DEPS,
5321)
5322
5323xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005324 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005325 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005326 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005327 "test/vadd-microkernel-tester.h",
5328 ] + MICROKERNEL_TEST_HDRS,
5329 deps = MICROKERNEL_TEST_DEPS,
5330)
5331
5332xnnpack_unit_test(
5333 name = "u8_clamp_test",
5334 srcs = [
5335 "test/u8-clamp.cc",
5336 "test/clamp-microkernel-tester.h",
5337 ] + MICROKERNEL_TEST_HDRS,
5338 deps = MICROKERNEL_TEST_DEPS,
5339)
5340
5341xnnpack_unit_test(
5342 name = "u8_lut32norm_test",
5343 srcs = [
5344 "test/u8-lut32norm.cc",
5345 "test/lut-norm-microkernel-tester.h",
5346 ] + MICROKERNEL_TEST_HDRS,
5347 deps = MICROKERNEL_TEST_DEPS,
5348)
5349
5350xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005351 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005352 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005353 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005354 "test/maxpool-microkernel-tester.h",
5355 ] + MICROKERNEL_TEST_HDRS,
5356 deps = MICROKERNEL_TEST_DEPS,
5357)
5358
5359xnnpack_unit_test(
5360 name = "u8_rmax_test",
5361 srcs = [
5362 "test/u8-rmax.cc",
5363 "test/rmax-microkernel-tester.h",
5364 ] + MICROKERNEL_TEST_HDRS,
5365 deps = MICROKERNEL_TEST_DEPS,
5366)
5367
5368xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005369 name = "x32_fill_test",
5370 srcs = [
5371 "test/x32-fill.cc",
5372 "test/fill-microkernel-tester.h",
5373 ] + MICROKERNEL_TEST_HDRS,
5374 deps = MICROKERNEL_TEST_DEPS,
5375)
5376
5377xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005378 name = "x32_packx_test",
5379 srcs = [
5380 "test/x32-packx.cc",
5381 "test/pack-microkernel-tester.h",
5382 "src/xnnpack/AlignedAllocator.h",
5383 ] + MICROKERNEL_TEST_HDRS,
5384 deps = MICROKERNEL_TEST_DEPS,
5385)
5386
5387xnnpack_unit_test(
5388 name = "x32_pad_test",
5389 srcs = [
5390 "test/x32-pad.cc",
5391 "test/pad-microkernel-tester.h",
5392 ] + MICROKERNEL_TEST_HDRS,
5393 deps = MICROKERNEL_TEST_DEPS,
5394)
5395
5396xnnpack_unit_test(
5397 name = "x32_unpool_test",
5398 srcs = [
5399 "test/x32-unpool.cc",
5400 "test/unpool-microkernel-tester.h",
5401 ] + MICROKERNEL_TEST_HDRS,
5402 deps = MICROKERNEL_TEST_DEPS,
5403)
5404
5405xnnpack_unit_test(
5406 name = "x32_zip_test",
5407 srcs = [
5408 "test/x32-zip.cc",
5409 "test/zip-microkernel-tester.h",
5410 ] + MICROKERNEL_TEST_HDRS,
5411 deps = MICROKERNEL_TEST_DEPS,
5412)
5413
5414xnnpack_unit_test(
5415 name = "x8_lut_test",
5416 srcs = [
5417 "test/x8-lut.cc",
5418 "test/lut-microkernel-tester.h",
5419 ] + MICROKERNEL_TEST_HDRS,
5420 deps = MICROKERNEL_TEST_DEPS,
5421)
5422
5423xnnpack_unit_test(
5424 name = "x8_zip_test",
5425 srcs = [
5426 "test/x8-zip.cc",
5427 "test/zip-microkernel-tester.h",
5428 ] + MICROKERNEL_TEST_HDRS,
5429 deps = MICROKERNEL_TEST_DEPS,
5430)
5431
Marat Dukhan20c3b922020-03-10 03:45:06 -07005432########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005433
5434xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005435 name = "operator_size_test",
5436 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005437 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005438)
5439
Marat Dukhan20c3b922020-03-10 03:45:06 -07005440xnnpack_binary(
5441 name = "subgraph_size_test",
5442 srcs = ["test/subgraph-size.c"],
5443 deps = [":XNNPACK"],
5444)
5445
5446########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005447
5448xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005449 name = "abs_nc_test",
5450 srcs = [
5451 "test/abs-nc.cc",
5452 "test/abs-operator-tester.h",
5453 ],
5454 deps = OPERATOR_TEST_DEPS,
5455)
5456
5457xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005458 name = "add_nd_test",
5459 srcs = [
5460 "test/add-nd.cc",
5461 "test/binary-elementwise-operator-tester.h",
5462 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005463 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005464)
5465
5466xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005467 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005468 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005469 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005470 "test/argmax-pooling-operator-tester.h",
5471 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005472 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005473)
5474
5475xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005476 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005477 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005478 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005479 "test/average-pooling-operator-tester.h",
5480 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005481 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005482)
5483
5484xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005485 name = "bankers_rounding_nc_test",
5486 srcs = [
5487 "test/bankers-rounding-nc.cc",
5488 "test/bankers-rounding-operator-tester.h",
5489 ],
5490 deps = OPERATOR_TEST_DEPS,
5491)
5492
5493xnnpack_unit_test(
5494 name = "ceiling_nc_test",
5495 srcs = [
5496 "test/ceiling-nc.cc",
5497 "test/ceiling-operator-tester.h",
5498 ],
5499 deps = OPERATOR_TEST_DEPS,
5500)
5501
5502xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005503 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005504 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005505 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005506 "test/channel-shuffle-operator-tester.h",
