blob: 04cd3397db01559a305e560233120072f0db6d7f [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 Dukhan08b7a972020-07-14 18:17:29 -0700361 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
362 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
363 "src/qu8-dwconv/up1x9-minmax-scalar.c",
364 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
365 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
366 "src/qu8-gemm/2x2-minmax-scalar.c",
367 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700368 "src/qu8-requantization/fp32-scalar-lrintf.c",
369 "src/qu8-requantization/fp32-scalar-magic.c",
370 "src/qu8-requantization/precise-scalar-signed64.c",
371 "src/qu8-requantization/precise-scalar-unsigned32.c",
372 "src/qu8-requantization/precise-scalar-unsigned64.c",
373 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700374 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700375 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700376 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700377 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700378 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700379 "src/x32-fill/scalar-float.c",
380 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700381 "src/x32-packx/x2-scalar.c",
382 "src/x32-packx/x3-scalar.c",
383 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700384 "src/x32-pad/scalar-int.c",
385 "src/x32-pad/scalar-float.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700386 "src/x32-unpool/scalar.c",
387 "src/x32-zip/x2-scalar.c",
388 "src/x32-zip/x3-scalar.c",
389 "src/x32-zip/x4-scalar.c",
390 "src/x32-zip/xm-scalar.c",
391 "src/x8-lut/scalar.c",
392 "src/x8-zip/x2-scalar.c",
393 "src/x8-zip/x3-scalar.c",
394 "src/x8-zip/x4-scalar.c",
395 "src/x8-zip/xm-scalar.c",
396]
397
Marat Dukhan436ebe62019-12-04 15:10:12 -0800398WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700399 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
400 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700401 "src/f32-clamp/gen/wasm-x1.c",
402 "src/f32-clamp/gen/wasm-x2.c",
403 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700404 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
405 "src/f32-dwconv/gen/up1x4-wasm.c",
406 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
407 "src/f32-dwconv/gen/up1x9-wasm.c",
408 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
409 "src/f32-dwconv/gen/up1x25-wasm.c",
410 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
411 "src/f32-dwconv/gen/up2x4-wasm.c",
412 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
413 "src/f32-dwconv/gen/up2x9-wasm.c",
414 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
415 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700416 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
417 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
418 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
419 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700420 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
421 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700422 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
423 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
424 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
425 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700426 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
427 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700428 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
429 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700430 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
431 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
432 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700433 "src/f32-gemm/gen/1x4-wasm.c",
434 "src/f32-gemm/gen/2x4-wasm.c",
435 "src/f32-gemm/gen/4x2-wasm.c",
436 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700437 "src/f32-gemm/gen/1x4-relu-wasm.c",
438 "src/f32-gemm/gen/2x4-relu-wasm.c",
439 "src/f32-gemm/gen/4x2-relu-wasm.c",
440 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700441 "src/f32-gemm/gen/1x4-minmax-wasm.c",
442 "src/f32-gemm/gen/2x4-minmax-wasm.c",
443 "src/f32-gemm/gen/4x2-minmax-wasm.c",
444 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700445 "src/f32-hswish/gen/hswish-wasm-x1.c",
446 "src/f32-hswish/gen/hswish-wasm-x2.c",
447 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700448 "src/f32-igemm/gen/1x4-wasm.c",
449 "src/f32-igemm/gen/2x4-wasm.c",
450 "src/f32-igemm/gen/4x2-wasm.c",
451 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700452 "src/f32-igemm/gen/1x4-relu-wasm.c",
453 "src/f32-igemm/gen/2x4-relu-wasm.c",
454 "src/f32-igemm/gen/4x2-relu-wasm.c",
455 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700456 "src/f32-igemm/gen/1x4-minmax-wasm.c",
457 "src/f32-igemm/gen/2x4-minmax-wasm.c",
458 "src/f32-igemm/gen/4x2-minmax-wasm.c",
459 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700460 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
461 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
462 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700463 "src/f32-prelu/gen/wasm-2x1.c",
464 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700465 "src/f32-relu/gen/wasm-x1.c",
466 "src/f32-relu/gen/wasm-x2.c",
467 "src/f32-relu/gen/wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700468 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
469 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
470 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
471 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
472 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
473 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
474 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
475 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
476 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
477 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
478 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
479 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800480 "src/f32-vbinary/gen/vmax-wasm-x1.c",
481 "src/f32-vbinary/gen/vmax-wasm-x2.c",
482 "src/f32-vbinary/gen/vmax-wasm-x4.c",
483 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
484 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
485 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
486 "src/f32-vbinary/gen/vmin-wasm-x1.c",
487 "src/f32-vbinary/gen/vmin-wasm-x2.c",
488 "src/f32-vbinary/gen/vmin-wasm-x4.c",
489 "src/f32-vbinary/gen/vminc-wasm-x1.c",
490 "src/f32-vbinary/gen/vminc-wasm-x2.c",
491 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700492 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
493 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
494 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
495 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
496 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
497 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
498 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
499 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
500 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
501 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
502 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
503 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
504 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
505 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
506 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
507 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
508 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
509 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700510 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
511 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
512 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700513 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
514 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
515 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800516]
517
Marat Dukhan290055c2020-06-09 12:24:29 -0700518WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700519 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
520 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
521 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700522 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
523 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
524 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
525 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700526 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
527 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
528 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
529 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700530 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
531 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
532 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
533 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
534 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
535 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700536 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
537 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
538 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
539 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
540 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
541 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
542 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
543 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
544 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
545 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
546 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
547 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
548 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
549 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
550 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
551 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
552 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
553 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
554 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
555 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
556 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
557 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
558 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
559 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700560 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
561 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
562 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700563 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700564 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
565 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
566 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
567 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700568 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
569 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
570 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
571 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
572 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
573 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
574 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
575 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
576 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
577 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
578 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
579 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
580 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
581 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
582 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
583 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
584 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
585 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
586 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
587 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
588 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
589 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
590 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
591 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
592 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
593 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
594 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
595 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
596 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
597 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700598 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
599 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700600 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
601 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
602 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
603 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
604 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
605 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
606 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
607 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
608 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
609 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
610 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
611 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
612 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
613 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
614 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
615 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
616 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
617 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
618 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
619 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
620 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
621 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
622 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
623 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
624 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
625 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
626 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
627 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
628 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
629 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700630 "src/f32-ibilinear/gen/wasmsimd-c4.c",
631 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700632 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
633 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
634 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700635 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700636 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
637 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
638 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
639 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700640 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
641 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
642 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
643 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
644 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
645 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
646 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
647 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
648 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
649 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
650 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
651 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
652 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
653 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
654 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
655 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
656 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
657 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
658 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
659 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
660 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
661 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
662 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
663 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
664 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
665 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
666 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
667 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
668 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
669 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700670 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
671 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700672 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
673 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
674 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
675 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700676 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
677 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700678 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
679 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
680 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700681 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
682 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700683 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
684 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
685 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
686 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700687 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
688 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700689 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
690 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
691 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700692 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
693 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700694 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
695 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
696 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
697 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
698 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
699 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
700 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700701 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
702 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700703 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
704 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
705 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
706 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700707 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
708 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
709 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
710 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
711 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
712 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
713 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
714 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
715 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
716 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
717 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
718 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700719 "src/f32-relu/gen/wasmsimd-x4.c",
720 "src/f32-relu/gen/wasmsimd-x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700721 "src/f32-rmax/wasmsimd-arm.c",
722 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700723 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
724 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
725 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
726 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
727 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
728 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700729 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
730 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
731 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
732 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
733 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
734 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700735 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
736 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
737 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
738 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700739 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
740 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700741 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
742 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
743 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
744 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700745 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
746 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700747 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
748 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
749 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
750 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700751 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
752 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700753 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
754 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
755 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
756 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700757 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
758 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700759 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
760 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
761 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
762 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700763 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
764 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
765 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
766 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700767 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
768 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
769 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
770 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700771 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
772 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
773 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
774 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700775 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
776 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
777 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
778 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700779 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
780 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700781 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
782 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
783 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
784 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700785 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
786 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700787 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
788 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
789 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
790 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700791 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
792 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700793 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
794 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
795 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
796 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700797 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
798 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700799 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
800 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700801 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
802 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700803 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
804 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
805 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
806 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700807 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
808 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700809 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
810 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
811 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
812 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700813 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
814 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700815 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
816 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
817 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
818 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700819 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
820 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
821 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
822 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700823 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
824 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
825 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