5507 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005508 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005509)
5510
5511xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005512 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005513 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005514 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005515 "test/clamp-operator-tester.h",
5516 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005517 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005518)
5519
5520xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005521 name = "constant_pad_nd_test",
5522 srcs = [
5523 "test/constant-pad-nd.cc",
5524 "test/constant-pad-operator-tester.h",
5525 ],
5526 deps = OPERATOR_TEST_DEPS,
5527)
5528
5529xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005530 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005531 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005532 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005533 "test/convolution-operator-tester.h",
5534 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005535 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005536)
5537
5538xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005539 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005540 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005541 "test/convolution-nchw.cc",
5542 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005543 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005544 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005545)
5546
5547xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005548 name = "copy_nc_test",
5549 srcs = [
5550 "test/copy-nc.cc",
5551 "test/copy-operator-tester.h",
5552 ],
5553 deps = OPERATOR_TEST_DEPS,
5554)
5555
5556xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005557 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005558 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005559 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005560 "test/deconvolution-operator-tester.h",
5561 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005562 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005563)
5564
5565xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005566 name = "divide_nd_test",
5567 srcs = [
5568 "test/binary-elementwise-operator-tester.h",
5569 "test/divide-nd.cc",
5570 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005571 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005572)
5573
5574xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005575 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005576 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005577 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005578 "test/fully-connected-operator-tester.h",
5579 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005580 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005581)
5582
5583xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005584 name = "floor_nc_test",
5585 srcs = [
5586 "test/floor-nc.cc",
5587 "test/floor-operator-tester.h",
5588 ],
5589 deps = OPERATOR_TEST_DEPS,
5590)
5591
5592xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005593 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005594 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005595 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005596 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005597 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005598 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005599)
5600
5601xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005602 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005603 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005604 "test/global-average-pooling-ncw.cc",
5605 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005606 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005607 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005608)
5609
5610xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005611 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005612 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005613 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005614 "test/hardswish-operator-tester.h",
5615 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005616 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005617)
5618
5619xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005620 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005621 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005622 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005623 "test/leaky-relu-operator-tester.h",
5624 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005625 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005626)
5627
5628xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005629 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005630 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005631 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005632 "test/max-pooling-operator-tester.h",
5633 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005634 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005635)
5636
5637xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005638 name = "maximum_nd_test",
5639 srcs = [
5640 "test/binary-elementwise-operator-tester.h",
5641 "test/maximum-nd.cc",
5642 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005643 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005644)
5645
5646xnnpack_unit_test(
5647 name = "minimum_nd_test",
5648 srcs = [
5649 "test/binary-elementwise-operator-tester.h",
5650 "test/minimum-nd.cc",
5651 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005652 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005653)
5654
5655xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005656 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005657 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005658 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005659 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005660 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005661 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005662)
5663
5664xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005665 name = "negate_nc_test",
5666 srcs = [
5667 "test/negate-nc.cc",
5668 "test/negate-operator-tester.h",
5669 ],
5670 deps = OPERATOR_TEST_DEPS,
5671)
5672
5673xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005674 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005675 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005676 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005677 "test/prelu-operator-tester.h",
5678 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005679 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005680)
5681
5682xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005683 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005684 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005685 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005686 "test/resize-bilinear-operator-tester.h",
5687 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005688 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005689)
5690
5691xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005692 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005693 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005694 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005695 "test/sigmoid-operator-tester.h",
5696 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005697 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005698)
5699
5700xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005701 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005702 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005703 "test/softmax-nc.cc",
5704 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005705 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005706 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005707)
5708
5709xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005710 name = "square_nc_test",
5711 srcs = [
5712 "test/square-nc.cc",
5713 "test/square-operator-tester.h",
5714 ],
5715 deps = OPERATOR_TEST_DEPS,
5716)
5717
5718xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005719 name = "square_root_nc_test",
5720 srcs = [
5721 "test/square-root-nc.cc",
5722 "test/square-root-operator-tester.h",
5723 ],
5724 deps = OPERATOR_TEST_DEPS,
5725)
5726
5727xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005728 name = "squared_difference_nd_test",
5729 srcs = [
5730 "test/binary-elementwise-operator-tester.h",
5731 "test/squared-difference-nd.cc",
5732 ],
5733 deps = OPERATOR_TEST_DEPS,
5734)
5735
5736xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005737 name = "subtract_nd_test",
5738 srcs = [
5739 "test/binary-elementwise-operator-tester.h",
5740 "test/subtract-nd.cc",
5741 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005742 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005743)
5744
5745xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005746 name = "truncation_nc_test",
5747 srcs = [
5748 "test/truncation-nc.cc",
5749 "test/truncation-operator-tester.h",
5750 ],
5751 deps = OPERATOR_TEST_DEPS,
5752)
5753
5754xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005755 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005756 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005757 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005758 "test/unpooling-operator-tester.h",
5759 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005760 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005761)
5762
Chao Mei6ddfc602020-05-13 22:29:36 -07005763############################### Misc unit tests ###############################
5764
5765xnnpack_unit_test(
5766 name = "memory_planner_test",
5767 srcs = [
5768 "test/memory-planner-test.cc",
5769 ],
5770 deps = [
5771 ":XNNPACK",
5772 ":memory_planner",
5773 ],
5774)
5775
Marat Dukhan08c4a432019-10-03 09:29:21 -07005776############################# Build configurations #############################
5777
Marat Dukhanb8642352019-10-30 15:43:02 -07005778# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005779config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005780 name = "xnn_enable_assembly_explicit_true",
5781 define_values = {"xnn_enable_assembly": "true"},
5782)
5783
5784# Disables usage of assembly kernels.
5785config_setting(
5786 name = "xnn_enable_assembly_explicit_false",
5787 define_values = {"xnn_enable_assembly": "false"},
5788)
5789
Marat Dukhan9de90e02020-06-18 16:04:12 -07005790# Enables usage of sparse inference.
5791config_setting(
5792 name = "xnn_enable_sparse_explicit_true",
5793 define_values = {"xnn_enable_sparse": "true"},
5794)
5795
5796# Disables usage of sparse inference.
5797config_setting(
5798 name = "xnn_enable_sparse_explicit_false",
5799 define_values = {"xnn_enable_sparse": "false"},
5800)
5801
Marat Dukhan05702cf2020-03-26 15:41:33 -07005802# Disables usage of HMP-aware optimizations.