826 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
827 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
828 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
829 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
830 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
831 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
832 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
833 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
834 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
835 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
836 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700837 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
838 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700839 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
840 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
841 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
842 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
843 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
844 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700845 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700846 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700847 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700848 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700849 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700850 "src/x32-zip/x2-wasmsimd.c",
851 "src/x32-zip/x3-wasmsimd.c",
852 "src/x32-zip/x4-wasmsimd.c",
853 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700854 "src/math/roundne-wasmsimd-addsub.c",
855 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700856 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700857 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700858 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700859 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700860 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -0700861 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
862 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700863]
864
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800865PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -0700866 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700867 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700868 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700869 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700870 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6f278b52020-06-10 16:13:11 -0700871 "src/f32-gavgpool-cw/psimd-x4.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700872 "src/f32-spmm/gen/4x1-minmax-psimd.c",
873 "src/f32-spmm/gen/8x1-minmax-psimd.c",
874 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700875]
876
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800877PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700878 "src/qu8-requantization/precise-psimd.c",
879 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800880]
881
Marat Dukhan08c4a432019-10-03 09:29:21 -0700882# ISA-specific micro-kernels
883NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -0700884 "src/f32-argmaxpool/4x-neon-c4.c",
885 "src/f32-argmaxpool/9p8x-neon-c4.c",
886 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700887 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
888 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700889 "src/f32-clamp/gen/neon-x4.c",
890 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700891 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
892 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700893 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
894 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700895 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
896 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700897 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
898 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700899 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
900 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
901 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
902 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700903 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
904 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
905 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
906 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700907 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
908 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
909 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
910 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700911 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700912 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
913 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700914 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
915 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
916 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
917 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
918 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
919 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
920 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
921 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
922 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
923 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
924 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
925 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
926 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
927 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
928 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
929 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
930 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
931 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
932 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
933 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
934 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
935 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
936 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
937 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
938 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
939 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
940 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
941 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
942 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
943 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
944 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700945 "src/f32-hswish/gen/hswish-neon-x4.c",
946 "src/f32-hswish/gen/hswish-neon-x8.c",
947 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700948 "src/f32-ibilinear/gen/neon-c4.c",
949 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700950 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
951 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
952 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
953 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
954 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
955 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
956 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
957 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
958 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
959 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
960 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
961 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
962 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
963 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
964 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
965 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700966 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
967 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
968 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700969 "src/f32-ppmm/gen/4x8-minmax-neon.c",
970 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700971 "src/f32-prelu/gen/neon-1x4.c",
972 "src/f32-prelu/gen/neon-1x8.c",
973 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800974 "src/f32-prelu/gen/neon-2x4.c",
975 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700976 "src/f32-prelu/gen/neon-2x16.c",
977 "src/f32-prelu/gen/neon-4x4.c",
978 "src/f32-prelu/gen/neon-4x8.c",
979 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -0800980 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
981 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
982 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
983 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
984 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
985 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
986 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
987 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
988 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
989 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
990 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
991 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
992 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
993 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
994 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
995 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
996 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
997 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
998 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
999 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1000 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1001 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1002 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1003 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001004 "src/f32-relu/gen/neon-x4.c",
1005 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001006 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001007 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001008 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1009 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1010 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1011 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1012 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1013 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1014 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1015 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1016 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1017 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1018 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1019 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1020 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1021 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1022 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1023 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1024 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1025 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001026 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1027 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1028 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1029 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001030 "src/f32-vbinary/gen/vmax-neon-x4.c",
1031 "src/f32-vbinary/gen/vmax-neon-x8.c",
1032 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1033 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1034 "src/f32-vbinary/gen/vmin-neon-x4.c",
1035 "src/f32-vbinary/gen/vmin-neon-x8.c",
1036 "src/f32-vbinary/gen/vminc-neon-x4.c",
1037 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001038 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1039 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1040 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1041 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1042 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1043 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001044 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1045 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1046 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1047 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001048 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1049 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1050 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1051 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001052 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1053 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001054 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1055 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001056 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1057 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1058 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1059 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1060 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1061 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1062 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1063 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001064 "src/f32-vunary/gen/vabs-neon-x4.c",
1065 "src/f32-vunary/gen/vabs-neon-x8.c",
1066 "src/f32-vunary/gen/vneg-neon-x4.c",
1067 "src/f32-vunary/gen/vneg-neon-x8.c",
1068 "src/f32-vunary/gen/vsqr-neon-x4.c",
1069 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001070 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1071 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1072 "src/qu8-dwconv/up8x9-minmax-neon.c",
1073 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1074 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1075 "src/qu8-gemm/4x8-minmax-neon.c",
1076 "src/qu8-gemm/8x8-minmax-neon.c",
1077 "src/qu8-igemm/4x8-minmax-neon.c",
1078 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001079 "src/qu8-requantization/precise-neon.c",
1080 "src/qu8-requantization/fp32-neon.c",
1081 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001082 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001083 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001084 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001085 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001086 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001087 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001088 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001089 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001090 "src/x32-zip/x2-neon.c",
1091 "src/x32-zip/x3-neon.c",
1092 "src/x32-zip/x4-neon.c",
1093 "src/x32-zip/xm-neon.c",
1094 "src/x8-zip/x2-neon.c",
1095 "src/x8-zip/x3-neon.c",
1096 "src/x8-zip/x4-neon.c",
1097 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001098 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001099 "src/math/roundd-neon-addsub.c",
1100 "src/math/roundd-neon-cvt.c",
1101 "src/math/roundu-neon-addsub.c",
1102 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001103 "src/math/roundz-neon-addsub.c",
1104 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001105 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001106 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1107 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1108 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1109 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1110 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1111 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001112 "src/math/sqrt-neon-nr1rsqrts.c",
1113 "src/math/sqrt-neon-nr2rsqrts.c",
1114 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001115]
1116
1117NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001118 "src/f32-ibilinear/gen/neonfma-c4.c",
1119 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001120 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1121 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1122 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1123 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1124 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1125 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1126 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1127 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1128 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001129 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1130 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1131 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1132 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001133 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1134 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1135 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1136 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001137 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1138 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1139 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1140 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001141 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1142 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1143 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1144 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1145 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1146 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1147 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1148 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1149 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1150 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1151 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1152 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1153 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1154 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1155 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1156 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1157 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1158 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001159 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1160 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001161 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1162 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1163 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1164 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1165 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1166 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1167 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1168 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1169 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1170 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1171 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1172 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1173 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1174 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1175 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1176 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1177 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1178 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1179 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1180 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1181 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1182 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1183 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1184 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001185 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1186 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1187 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1188 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1189 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1190 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1191 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1192 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1193 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1194 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1195 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1196 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1197 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1198 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1199 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1200 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1201 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1202 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1203 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1204 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1205 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1206 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1207 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1208 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1209 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1210 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1211 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1212 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1213 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1214 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1215 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1216 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1217 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1218 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1219 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1220 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1221 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1222 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1223 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1224 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1225 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1226 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1227 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1228 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1229 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1230 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1231 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1232 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1233 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1234 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1235 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1236 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1237 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1238 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001239 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1240 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001241 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1242 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1243 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1244 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1245 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1246 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1247 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1248 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1249 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1250 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1251 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1252 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1253 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1254 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1255 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1256 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1257 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1258 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1259 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1260 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001261 "src/math/exp-neonfma-lut64-p2.c",
1262 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001263 "src/math/expminus-neonfma-lut2048-p1.c",
1264 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001265 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001266 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1267 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1268 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1269 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1270 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1271 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1272 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1273 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1274 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1275 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1276 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1277 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1278 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1279 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1280 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1281 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1282 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1283 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001284 "src/math/sqrt-neonfma-nr1fma.c",
1285 "src/math/sqrt-neonfma-nr2fma.c",
1286 "src/math/sqrt-neonfma-nr3fma.c",
1287 "src/math/sqrt-neonfma-nr2fma1adj.c",
1288 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001289]
1290
1291AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001292 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1293 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1294 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1295 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1296 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1297 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001298 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1299 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1300 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1301 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1302 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1303 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1304 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1305 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1306 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1307 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1308 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1309 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1310 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1311 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1312 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1313 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1314 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1315 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1316 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1317 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001318 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1319 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001320 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1321 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001322 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1323 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001324 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1325 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001326 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1327 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1328 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1329 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1330 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001331 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1332 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1333 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1334 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1335 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1336 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1337 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1338 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1339 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1340 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1341 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1342 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1343 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1344 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1345 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1346 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1347 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1348 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001349 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1350 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1351 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1352 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1353 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1354 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1355 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1356 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1357 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1358 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1359 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1360 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1361 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1362 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1363 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1364 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1365 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1366 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001367 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1368 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001369 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1370 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1371 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1372 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1373 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1374 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001375]
1376
Marat Dukhan8853b822020-05-07 12:19:01 -07001377NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001378 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1379 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1380 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1381 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1382 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1383 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1384 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1385 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001386 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001387 "src/math/roundd-neonv8.c",
1388 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001389 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001390]
1391
Marat Dukhan08c4a432019-10-03 09:29:21 -07001392AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001393 "src/f16-clamp/gen/neonfp16arith-x16.c",
1394 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001395 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1396 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1397 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1398 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1399 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1400 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1401 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1402 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1403 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1404 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1405 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1406 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001407 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1408 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001409 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1410 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001411 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001412 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001413 "src/f16-relu/gen/neonfp16arith-x16.c",
1414 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001415 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1416 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1417 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1418 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1419 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1420 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1421 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1422 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1423 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1424 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1425 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1426 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1427 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1428 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1429 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1430 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1431 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1432 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1433 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1434 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1435 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1436 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1437 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1438 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1439 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1440 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1441 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1442 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001443 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1444 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001445 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1446 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1447 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1448 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1449 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1450 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1451 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1452 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001453 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1454 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1455 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1456 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001457 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1458 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1459 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1460 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1461 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1462 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1463 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1464 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1465 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1466 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1467 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1468 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001469 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1470 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1471 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1472 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1473 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1474 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1475 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1476 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001477]
1478
1479SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001480 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1481 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001482 "src/f32-clamp/gen/sse-x4.c",
1483 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001484 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1485 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001486 "src/f32-dwconv-chw/3x3p1-sse.c",
1487 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001488 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1489 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1490 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1491 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1492 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1493 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1494 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1495 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1496 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1497 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1498 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1499 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001500 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001501 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1502 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001503 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1504 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1505 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1506 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1507 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1508 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1509 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1510 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1511 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1512 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1513 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1514 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1515 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001516 "src/f32-hswish/gen/hswish-sse-x4.c",
1517 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001518 "src/f32-ibilinear/gen/sse-c4.c",
1519 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001520 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1521 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1522 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1523 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1524 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1525 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1526 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001527 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1528 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1529 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001530 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001531 "src/f32-prelu/gen/sse-2x4.c",
1532 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001533 "src/f32-relu/gen/sse-x4.c",
1534 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001535 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001536 "src/f32-spmm/gen/4x1-minmax-sse.c",
1537 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001538 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001539 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1540 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1541 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1542 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1543 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1544 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1545 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1546 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001547 "src/f32-vbinary/gen/vmax-sse-x4.c",
1548 "src/f32-vbinary/gen/vmax-sse-x8.c",
1549 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1550 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1551 "src/f32-vbinary/gen/vmin-sse-x4.c",
1552 "src/f32-vbinary/gen/vmin-sse-x8.c",
1553 "src/f32-vbinary/gen/vminc-sse-x4.c",
1554 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001555 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1556 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1557 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1558 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1559 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1560 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1561 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1562 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001563 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1564 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1565 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1566 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001567 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1568 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1569 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1570 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001571 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1572 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001573 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1574 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001575 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1576 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001577 "src/f32-vunary/gen/vabs-sse-x4.c",
1578 "src/f32-vunary/gen/vabs-sse-x8.c",
1579 "src/f32-vunary/gen/vneg-sse-x4.c",
1580 "src/f32-vunary/gen/vneg-sse-x8.c",
1581 "src/f32-vunary/gen/vsqr-sse-x4.c",
1582 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001583 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001584 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001585 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001586 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001587 "src/math/roundd-sse-addsub.c",
1588 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001589 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001590 "src/math/sqrt-sse-hh1mac.c",
1591 "src/math/sqrt-sse-nr1mac.c",
1592 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001593]
1594
1595SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001596 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1597 "src/f32-argmaxpool/4x-sse2-c4.c",
1598 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001599 "src/f32-prelu/gen/sse2-2x4.c",
1600 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001601 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1602 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1603 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1604 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1605 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1606 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1607 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1608 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1609 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1610 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1611 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1612 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001613 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001614 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001615 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001616 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001617 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1618 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001619 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1620 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001621 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1622 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1623 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1624 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1625 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1626 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1627 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1628 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001629 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1630 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1631 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1632 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1633 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1634 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1635 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1636 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001637 "src/qu8-requantization/precise-sse2.c",
1638 "src/qu8-requantization/fp32-sse2.c",
1639 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001640 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001641 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001642 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001643 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001644 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001645 "src/x32-zip/x2-sse2.c",
1646 "src/x32-zip/x3-sse2.c",
1647 "src/x32-zip/x4-sse2.c",
1648 "src/x32-zip/xm-sse2.c",
1649 "src/x8-zip/x2-sse2.c",
1650 "src/x8-zip/x3-sse2.c",
1651 "src/x8-zip/x4-sse2.c",
1652 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001653 "src/math/exp-sse2-p5.c",
1654 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001655 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001656 "src/math/roundd-sse2-cvt.c",
1657 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001658 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001659 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001660]
1661
1662SSSE3_UKERNELS = [
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001663 "src/qu8-requantization/precise-ssse3.c",
1664 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001665]
1666
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001667SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001668 "src/f32-prelu/gen/sse41-2x4.c",
1669 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001670 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001671 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001672 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001673 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001674 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1675 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001676 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1677 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001678 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1679 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1680 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1681 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1682 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1683 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1684 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1685 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001686 "src/qu8-requantization/precise-sse4.c",
1687 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001688 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001689 "src/math/roundd-sse41.c",
1690 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001691 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001692]
1693
Marat Dukhan08c4a432019-10-03 09:29:21 -07001694AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001695 "src/f32-clamp/gen/avx-x8.c",
1696 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001697 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1698 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1699 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1700 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1701 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1702 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1703 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1704 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1705 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1706 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1707 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1708 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1709 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1710 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1711 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1712 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1713 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1714 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1715 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1716 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1717 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1718 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1719 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1720 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1721 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1722 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1723 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1724 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1725 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1726 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001727 "src/f32-hswish/gen/hswish-avx-x8.c",
1728 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001729 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1730 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1731 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1732 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1733 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1734 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1735 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1736 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1737 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001738 "src/f32-prelu/gen/avx-2x8.c",
1739 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001740 "src/f32-relu/gen/avx-x16.c",
1741 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001742 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001743 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1744 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1745 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1746 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1747 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1748 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1749 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1750 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001751 "src/f32-vbinary/gen/vmax-avx-x8.c",
1752 "src/f32-vbinary/gen/vmax-avx-x16.c",
1753 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1754 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1755 "src/f32-vbinary/gen/vmin-avx-x8.c",
1756 "src/f32-vbinary/gen/vmin-avx-x16.c",
1757 "src/f32-vbinary/gen/vminc-avx-x8.c",