5803config_setting(
5804 name = "xnn_enable_hmp_explicit_false",
5805 define_values = {"xnn_enable_hmp": "false"},
5806)
5807
Chao Mei6ddfc602020-05-13 22:29:36 -07005808# Enable usage of optimized memory allocation
5809config_setting(
5810 name = "xnn_enable_memopt_explicit_true",
5811 define_values = {"xnnpack_enable_memopt": "true"},
5812)
5813
5814# Disable usage of optimized memory allocation
5815config_setting(
5816 name = "xnn_enable_memopt_explicit_false",
5817 define_values = {"xnnpack_enable_memopt": "false"},
5818)
5819
Marat Dukhanb8642352019-10-30 15:43:02 -07005820# Builds with -c dbg
5821config_setting(
5822 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005823 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005824 "compilation_mode": "dbg",
5825 },
5826)
5827
5828# Builds with -c opt
5829config_setting(
5830 name = "optimized_build",
5831 values = {
5832 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005833 },
5834)
5835
5836config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005837 name = "linux_k8",
5838 values = {"cpu": "k8"},
5839)
5840
5841config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005842 name = "linux_arm",
5843 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005844)
5845
5846config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005847 name = "linux_armeabi",
5848 values = {"cpu": "armeabi"},
5849)
5850
5851config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005852 name = "linux_armhf",
5853 values = {"cpu": "armhf"},
5854)
5855
5856config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005857 name = "linux_armv7a",
5858 values = {"cpu": "armv7a"},
5859)
5860
5861config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005862 name = "linux_aarch64",
5863 values = {"cpu": "aarch64"},
5864)
5865
5866config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005867 name = "android",
5868 values = {"crosstool_top": "//external:android/crosstool"},
5869)
5870
5871config_setting(
5872 name = "android_armv7",
5873 values = {
5874 "crosstool_top": "//external:android/crosstool",
5875 "cpu": "armeabi-v7a",
5876 },
5877)
5878
5879config_setting(
5880 name = "android_arm64",
5881 values = {
5882 "crosstool_top": "//external:android/crosstool",
5883 "cpu": "arm64-v8a",
5884 },
5885)
5886
5887config_setting(
5888 name = "android_x86",
5889 values = {
5890 "crosstool_top": "//external:android/crosstool",
5891 "cpu": "x86",
5892 },
5893)
5894
5895config_setting(
5896 name = "android_x86_64",
5897 values = {
5898 "crosstool_top": "//external:android/crosstool",
5899 "cpu": "x86_64",
5900 },
5901)
5902
5903config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005904 name = "windows_x86_64",
5905 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005906)
5907
5908config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005909 name = "windows_x86_64_clang",
5910 values = {
5911 "compiler": "clang-cl",
5912 "cpu": "x64_windows",
5913 },
5914)
5915
5916config_setting(
5917 name = "windows_x86_64_mingw",
5918 values = {
5919 "compiler": "mingw-gcc",
5920 "cpu": "x64_windows",
5921 },
5922)
5923
5924config_setting(
5925 name = "windows_x86_64_msys",
5926 values = {
5927 "compiler": "msys-gcc",
5928 "cpu": "x64_windows",
5929 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005930)
5931
5932config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005933 name = "macos_x86_64",
5934 values = {
5935 "apple_platform_type": "macos",
5936 "cpu": "darwin",
5937 },
5938)
5939
5940config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005941 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005942 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005943)
5944
5945config_setting(
5946 name = "emscripten_wasm",
5947 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005948 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005949 "cpu": "wasm",
5950 },
5951)
5952
5953config_setting(
5954 name = "emscripten_wasmsimd",
5955 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005956 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005957 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005958 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005959 },
5960)
5961
5962config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005963 name = "ios_armv7",
5964 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005965 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005966 "cpu": "ios_armv7",
5967 },
5968)
5969
5970config_setting(
5971 name = "ios_arm64",
5972 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005973 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005974 "cpu": "ios_arm64",
5975 },
5976)
5977
5978config_setting(
5979 name = "ios_arm64e",
5980 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005981 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005982 "cpu": "ios_arm64e",
5983 },
5984)
5985
5986config_setting(
5987 name = "ios_x86",
5988 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005989 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005990 "cpu": "ios_i386",
5991 },
5992)
5993
5994config_setting(
5995 name = "ios_x86_64",
5996 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005997 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005998 "cpu": "ios_x86_64",
5999 },
6000)
6001
6002config_setting(
6003 name = "watchos_armv7k",
6004 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006005 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006006 "cpu": "watchos_armv7k",
6007 },
6008)
6009
6010config_setting(
6011 name = "watchos_arm64_32",
6012 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006013 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006014 "cpu": "watchos_arm64_32",
6015 },
6016)
6017
6018config_setting(
6019 name = "watchos_x86",
6020 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006021 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006022 "cpu": "watchos_i386",
6023 },
6024)
6025
6026config_setting(
6027 name = "watchos_x86_64",
6028 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006029 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006030 "cpu": "watchos_x86_64",
6031 },
6032)
6033
6034config_setting(
6035 name = "tvos_arm64",
6036 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006037 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006038 "cpu": "tvos_arm64",
6039 },
6040)
6041
6042config_setting(
6043 name = "tvos_x86_64",
6044 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006045 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006046 "cpu": "tvos_x86_64",
6047 },
6048)