1758 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001759 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1760 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1761 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1762 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1763 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1764 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1765 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1766 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001767 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1768 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1769 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1770 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001771 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1772 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1773 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1774 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001775 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1776 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001777 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001778 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1779 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1780 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1781 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1782 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1783 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1784 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1785 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001786 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1787 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001788 "src/f32-vunary/gen/vabs-avx-x8.c",
1789 "src/f32-vunary/gen/vabs-avx-x16.c",
1790 "src/f32-vunary/gen/vneg-avx-x8.c",
1791 "src/f32-vunary/gen/vneg-avx-x16.c",
1792 "src/f32-vunary/gen/vsqr-avx-x8.c",
1793 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001794]
1795
Marat Dukhanfda12b82019-11-21 12:27:59 -08001796FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001797 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1798 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1799 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1800 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1801 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1802 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1803 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1804 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1805 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1806 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1807 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1808 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1809 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
1810 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
1811 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
1812 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
1813 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
1814 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
1815 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
1816 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
1817 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
1818 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
1819 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
1820 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
1821 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
1822 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
1823 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
1824 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
1825 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
1826 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
1827 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
1828 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
1829 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
1830 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
1831 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
1832 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
1833 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
1834 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
1835 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
1836 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001837 "src/f32-hswish/gen/hswish-fma3-x8.c",
1838 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001839 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
1840 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
1841 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
1842 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
1843 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
1844 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
1845 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
1846 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
1847 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
1848 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
1849 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
1850 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
1851 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
1852 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001853 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
1854 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
1855 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
1856 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
1857 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
1858 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
1859 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
1860 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001861 "src/math/sqrt-fma3-nr1fma.c",
1862 "src/math/sqrt-fma3-nr2fma.c",
1863 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08001864]
1865
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001866AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001867 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
1868 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
1869 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
1870 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
1871 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
1872 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
1873 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
1874 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
1875 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
1876 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
1877 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
1878 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
1879 "src/f32-raddextexp/gen/avx2-p5-x64.c",
1880 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
1881 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
1882 "src/f32-raddextexp/gen/avx2-p5-x72.c",
1883 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
1884 "src/f32-raddextexp/gen/avx2-p5-x80.c",
1885 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
1886 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
1887 "src/f32-raddextexp/gen/avx2-p5-x96.c",
1888 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
1889 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
1890 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
1891 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
1892 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
1893 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
1894 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
1895 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
1896 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
1897 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
1898 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
1899 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
1900 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
1901 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
1902 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08001903 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
1904 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
1905 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
1906 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
1907 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
1908 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
1909 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
1910 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
1911 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
1912 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
1913 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
1914 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
1915 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
1916 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
1917 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
1918 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
1919 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
1920 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
1921 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
1922 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
1923 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
1924 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
1925 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
1926 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
1927 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
1928 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
1929 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
1930 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
1931 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
1932 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08001933 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
1934 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
1935 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
1936 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
1937 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
1938 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
1939 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
1940 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
1941 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
1942 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
1943 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
1944 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
1945 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
1946 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
1947 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
1948 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
1949 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
1950 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
1951 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
1952 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
1953 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
1954 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
1955 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
1956 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001957 "src/math/exp-avx2-p5.c",
1958 "src/math/exp-avx2-perm-p3.c",
1959 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07001960 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07001961 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08001962 "src/math/sigmoid-avx2-rr2-p5-div.c",
1963 "src/math/sigmoid-avx2-rr1-p5-div.c",
1964 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
1965 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
1966 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
1967 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07001968]
1969
Marat Dukhan08c4a432019-10-03 09:29:21 -07001970AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001971 "src/f32-clamp/gen/avx512f-x16.c",
1972 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001973 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
1974 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
1975 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
1976 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
1977 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
1978 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
1979 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
1980 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
1981 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
1982 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
1983 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
1984 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
1985 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
1986 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
1987 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
1988 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
1989 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
1990 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
1991 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
1992 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
1993 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
1994 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
1995 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
1996 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001997 "src/f32-hswish/gen/hswish-avx512f-x16.c",
1998 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001999 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2000 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2001 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2002 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2003 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2004 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002005 "src/f32-prelu/gen/avx512f-2x16.c",
2006 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002007 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2008 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2009 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2010 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2011 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2012 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2013 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2014 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2015 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2016 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2017 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2018 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2019 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2020 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2021 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2022 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2023 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2024 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2025 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2026 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2027 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2028 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2029 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2030 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2031 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2032 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2033 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2034 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2035 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2036 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2037 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2038 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2039 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2040 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2041 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2042 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002043 "src/f32-relu/gen/avx512f-x16.c",
2044 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002045 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002046 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2047 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2048 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2049 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2050 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2051 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2052 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2053 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002054 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2055 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2056 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2057 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2058 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2059 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2060 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2061 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002062 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2063 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2064 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2065 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2066 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2067 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2068 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2069 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002070 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2071 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2072 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2073 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002074 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2075 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2076 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2077 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002078 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2079 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002080 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002081 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2082 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2083 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2084 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2085 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2086 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2087 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2088 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2089 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2090 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2091 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2092 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2093 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2094 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2095 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2096 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2097 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2098 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2099 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2100 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2101 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2102 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2103 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2104 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002105 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2106 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2107 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2108 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2109 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2110 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2111 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2112 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002113 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2114 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2115 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2116 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2117 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2118 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2119 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2120 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002121 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2122 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2123 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2124 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2125 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2126 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002127 "src/math/exp-avx512f-p5-scalef.c",
2128 "src/math/exp-avx512f-p5.c",
2129 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002130 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002131 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002132 "src/math/sqrt-avx512f-nr1fma.c",
2133 "src/math/sqrt-avx512f-nr2fma.c",
2134 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002135]
2136
2137AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002138 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002139 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002140 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2141 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002142 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002143 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002144 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002145 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002146 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2147 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002148 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2149 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2150 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2151 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002152]
2153
2154AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002155 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002156 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2157 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002158 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002159 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2160 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002161 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2162 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2163 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002164 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002165 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2166 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2167 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002168 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002169 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2170 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002171 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002172 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2173 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2174 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2175 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2176 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2177 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2178 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2179 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2180 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2181 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2182 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2183 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2184 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2185 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2186 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2187 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2188 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2189 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002190 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2191 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002192 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002193 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2194 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2195 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2196 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2197 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2198 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2199 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2200 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2201 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2202 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2203 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2204 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2205 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2206 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2207 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2208 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2209 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2210 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002211 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2212 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2213 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2214 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2215 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2216 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2217 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2218 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2219 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2220 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2221 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2222 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2223 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2224 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2225 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2226 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2227 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2228 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002229]
2230
Marat Dukhan1b354632020-03-23 12:50:22 -07002231INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002232 "src/xnnpack/argmaxpool.h",
2233 "src/xnnpack/avgpool.h",
2234 "src/xnnpack/clamp.h",
2235 "src/xnnpack/common.h",
2236 "src/xnnpack/conv.h",
2237 "src/xnnpack/dwconv.h",
2238 "src/xnnpack/gavgpool.h",
2239 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002240 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002241 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002242 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002243 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002244 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002245 "src/xnnpack/lut.h",
2246 "src/xnnpack/math.h",
2247 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002248 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002249 "src/xnnpack/packx.h",
2250 "src/xnnpack/pad.h",
2251 "src/xnnpack/params.h",
2252 "src/xnnpack/pavgpool.h",
2253 "src/xnnpack/ppmm.h",
2254 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002255 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002256 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002257 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002258 "src/xnnpack/rmax.h",
2259 "src/xnnpack/scalar-utils.h",
2260 "src/xnnpack/spmm.h",
2261 "src/xnnpack/unpool.h",
2262 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002263 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002264 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002265 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002266 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002267 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002268 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002269 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002270]
2271
2272INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002273 "include/xnnpack.h",
2274 "src/xnnpack/allocator.h",
2275 "src/xnnpack/compute.h",
2276 "src/xnnpack/im2col.h",
2277 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002278 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002279 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002280 "src/xnnpack/operator.h",
2281 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002282 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002283 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002284 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002285 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002286]
2287
Marat Dukhan1b354632020-03-23 12:50:22 -07002288ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002289 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002290]
2291
Marat Dukhan1b354632020-03-23 12:50:22 -07002292MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002293 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002294 "include/xnnpack.h",
2295]
2296
Marat Dukhan1b354632020-03-23 12:50:22 -07002297MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002298 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002299 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002300 "src/xnnpack/requantization.h",
2301 "include/xnnpack.h",
2302]
2303
2304OPERATOR_TEST_PARAMS_HDRS = [
2305 "src/xnnpack/params.h",
2306 "src/xnnpack/common.h",
2307]
2308
2309WEIGHTS_PACK_HDRS = [
2310 "src/xnnpack/pack.h",
2311 "src/xnnpack/operator.h",
2312 "src/xnnpack/compute.h",
2313]
2314
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002315LOGGING_COPTS = select({
2316 # No logging in optimized mode
2317 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2318 # Full logging in debug mode
2319 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2320 # Error-only logging in default (fastbuild) mode
2321 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2322})
2323
Marat Dukhan3b59de22020-06-03 20:15:19 -07002324LOGGING_SRCS = select({
2325 # No logging in optimized mode
2326 ":optimized_build": [],
2327 "//conditions:default": [
2328 "src/operator-strings.c",
2329 "src/subgraph-strings.c",
2330 ],
2331})
2332
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002333LOGGING_HDRS = [
2334 "src/xnnpack/log.h",
2335]
2336
Marat Dukhan08c4a432019-10-03 09:29:21 -07002337xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002338 name = "tables",
2339 srcs = TABLE_SRCS,
2340 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002341 gcc_copts = xnnpack_gcc_std_copts(),
2342 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002343)
2344
2345xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002346 name = "scalar_ukernels",
2347 srcs = SCALAR_UKERNELS,
2348 hdrs = INTERNAL_HDRS,
2349 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002350 gcc_copts = xnnpack_gcc_std_copts(),
2351 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002352 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002353 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002354 "@FP16",
2355 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002356 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002357 ],
2358)
2359
2360xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002361 name = "scalar_ukernels_test_mode",
2362 srcs = SCALAR_UKERNELS,
2363 hdrs = INTERNAL_HDRS,
2364 aarch32_copts = ["-marm"],
2365 copts = [
2366 "-UNDEBUG",
2367 "-DXNN_TEST_MODE=1",
2368 ],
2369 gcc_copts = xnnpack_gcc_std_copts(),
2370 msvc_copts = xnnpack_msvc_std_copts(),
2371 deps = [
2372 ":tables",
2373 "@FP16",
2374 "@FXdiv",
2375 "@pthreadpool",
2376 ],
2377)
2378
2379xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002380 name = "wasm_ukernels",
2381 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002382 gcc_copts = xnnpack_gcc_std_copts(),
2383 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002384 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002385 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002386 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002387 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002388 "@FP16",
2389 "@FXdiv",
2390 "@pthreadpool",
2391 ],
2392)
2393
2394xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002395 name = "wasm_ukernels_test_mode",
2396 hdrs = INTERNAL_HDRS,
2397 copts = [
2398 "-UNDEBUG",
2399 "-DXNN_TEST_MODE=1",
2400 ],
2401 gcc_copts = xnnpack_gcc_std_copts(),
2402 msvc_copts = xnnpack_msvc_std_copts(),
2403 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002404 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002405 deps = [
2406 ":tables",
2407 "@FP16",
2408 "@FXdiv",
2409 "@pthreadpool",
2410 ],
2411)
2412
2413xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002414 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002415 hdrs = INTERNAL_HDRS,
2416 aarch32_copts = [
2417 "-marm",
2418 "-mfpu=neon",
2419 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002420 gcc_copts = xnnpack_gcc_std_copts(),
2421 gcc_x86_copts = ["-msse2"],
2422 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002423 optimized_copts = [
2424 "-O3",
2425 "-ffast-math",
2426 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002427 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002428 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002429 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002430 "@FP16",
2431 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002432 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002433 ],
2434)
2435
2436xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002437 name = "psimd_fastmath_ukernels_test_mode",
2438 hdrs = INTERNAL_HDRS,
2439 aarch32_copts = [
2440 "-marm",
2441 "-mfpu=neon",
2442 ],
2443 copts = [
2444 "-UNDEBUG",
2445 "-DXNN_TEST_MODE=1",
2446 ],
2447 gcc_copts = xnnpack_gcc_std_copts(),
2448 gcc_x86_copts = ["-msse2"],
2449 msvc_copts = xnnpack_msvc_std_copts(),
2450 optimized_copts = [
2451 "-O3",
2452 "-ffast-math",
2453 ],
2454 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2455 deps = [
2456 ":tables",
2457 "@FP16",
2458 "@psimd",
2459 "@pthreadpool",
2460 ],
2461)
2462
2463xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002464 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002465 hdrs = INTERNAL_HDRS,
2466 aarch32_copts = [
2467 "-marm",
2468 "-mfpu=neon",
2469 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002470 gcc_copts = xnnpack_gcc_std_copts(),
2471 gcc_x86_copts = ["-msse2"],
2472 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002473 optimized_copts = [
2474 "-O3",
2475 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002476 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002477 deps = [
2478 ":tables",
2479 "@FP16",
2480 "@psimd",
2481 "@pthreadpool",
2482 ],
2483)
2484
2485xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002486 name = "psimd_accmath_ukernels_test_mode",
2487 hdrs = INTERNAL_HDRS,
2488 aarch32_copts = [
2489 "-marm",
2490 "-mfpu=neon",
2491 ],
2492 copts = [
2493 "-UNDEBUG",
2494 "-DXNN_TEST_MODE=1",
2495 ],
2496 gcc_copts = xnnpack_gcc_std_copts(),
2497 gcc_x86_copts = ["-msse2"],
2498 msvc_copts = xnnpack_msvc_std_copts(),
2499 optimized_copts = [
2500 "-O3",
2501 ],
2502 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2503 deps = [
2504 ":tables",
2505 "@FP16",
2506 "@psimd",
2507 "@pthreadpool",
2508 ],
2509)
2510
2511xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002512 name = "neon_ukernels",
2513 hdrs = INTERNAL_HDRS,
2514 aarch32_copts = [
2515 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002516 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002517 "-mfpu=neon",
2518 ],
2519 aarch32_srcs = NEON_UKERNELS,
2520 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002521 gcc_copts = xnnpack_gcc_std_copts(),
2522 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002523 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002524 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002525 "@FP16",
2526 "@pthreadpool",
2527 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002528)
2529
2530xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002531 name = "neon_ukernels_test_mode",
2532 hdrs = INTERNAL_HDRS,
2533 aarch32_copts = [
2534 "-marm",
2535 "-march=armv7-a",
2536 "-mfpu=neon",
2537 ],
2538 aarch32_srcs = NEON_UKERNELS,
2539 aarch64_srcs = NEON_UKERNELS,
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 "@pthreadpool",
2550 ],
2551)
2552
2553xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002554 name = "neonfma_ukernels",
2555 hdrs = INTERNAL_HDRS,
2556 aarch32_copts = [
2557 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002558 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002559 "-mfpu=neon-vfpv4",
2560 ],
2561 aarch32_srcs = NEONFMA_UKERNELS,
2562 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002563 apple_aarch32_copts = [
2564 "-mcpu=swift",
2565 "-mtune=generic",
2566 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002567 gcc_copts = xnnpack_gcc_std_copts(),
2568 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002569 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002570 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002571 "@FP16",
2572 "@pthreadpool",
2573 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002574)
2575
2576xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002577 name = "neonfma_ukernels_test_mode",
2578 hdrs = INTERNAL_HDRS,
2579 aarch32_copts = [
2580 "-marm",
2581 "-march=armv7-a",
2582 "-mfpu=neon-vfpv4",
2583 ],
2584 aarch32_srcs = NEONFMA_UKERNELS,
2585 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002586 apple_aarch32_copts = [
2587 "-mcpu=swift",
2588 "-mtune=generic",
2589 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002590 copts = [
2591 "-UNDEBUG",
2592 "-DXNN_TEST_MODE=1",
2593 ],
2594 gcc_copts = xnnpack_gcc_std_copts(),
2595 msvc_copts = xnnpack_msvc_std_copts(),
2596 deps = [
2597 ":tables",
2598 "@FP16",
2599 "@pthreadpool",
2600 ],
2601)
2602
2603xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002604 name = "neonv8_ukernels",
2605 hdrs = INTERNAL_HDRS,
2606 aarch32_copts = [
2607 "-marm",
2608 "-march=armv8-a",
2609 "-mfpu=neon-fp-armv8",
2610 ],
2611 aarch32_srcs = NEONV8_UKERNELS,
2612 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002613 apple_aarch32_copts = [
2614 "-mcpu=cyclone",
2615 "-mtune=generic",
2616 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002617 gcc_copts = xnnpack_gcc_std_copts(),
2618 msvc_copts = xnnpack_msvc_std_copts(),
2619 deps = [
2620 ":tables",
2621 "@FP16",
2622 "@pthreadpool",
2623 ],
2624)
2625
2626xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002627 name = "neonv8_ukernels_test_mode",
2628 hdrs = INTERNAL_HDRS,
2629 aarch32_copts = [
2630 "-marm",
2631 "-march=armv8-a",
2632 "-mfpu=neon-fp-armv8",
2633 ],
2634 aarch32_srcs = NEONV8_UKERNELS,
2635 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002636 apple_aarch32_copts = [
2637 "-mcpu=cyclone",
2638 "-mtune=generic",
2639 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002640 copts = [
2641 "-UNDEBUG",
2642 "-DXNN_TEST_MODE=1",
2643 ],
2644 gcc_copts = xnnpack_gcc_std_copts(),
2645 msvc_copts = xnnpack_msvc_std_copts(),
2646 deps = [
2647 ":tables",
2648 "@FP16",
2649 "@pthreadpool",
2650 ],
2651)
2652
2653xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002654 name = "neonfp16arith_ukernels",
2655 hdrs = INTERNAL_HDRS,
2656 aarch64_copts = ["-march=armv8.2-a+fp16"],
2657 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002658 gcc_copts = xnnpack_gcc_std_copts(),
2659 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002660 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002661 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002662 "@FP16",
2663 "@pthreadpool",
2664 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002665)
2666
2667xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002668 name = "neonfp16arith_ukernels_test_mode",
2669 hdrs = INTERNAL_HDRS,
2670 aarch64_copts = ["-march=armv8.2-a+fp16"],
2671 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2672 copts = [
2673 "-UNDEBUG",
2674 "-DXNN_TEST_MODE=1",
2675 ],
2676 gcc_copts = xnnpack_gcc_std_copts(),
2677 msvc_copts = xnnpack_msvc_std_copts(),
2678 deps = [
2679 ":tables",
2680 "@FP16",
2681 "@pthreadpool",
2682 ],
2683)
2684
2685xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002686 name = "sse2_ukernels",
2687 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002688 gcc_copts = xnnpack_gcc_std_copts(),
2689 gcc_x86_copts = ["-msse2"],
2690 msvc_copts = xnnpack_msvc_std_copts(),
2691 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002692 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002693 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002694 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002695 "@FP16",
2696 "@pthreadpool",
2697 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002698)
2699
2700xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002701 name = "sse2_ukernels_test_mode",
2702 hdrs = INTERNAL_HDRS,
2703 copts = [
2704 "-UNDEBUG",
2705 "-DXNN_TEST_MODE=1",
2706 ],
2707 gcc_copts = xnnpack_gcc_std_copts(),
2708 gcc_x86_copts = ["-msse2"],
2709 msvc_copts = xnnpack_msvc_std_copts(),
2710 msvc_x86_32_copts = ["/arch:SSE2"],
2711 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2712 deps = [
2713 ":tables",
2714 "@FP16",
2715 "@pthreadpool",
2716 ],
2717)
2718
2719xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002720 name = "ssse3_ukernels",
2721 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002722 gcc_copts = xnnpack_gcc_std_copts(),
2723 gcc_x86_copts = ["-mssse3"],
2724 msvc_copts = xnnpack_msvc_std_copts(),
2725 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002726 x86_srcs = SSSE3_UKERNELS,
2727 deps = [
2728 ":tables",
2729 "@FP16",
2730 "@pthreadpool",
2731 ],
2732)
2733
2734xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002735 name = "ssse3_ukernels_test_mode",
2736 hdrs = INTERNAL_HDRS,
2737 copts = [
2738 "-UNDEBUG",
2739 "-DXNN_TEST_MODE=1",
2740 ],
2741 gcc_copts = xnnpack_gcc_std_copts(),
2742 gcc_x86_copts = ["-mssse3"],
2743 msvc_copts = xnnpack_msvc_std_copts(),
2744 msvc_x86_32_copts = ["/arch:SSE2"],
2745 x86_srcs = SSSE3_UKERNELS,
2746 deps = [
2747 ":tables",
2748 "@FP16",
2749 "@pthreadpool",
2750 ],
2751)
2752
2753xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002754 name = "sse41_ukernels",
2755 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002756 gcc_copts = xnnpack_gcc_std_copts(),
2757 gcc_x86_copts = ["-msse4.1"],
2758 msvc_copts = xnnpack_msvc_std_copts(),
2759 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002760 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002761 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002762 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002763 "@FP16",
2764 "@pthreadpool",
2765 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002766)
2767
2768xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002769 name = "sse41_ukernels_test_mode",
2770 hdrs = INTERNAL_HDRS,
2771 copts = [
2772 "-UNDEBUG",
2773 "-DXNN_TEST_MODE=1",
2774 ],
2775 gcc_copts = xnnpack_gcc_std_copts(),
2776 gcc_x86_copts = ["-msse4.1"],
2777 msvc_copts = xnnpack_msvc_std_copts(),
2778 msvc_x86_32_copts = ["/arch:SSE2"],
2779 x86_srcs = SSE41_UKERNELS,
2780 deps = [
2781 ":tables",
2782 "@FP16",
2783 "@pthreadpool",
2784 ],
2785)
2786
2787xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002788 name = "avx_ukernels",
2789 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002790 gcc_copts = xnnpack_gcc_std_copts(),
2791 gcc_x86_copts = ["-mavx"],
2792 msvc_copts = xnnpack_msvc_std_copts(),
2793 msvc_x86_32_copts = ["/arch:AVX"],
2794 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002795 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002796 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002797 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002798 "@FP16",
2799 "@pthreadpool",
2800 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002801)
2802
2803xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002804 name = "avx_ukernels_test_mode",
2805 hdrs = INTERNAL_HDRS,
2806 copts = [
2807 "-UNDEBUG",
2808 "-DXNN_TEST_MODE=1",
2809 ],
2810 gcc_copts = xnnpack_gcc_std_copts(),
2811 gcc_x86_copts = ["-mavx"],
2812 msvc_copts = xnnpack_msvc_std_copts(),
2813 msvc_x86_32_copts = ["/arch:AVX"],
2814 msvc_x86_64_copts = ["/arch:AVX"],
2815 x86_srcs = AVX_UKERNELS,
2816 deps = [
2817 ":tables",
2818 "@FP16",
2819 "@pthreadpool",
2820 ],
2821)
2822
2823xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08002824 name = "fma3_ukernels",
2825 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002826 gcc_copts = xnnpack_gcc_std_copts(),
2827 gcc_x86_copts = ["-mfma"],
2828 msvc_copts = xnnpack_msvc_std_copts(),
2829 msvc_x86_32_copts = ["/arch:AVX"],
2830 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08002831 x86_srcs = FMA3_UKERNELS,
2832 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002833 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002834 "@FP16",
2835 "@pthreadpool",
2836 ],
2837)
2838
2839xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002840 name = "fma3_ukernels_test_mode",
2841 hdrs = INTERNAL_HDRS,
2842 copts = [
2843 "-UNDEBUG",
2844 "-DXNN_TEST_MODE=1",
2845 ],
2846 gcc_copts = xnnpack_gcc_std_copts(),
2847 gcc_x86_copts = ["-mfma"],
2848 msvc_copts = xnnpack_msvc_std_copts(),
2849 msvc_x86_32_copts = ["/arch:AVX"],
2850 msvc_x86_64_copts = ["/arch:AVX"],
2851 x86_srcs = FMA3_UKERNELS,
2852 deps = [
2853 ":tables",
2854 "@FP16",
2855 "@pthreadpool",
2856 ],
2857)
2858
2859xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002860 name = "avx2_ukernels",
2861 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002862 gcc_copts = xnnpack_gcc_std_copts(),
2863 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002864 "-mfma",
2865 "-mavx2",
2866 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002867 msvc_copts = xnnpack_msvc_std_copts(),
2868 msvc_x86_32_copts = ["/arch:AVX2"],
2869 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002870 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002871 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002872 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002873 "@FP16",
2874 "@pthreadpool",
2875 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002876)
2877
2878xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002879 name = "avx2_ukernels_test_mode",
2880 hdrs = INTERNAL_HDRS,
2881 copts = [
2882 "-UNDEBUG",
2883 "-DXNN_TEST_MODE=1",
2884 ],
2885 gcc_copts = xnnpack_gcc_std_copts(),
2886 gcc_x86_copts = [
2887 "-mfma",
2888 "-mavx2",
2889 ],
2890 msvc_copts = xnnpack_msvc_std_copts(),
2891 msvc_x86_32_copts = ["/arch:AVX2"],
2892 msvc_x86_64_copts = ["/arch:AVX2"],
2893 x86_srcs = AVX2_UKERNELS,
2894 deps = [
2895 ":tables",
2896 "@FP16",
2897 "@pthreadpool",
2898 ],
2899)
2900
2901xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002902 name = "avx512f_ukernels",
2903 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002904 gcc_copts = xnnpack_gcc_std_copts(),
2905 gcc_x86_copts = ["-mavx512f"],
2906 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2907 msvc_copts = xnnpack_msvc_std_copts(),
2908 msvc_x86_32_copts = ["/arch:AVX512"],
2909 msvc_x86_64_copts = ["/arch:AVX512"],
2910 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002911 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002912 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002913 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002914 "@FP16",
2915 "@pthreadpool",
2916 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002917)
2918
2919xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002920 name = "avx512f_ukernels_test_mode",
2921 hdrs = INTERNAL_HDRS,
2922 copts = [
2923 "-UNDEBUG",
2924 "-DXNN_TEST_MODE=1",
2925 ],
2926 gcc_copts = xnnpack_gcc_std_copts(),
2927 gcc_x86_copts = ["-mavx512f"],
2928 mingw_copts = ["-fno-asynchronous-unwind-tables"],
2929 msvc_copts = xnnpack_msvc_std_copts(),
2930 msvc_x86_32_copts = ["/arch:AVX512"],
2931 msvc_x86_64_copts = ["/arch:AVX512"],
2932 msys_copts = ["-fno-asynchronous-unwind-tables"],
2933 x86_srcs = AVX512F_UKERNELS,
2934 deps = [
2935 ":tables",
2936 "@FP16",
2937 "@pthreadpool",
2938 ],
2939)
2940
2941xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002942 name = "asm_ukernels",
2943 hdrs = ["src/xnnpack/assembly.h"],
2944 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07002945 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002946 aarch64_srcs = AARCH64_ASM_UKERNELS,
2947)
2948
Marat Dukhan3b59de22020-06-03 20:15:19 -07002949xnnpack_cc_library(
2950 name = "logging_utils",
2951 srcs = LOGGING_SRCS,
2952 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
2953 copts = LOGGING_COPTS + [
2954 "-Isrc",
2955 "-Iinclude",
2956 ] + select({
2957 ":debug_build": [],
2958 "//conditions:default": xnnpack_min_size_copts(),
2959 }),
2960 gcc_copts = xnnpack_gcc_std_copts(),
2961 msvc_copts = xnnpack_msvc_std_copts(),
2962 visibility = xnnpack_visibility(),
2963 deps = [
2964 "@FP16",
2965 "@clog",
2966 "@pthreadpool",
2967 ],
2968)
2969
Marat Dukhan08c4a432019-10-03 09:29:21 -07002970xnnpack_aggregate_library(
2971 name = "ukernels",
2972 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002973 ":neon_ukernels",
2974 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002975 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002976 ":asm_ukernels",
2977 ],
2978 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002979 ":neon_ukernels",
2980 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07002981 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002982 ":neonfp16arith_ukernels",
2983 ":asm_ukernels",
2984 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002985 generic_deps = [
2986 ":scalar_ukernels",
2987 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002988 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002989 ":psimd_fastmath_ukernels",
2990 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002991 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002992 wasm_deps = [
2993 ":wasm_ukernels",
2994 ],
2995 wasmsimd_deps = [
2996 ":wasm_ukernels",
2997 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002998 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002999 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003000 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003001 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003002 ":avx_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003003 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003004 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003005 ":avx512f_ukernels",
3006 ],
3007)
3008
Marat Dukhan33fcf782020-05-24 14:27:15 -07003009xnnpack_aggregate_library(
3010 name = "ukernels_test_mode",
3011 aarch32_deps = [
3012 ":neon_ukernels_test_mode",
3013 ":neonfma_ukernels_test_mode",
3014 ":neonv8_ukernels_test_mode",
3015 ":asm_ukernels",
3016 ],
3017 aarch64_deps = [
3018 ":neon_ukernels_test_mode",
3019 ":neonfma_ukernels_test_mode",
3020 ":neonv8_ukernels_test_mode",
3021 ":neonfp16arith_ukernels_test_mode",
3022 ":asm_ukernels",
3023 ],
3024 generic_deps = [
3025 ":scalar_ukernels_test_mode",
3026 ],
3027 psimd_deps = [
3028 ":psimd_fastmath_ukernels_test_mode",
3029 ":psimd_accmath_ukernels_test_mode",
3030 ],
3031 wasm_deps = [
3032 ":wasm_ukernels_test_mode",
3033 ],
3034 wasmsimd_deps = [
3035 ":wasm_ukernels_test_mode",
3036 ],
3037 x86_deps = [
3038 ":sse2_ukernels_test_mode",
3039 ":ssse3_ukernels_test_mode",
3040 ":sse41_ukernels_test_mode",
3041 ":avx_ukernels_test_mode",
3042 ":fma3_ukernels_test_mode",
3043 ":avx2_ukernels_test_mode",
3044 ":avx512f_ukernels_test_mode",
3045 ],
3046)
3047
Marat Dukhan86df3842020-06-17 14:03:31 -07003048# This target must be compatible with Android NDK r17 and compile with Clang 6.
3049# It is similar to :ukernels target, but excludes microkernels which require
3050# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003051xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003052 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003053 aarch32_deps = [
3054 ":neon_ukernels",
3055 ":neonfma_ukernels",
3056 ":neonv8_ukernels",
3057 ":asm_ukernels",
3058 ],
3059 aarch64_deps = [
3060 ":neon_ukernels",
3061 ":neonfma_ukernels",
3062 ":neonv8_ukernels",
3063 ":asm_ukernels",
3064 ],
3065 generic_deps = [
3066 ":scalar_ukernels",
3067 ],
3068 psimd_deps = [
3069 ":psimd_fastmath_ukernels",
3070 ":psimd_accmath_ukernels",
3071 ],
3072 wasm_deps = [
3073 ":wasm_ukernels",
3074 ],
3075 wasmsimd_deps = [
3076 ":wasm_ukernels",
3077 ],
3078 x86_deps = [
3079 ":sse2_ukernels",
3080 ":ssse3_ukernels",
3081 ":sse41_ukernels",
3082 ":avx_ukernels",
3083 ":fma3_ukernels",
3084 ":avx2_ukernels",
3085 ":avx512f_ukernels",
3086 ],
3087)
3088
Marat Dukhan08c4a432019-10-03 09:29:21 -07003089xnnpack_cc_library(
3090 name = "im2col",
3091 srcs = ["src/im2col.c"],
3092 hdrs = [
3093 "src/xnnpack/common.h",
3094 "src/xnnpack/im2col.h",
3095 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003096 gcc_copts = xnnpack_gcc_std_copts(),
3097 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003098)
3099
3100xnnpack_cc_library(
3101 name = "indirection",
3102 srcs = ["src/indirection.c"],
3103 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003104 gcc_copts = xnnpack_gcc_std_copts(),
3105 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003106 deps = [
3107 "@FP16",
3108 "@FXdiv",
3109 "@pthreadpool",
3110 ],
3111)
3112
3113xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003114 name = "indirection_test_mode",
3115 srcs = ["src/indirection.c"],
3116 hdrs = INTERNAL_HDRS,
3117 copts = [
3118 "-UNDEBUG",
3119 "-DXNN_TEST_MODE=1",
3120 ],
3121 gcc_copts = xnnpack_gcc_std_copts(),
3122 msvc_copts = xnnpack_msvc_std_copts(),
3123 deps = [
3124 "@FP16",
3125 "@FXdiv",
3126 "@pthreadpool",
3127 ],
3128)
3129
3130xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003131 name = "packing",
3132 srcs = ["src/packing.c"],
3133 hdrs = INTERNAL_HDRS,
3134 gcc_copts = xnnpack_gcc_std_copts(),
3135 msvc_copts = xnnpack_msvc_std_copts(),
3136 deps = [
3137 "@FP16",
3138 "@FXdiv",
3139 "@pthreadpool",
3140 ],
3141)
3142
3143xnnpack_cc_library(
3144 name = "packing_test_mode",
3145 srcs = ["src/packing.c"],
3146 hdrs = INTERNAL_HDRS,
3147 copts = [
3148 "-UNDEBUG",
3149 "-DXNN_TEST_MODE=1",
3150 ],
3151 gcc_copts = xnnpack_gcc_std_copts(),
3152 msvc_copts = xnnpack_msvc_std_copts(),
3153 deps = [
3154 "@FP16",
3155 "@FXdiv",
3156 "@pthreadpool",
3157 ],
3158)
3159
3160xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003161 name = "operator_run",
3162 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003163 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003164 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003165 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3166 "//conditions:default": [],
3167 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003168 gcc_copts = xnnpack_gcc_std_copts(),
3169 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003170 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003171 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003172 "@FP16",
3173 "@FXdiv",
3174 "@clog",
3175 "@pthreadpool",
3176 ],
3177)
3178
Chao Mei6ddfc602020-05-13 22:29:36 -07003179xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003180 name = "operator_run_test_mode",
3181 srcs = ["src/operator-run.c"],
3182 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3183 copts = LOGGING_COPTS + [
3184 "-UNDEBUG",
3185 "-DXNN_TEST_MODE=1",
3186 ] + select({
3187 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3188 "//conditions:default": [],
3189 }),
3190 gcc_copts = xnnpack_gcc_std_copts(),
3191 msvc_copts = xnnpack_msvc_std_copts(),
3192 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003193 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003194 "@FP16",
3195 "@FXdiv",
3196 "@clog",
3197 "@pthreadpool",
3198 ],
3199)
3200
3201xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003202 name = "memory_planner",
3203 srcs = ["src/memory-planner.c"],
3204 hdrs = INTERNAL_HDRS,
3205 defines = select({
3206 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3207 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3208 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3209 }),
3210 gcc_copts = xnnpack_gcc_std_copts(),
3211 msvc_copts = xnnpack_msvc_std_copts(),
3212 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003213 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003214 "@pthreadpool",
3215 ],
3216)
3217
Marat Dukhan33fcf782020-05-24 14:27:15 -07003218xnnpack_cc_library(
3219 name = "memory_planner_test_mode",
3220 srcs = ["src/memory-planner.c"],
3221 hdrs = INTERNAL_HDRS,
3222 copts = [
3223 "-UNDEBUG",
3224 "-DXNN_TEST_MODE=1",
3225 ],
3226 defines = select({
3227 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3228 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3229 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3230 }),
3231 gcc_copts = xnnpack_gcc_std_copts(),
3232 msvc_copts = xnnpack_msvc_std_copts(),
3233 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003234 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003235 "@pthreadpool",
3236 ],
3237)
3238
Marat Dukhan08c4a432019-10-03 09:29:21 -07003239cc_library(
3240 name = "enable_assembly",
3241 defines = select({
3242 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3243 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003244 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003245 }),
3246)
3247
Marat Dukhan9de90e02020-06-18 16:04:12 -07003248cc_library(
3249 name = "enable_sparse",
3250 defines = select({
3251 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3252 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003253 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003254 }),
3255)
3256
Marat Dukhancf056b22019-10-07 10:26:29 -07003257xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003258 name = "operators",
3259 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003260 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003261 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003262 ],
3263 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003264 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003265 "-Isrc",
3266 "-Iinclude",
3267 ] + select({
3268 ":debug_build": [],
3269 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003270 }) + select({
3271 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3272 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003273 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003274 gcc_copts = xnnpack_gcc_std_copts(),
3275 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003276 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003277 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003278 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003279 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003280 "@FP16",
3281 "@FXdiv",
3282 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003283 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003284 ],
3285)
3286
Marat Dukhan10a38082020-04-17 03:58:35 -07003287xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003288 name = "operators_test_mode",
3289 srcs = OPERATOR_SRCS + [
3290 "src/memory.c",
3291 "src/operator-delete.c",
3292 ],
3293 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3294 copts = LOGGING_COPTS + [
3295 "-Isrc",
3296 "-Iinclude",
3297 "-UNDEBUG",
3298 "-DXNN_TEST_MODE=1",
3299 ] + select({
3300 ":debug_build": [],
3301 "//conditions:default": xnnpack_min_size_copts(),
3302 }) + select({
3303 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3304 "//conditions:default": [],
3305 }),
3306 gcc_copts = xnnpack_gcc_std_copts(),
3307 msvc_copts = xnnpack_msvc_std_copts(),
3308 deps = [
3309 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003310 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003311 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003312 "@FP16",
3313 "@FXdiv",
3314 "@clog",
3315 "@pthreadpool",
3316 ],
3317)
3318
3319xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003320 name = "XNNPACK",
3321 srcs = [
3322 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003323 "src/runtime.c",
3324 "src/subgraph.c",
3325 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003326 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003327 hdrs = ["include/xnnpack.h"],
3328 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003329 "-Isrc",
3330 "-Iinclude",
3331 ] + select({
3332 ":debug_build": [],
3333 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003334 }) + select({
3335 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3336 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003337 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003338 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003339 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003340 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003341 visibility = xnnpack_visibility(),
3342 deps = [
3343 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003344 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003345 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003346 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003347 ":operator_run",
3348 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003349 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003350 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003351 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003352 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003353 ] + select({
3354 ":emscripten": [],
3355 "//conditions:default": ["@cpuinfo"],
3356 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003357)
3358
Marat Dukhan10a38082020-04-17 03:58:35 -07003359xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003360 name = "XNNPACK_test_mode",
3361 srcs = [
3362 "src/init.c",
3363 "src/runtime.c",
3364 "src/subgraph.c",
3365 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003366 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003367 hdrs = ["include/xnnpack.h"],
3368 copts = LOGGING_COPTS + [
3369 "-Isrc",
3370 "-Iinclude",
3371 "-UNDEBUG",
3372 "-DXNN_TEST_MODE=1",
3373 ] + select({
3374 ":debug_build": [],
3375 "//conditions:default": xnnpack_min_size_copts(),
3376 }) + select({
3377 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3378 "//conditions:default": [],
3379 }),
3380 gcc_copts = xnnpack_gcc_std_copts(),
3381 includes = ["include"],
3382 msvc_copts = xnnpack_msvc_std_copts(),
3383 visibility = xnnpack_visibility(),
3384 deps = [
3385 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003386 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003387 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003388 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003389 ":operator_run_test_mode",
3390 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003391 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003392 "@clog",
3393 "@FP16",
3394 "@pthreadpool",
3395 ] + select({
3396 ":emscripten": [],
3397 "//conditions:default": ["@cpuinfo"],
3398 }),
3399)
3400
3401xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003402 name = "xnnpack_f32",
3403 srcs = [
3404 "src/init.c",
3405 "src/runtime.c",
3406 "src/subgraph.c",
3407 "src/tensor.c",
3408 ] + SUBGRAPH_SRCS,
3409 hdrs = ["include/xnnpack.h"],
3410 copts = LOGGING_COPTS + [
3411 "-Isrc",
3412 "-Iinclude",
3413 ] + select({
3414 ":debug_build": [],
3415 "//conditions:default": xnnpack_min_size_copts(),
3416 }) + select({
3417 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3418 "//conditions:default": [],
3419 }),
3420 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003421 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003422 "XNN_NO_U8_OPERATORS",
3423 "XNN_NO_X8_OPERATORS",
3424 "XNN_NO_F16_OPERATORS",
3425 "XNN_NO_X16_OPERATORS",
3426 ],
3427 gcc_copts = xnnpack_gcc_std_copts(),
3428 includes = ["include"],
3429 msvc_copts = xnnpack_msvc_std_copts(),
3430 visibility = xnnpack_visibility(),
3431 deps = [
3432 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003433 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003434 ":logging_utils",
3435 ":memory_planner",
3436 ":operator_run",
3437 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003438 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003439 "@clog",
3440 "@FP16",
3441 "@pthreadpool",
3442 ] + select({
3443 ":emscripten": [],
3444 "//conditions:default": ["@cpuinfo"],
3445 }),
3446)
3447
3448xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003449 name = "xnnpack_operators_nhwc_f32",
3450 srcs = [
3451 "src/init.c",
3452 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003453 hdrs = ["include/xnnpack.h"],
3454 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003455 "-Isrc",
3456 "-Iinclude",
3457 ] + select({
3458 ":debug_build": [],
3459 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003460 }) + select({
3461 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3462 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003463 }),
3464 defines = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07003465 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003466 "XNN_NO_U8_OPERATORS",
3467 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003468 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003469 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003470 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003471 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003472 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003473 visibility = xnnpack_visibility(),
3474 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003475 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003476 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003477 ":operator_run",
3478 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003479 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003480 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003481 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003482 ] + select({
3483 ":emscripten": [],
3484 "//conditions:default": ["@cpuinfo"],
3485 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003486)
3487
Marat Dukhancf056b22019-10-07 10:26:29 -07003488xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003489 name = "bench_utils",
3490 srcs = ["bench/utils.cc"],
3491 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003492 deps = [
3493 "@com_google_benchmark//:benchmark",
3494 "@cpuinfo",
3495 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003496)
3497
Frank Barchard7e955972019-10-11 10:34:25 -07003498######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003499
3500xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003501 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003502 srcs = [
3503 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003504 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003505 "src/xnnpack/AlignedAllocator.h",
3506 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003507 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003508 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003509)
3510
3511xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07003512 name = "qu8_requantization_bench",
3513 srcs = [
3514 "bench/qu8-requantization.cc",
3515 "src/xnnpack/requantization-stubs.h",
3516 "src/xnnpack/AlignedAllocator.h",
3517 ] + MICROKERNEL_BENCHMARK_HDRS,
3518 deps = MICROKERNEL_BENCHMARK_DEPS,
3519)
3520
3521xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003522 name = "f16_igemm_bench",
3523 srcs = [
3524 "bench/f16-igemm.cc",
3525 "bench/conv.h",
3526 "bench/google/conv.h",
3527 "src/xnnpack/AlignedAllocator.h",
3528 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003529 deps = MICROKERNEL_BENCHMARK_DEPS + [
3530 ":indirection",
3531 ":packing",
3532 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003533)
3534
3535xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003536 name = "f16_gemm_bench",
3537 srcs = [
3538 "bench/f16-gemm.cc",
3539 "bench/gemm.h",
3540 "src/xnnpack/AlignedAllocator.h",
3541 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003542 deps = MICROKERNEL_BENCHMARK_DEPS + [
3543 ":packing",
3544 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003545)
3546
3547xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003548 name = "f16_spmm_bench",
3549 srcs = [
3550 "bench/f16-spmm.cc",
3551 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003552 "src/xnnpack/AlignedAllocator.h",
3553 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003554 deps = MICROKERNEL_BENCHMARK_DEPS,
3555)
3556
3557xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003558 name = "f32_igemm_bench",
3559 srcs = [
3560 "bench/f32-igemm.cc",
3561 "bench/conv.h",
3562 "src/xnnpack/AlignedAllocator.h",
3563 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003564 deps = MICROKERNEL_BENCHMARK_DEPS + [
3565 ":indirection",
3566 ":packing",
3567 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003568)
3569
3570xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07003571 name = "f16_relu_bench",
3572 srcs = [
3573 "bench/f16-relu.cc",
3574 "src/xnnpack/AlignedAllocator.h",
3575 ] + MICROKERNEL_BENCHMARK_HDRS,
3576 deps = MICROKERNEL_BENCHMARK_DEPS,
3577)
3578
3579xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003580 name = "f32_conv_hwc_bench",
3581 srcs = [
3582 "bench/f32-conv-hwc.cc",
3583 "bench/dconv.h",
3584 "src/xnnpack/AlignedAllocator.h",
3585 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003586 deps = MICROKERNEL_BENCHMARK_DEPS + [
3587 ":packing",
3588 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003589)
3590
3591xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003592 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003593 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003594 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003595 "bench/dconv.h",
3596 "src/xnnpack/AlignedAllocator.h",
3597 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003598 deps = MICROKERNEL_BENCHMARK_DEPS + [
3599 ":packing",
3600 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003601)
3602
3603xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003604 name = "f16_dwconv_bench",
3605 srcs = [
3606 "bench/f16-dwconv.cc",
3607 "bench/dwconv.h",
3608 "bench/google/dwconv.h",
3609 "src/xnnpack/AlignedAllocator.h",
3610 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003611 deps = MICROKERNEL_BENCHMARK_DEPS + [
3612 ":indirection",
3613 ":packing",
3614 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003615)
3616
3617xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003618 name = "f32_dwconv_bench",
3619 srcs = [
3620 "bench/f32-dwconv.cc",
3621 "bench/dwconv.h",
3622 "src/xnnpack/AlignedAllocator.h",
3623 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003624 deps = MICROKERNEL_BENCHMARK_DEPS + [
3625 ":indirection",
3626 ":packing",
3627 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003628)
3629
3630xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003631 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003632 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003633 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003634 "bench/dwconv.h",
3635 "src/xnnpack/AlignedAllocator.h",
3636 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003637 deps = MICROKERNEL_BENCHMARK_DEPS + [
3638 ":indirection",
3639 ":packing",
3640 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003641)
3642
3643xnnpack_benchmark(
3644 name = "f32_gemm_bench",
3645 srcs = [
3646 "bench/f32-gemm.cc",
3647 "bench/gemm.h",
3648 "src/xnnpack/AlignedAllocator.h",
3649 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003650 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003651 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003652)
3653
3654xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003655 name = "f32_hswish_bench",
3656 srcs = [
3657 "bench/f32-hswish.cc",
3658 "src/xnnpack/AlignedAllocator.h",
3659 ] + MICROKERNEL_BENCHMARK_HDRS,
3660 deps = MICROKERNEL_BENCHMARK_DEPS,
3661)
3662
3663xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003664 name = "f32_raddexpminusmax_bench",
3665 srcs = [
3666 "bench/f32-raddexpminusmax.cc",
3667 "src/xnnpack/AlignedAllocator.h",
3668 ] + MICROKERNEL_BENCHMARK_HDRS,
3669 deps = MICROKERNEL_BENCHMARK_DEPS,
3670)
3671
3672xnnpack_benchmark(
3673 name = "f32_raddextexp_bench",
3674 srcs = [
3675 "bench/f32-raddextexp.cc",
3676 "src/xnnpack/AlignedAllocator.h",
3677 ] + MICROKERNEL_BENCHMARK_HDRS,
3678 deps = MICROKERNEL_BENCHMARK_DEPS,
3679)
3680
3681xnnpack_benchmark(
3682 name = "f32_raddstoreexpminusmax_bench",
3683 srcs = [
3684 "bench/f32-raddstoreexpminusmax.cc",
3685 "src/xnnpack/AlignedAllocator.h",
3686 ] + MICROKERNEL_BENCHMARK_HDRS,
3687 deps = MICROKERNEL_BENCHMARK_DEPS,
3688)
3689
3690xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07003691 name = "f32_relu_bench",
3692 srcs = [
3693 "bench/f32-relu.cc",
3694 "src/xnnpack/AlignedAllocator.h",
3695 ] + MICROKERNEL_BENCHMARK_HDRS,
3696 deps = MICROKERNEL_BENCHMARK_DEPS,
3697)
3698
3699xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003700 name = "f32_rmax_bench",
3701 srcs = [
3702 "bench/f32-rmax.cc",
3703 "src/xnnpack/AlignedAllocator.h",
3704 ] + MICROKERNEL_BENCHMARK_HDRS,
3705 deps = MICROKERNEL_BENCHMARK_DEPS,
3706)
3707
3708xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003709 name = "f32_sigmoid_bench",
3710 srcs = [
3711 "bench/f32-sigmoid.cc",
3712 "src/xnnpack/AlignedAllocator.h",
3713 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003714 deps = MICROKERNEL_BENCHMARK_DEPS,
3715)
3716
3717xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003718 name = "f32_spmm_bench",
3719 srcs = [
3720 "bench/f32-spmm.cc",
3721 "bench/gemm.h",
3722 "src/xnnpack/AlignedAllocator.h",
3723 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003724 deps = MICROKERNEL_BENCHMARK_DEPS,
3725)
3726
3727xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003728 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003729 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003730 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003731 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003732 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003733 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003734)
3735
3736xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003737 name = "f32_vscaleexpminusmax_bench",
3738 srcs = [
3739 "bench/f32-vscaleexpminusmax.cc",
3740 "src/xnnpack/AlignedAllocator.h",
3741 ] + MICROKERNEL_BENCHMARK_HDRS,
3742 deps = MICROKERNEL_BENCHMARK_DEPS,
3743)
3744
3745xnnpack_benchmark(
3746 name = "f32_vscaleextexp_bench",
3747 srcs = [
3748 "bench/f32-vscaleextexp.cc",
3749 "src/xnnpack/AlignedAllocator.h",
3750 ] + MICROKERNEL_BENCHMARK_HDRS,
3751 deps = MICROKERNEL_BENCHMARK_DEPS,
3752)
3753
3754xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003755 name = "f32_vsqrt_bench",
3756 srcs = [
3757 "bench/f32-vsqrt.cc",
3758 "src/xnnpack/AlignedAllocator.h",
3759 ] + MICROKERNEL_BENCHMARK_HDRS,
3760 deps = MICROKERNEL_BENCHMARK_DEPS,
3761)
3762
3763xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003764 name = "f32_im2col_gemm_bench",
3765 srcs = [
3766 "bench/f32-im2col-gemm.cc",
3767 "bench/conv.h",
3768 "src/xnnpack/AlignedAllocator.h",
3769 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003770 deps = MICROKERNEL_BENCHMARK_DEPS + [
3771 ":im2col",
3772 ":packing",
3773 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003774)
3775
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003776xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07003777 name = "rounding_bench",
3778 srcs = [
3779 "bench/rounding.cc",
3780 "src/xnnpack/math-stubs.h",
3781 "src/xnnpack/AlignedAllocator.h",
3782 ] + MICROKERNEL_BENCHMARK_HDRS,
3783 deps = MICROKERNEL_BENCHMARK_DEPS,
3784)
3785
Marat Dukhan08c4a432019-10-03 09:29:21 -07003786########################### Benchmarks for operators ###########################
3787
3788xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003789 name = "average_pooling_bench",
3790 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07003791 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003792 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003793 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003794)
3795
3796xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003797 name = "bankers_rounding_bench",
3798 srcs = ["bench/bankers-rounding.cc"],
3799 copts = xnnpack_optional_tflite_copts(),
3800 tags = ["nowin32"],
3801 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3802)
3803
3804xnnpack_benchmark(
3805 name = "ceiling_bench",
3806 srcs = ["bench/ceiling.cc"],
3807 copts = xnnpack_optional_tflite_copts(),
3808 tags = ["nowin32"],
3809 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3810)
3811
3812xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003813 name = "channel_shuffle_bench",
3814 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003815 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003816)
3817
3818xnnpack_benchmark(
3819 name = "convolution_bench",
3820 srcs = ["bench/convolution.cc"],
3821 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003822 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003823 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003824)
3825
3826xnnpack_benchmark(
3827 name = "deconvolution_bench",
3828 srcs = ["bench/deconvolution.cc"],
3829 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003830 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003831 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003832)
3833
3834xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003835 name = "floor_bench",
3836 srcs = ["bench/floor.cc"],
3837 copts = xnnpack_optional_tflite_copts(),
3838 tags = ["nowin32"],
3839 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3840)
3841
3842xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003843 name = "global_average_pooling_bench",
3844 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003845 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003846)
3847
3848xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07003849 name = "hardswish_bench",
3850 srcs = ["bench/hardswish.cc"],
3851 copts = xnnpack_optional_tflite_copts(),
3852 tags = ["nowin32"],
3853 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3854)
3855
3856xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003857 name = "max_pooling_bench",
3858 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003859 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003860)
3861
3862xnnpack_benchmark(
3863 name = "sigmoid_bench",
3864 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08003865 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003866 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003867 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003868)
3869
3870xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07003871 name = "prelu_bench",
3872 srcs = ["bench/prelu.cc"],
3873 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07003874 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003875 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07003876)
3877
3878xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003879 name = "softmax_bench",
3880 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08003881 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07003882 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07003883 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003884)
3885
Marat Dukhan87727142020-06-24 15:24:10 -07003886xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07003887 name = "square_root_bench",
3888 srcs = ["bench/square-root.cc"],
3889 copts = xnnpack_optional_tflite_copts(),
3890 tags = ["nowin32"],
3891 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
3892)
3893
3894xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07003895 name = "truncation_bench",
3896 srcs = ["bench/truncation.cc"],
3897 deps = OPERATOR_BENCHMARK_DEPS,
3898)
3899
Marat Dukhanc068bb62019-10-04 13:24:39 -07003900############################# End-to-end benchmarks ############################
3901
3902cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003903 name = "fp32_mobilenet_v1",
3904 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003905 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003906 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003907 linkstatic = True,
3908 deps = [
3909 ":XNNPACK",
3910 "@pthreadpool",
3911 ],
3912)
3913
3914cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003915 name = "fp16_mobilenet_v1",
3916 srcs = ["models/fp16-mobilenet-v1.cc"],
3917 hdrs = ["models/models.h"],
3918 copts = xnnpack_std_cxxopts(),
3919 linkstatic = True,
3920 deps = [
3921 ":XNNPACK",
3922 "@FP16",
3923 "@pthreadpool",
3924 ],
3925)
3926
3927cc_library(
3928 name = "fp32_mobilenet_v2",
3929 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07003930 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003931 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07003932 linkstatic = True,
3933 deps = [
3934 ":XNNPACK",
3935 "@pthreadpool",
3936 ],
3937)
3938
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003939cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003940 name = "fp16_mobilenet_v2",
3941 srcs = ["models/fp16-mobilenet-v2.cc"],
3942 hdrs = ["models/models.h"],
3943 copts = xnnpack_std_cxxopts(),
3944 linkstatic = True,
3945 deps = [
3946 ":XNNPACK",
3947 "@FP16",
3948 "@pthreadpool",
3949 ],
3950)
3951
3952cc_library(
3953 name = "fp32_mobilenet_v3_large",
3954 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003955 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003956 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003957 linkstatic = True,
3958 deps = [
3959 ":XNNPACK",
3960 "@pthreadpool",
3961 ],
3962)
3963
3964cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07003965 name = "fp32_mobilenet_v3_small",
3966 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003967 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08003968 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003969 linkstatic = True,
3970 deps = [
3971 ":XNNPACK",
3972 "@pthreadpool",
3973 ],
3974)
3975
Marat Dukhanc068bb62019-10-04 13:24:39 -07003976xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07003977 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003978 srcs = [
3979 "bench/f32-dwconv-e2e.cc",
3980 "bench/end2end.h",
3981 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07003982 deps = MICROKERNEL_BENCHMARK_DEPS + [
3983 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003984 ":fp32_mobilenet_v1",
3985 ":fp32_mobilenet_v2",
3986 ":fp32_mobilenet_v3_large",
3987 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07003988 ],
3989)
3990
3991xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07003992 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08003993 srcs = [
3994 "bench/f32-gemm-e2e.cc",
3995 "bench/end2end.h",
3996 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07003997 deps = MICROKERNEL_BENCHMARK_DEPS + [
3998 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07003999 ":fp32_mobilenet_v1",
4000 ":fp32_mobilenet_v2",
4001 ":fp32_mobilenet_v3_large",
4002 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004003 ],
4004)
4005
4006xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004007 name = "end2end_bench",
4008 srcs = ["bench/end2end.cc"],
4009 deps = [
4010 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004011 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004012 ":fp16_mobilenet_v1",
4013 ":fp16_mobilenet_v2",
4014 ":fp32_mobilenet_v1",
4015 ":fp32_mobilenet_v2",
4016 ":fp32_mobilenet_v3_large",
4017 ":fp32_mobilenet_v3_small",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004018 "@pthreadpool",
4019 ],
4020)
4021
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004022#################### Accuracy evaluation for math functions ####################
4023
4024xnnpack_benchmark(
4025 name = "f32_exp_eval",
4026 srcs = [
4027 "eval/f32-exp.cc",
4028 "src/xnnpack/AlignedAllocator.h",
4029 ] + ACCURACY_EVAL_HDRS,
4030 deps = ACCURACY_EVAL_DEPS,
4031)
4032
Marat Dukhan515c9772019-10-17 18:07:57 -07004033xnnpack_benchmark(
4034 name = "f32_expminus_eval",
4035 srcs = [
4036 "eval/f32-expminus.cc",
4037 "src/xnnpack/AlignedAllocator.h",
4038 ] + ACCURACY_EVAL_HDRS,
4039 deps = ACCURACY_EVAL_DEPS,
4040)
4041
Marat Dukhan98ba4412019-10-23 02:14:28 -07004042xnnpack_benchmark(
4043 name = "f32_extexp_eval",
4044 srcs = [
4045 "eval/f32-extexp.cc",
4046 "src/xnnpack/AlignedAllocator.h",
4047 ] + ACCURACY_EVAL_HDRS,
4048 deps = ACCURACY_EVAL_DEPS,
4049)
4050
Marat Dukhan8853b822020-05-07 12:19:01 -07004051xnnpack_unit_test(
4052 name = "f32_roundne_eval",
4053 srcs = [
4054 "eval/f32-roundne.cc",
4055 "src/xnnpack/AlignedAllocator.h",
4056 "src/xnnpack/math-stubs.h",
4057 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004058 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004059 deps = MICROKERNEL_TEST_DEPS,
4060)
4061
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004062xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004063 name = "f32_roundd_eval",
4064 srcs = [
4065 "eval/f32-roundd.cc",
4066 "src/xnnpack/AlignedAllocator.h",
4067 "src/xnnpack/math-stubs.h",
4068 ] + MICROKERNEL_TEST_HDRS,
4069 automatic = False,
4070 deps = MICROKERNEL_TEST_DEPS,
4071)
4072
4073xnnpack_unit_test(
4074 name = "f32_roundu_eval",
4075 srcs = [
4076 "eval/f32-roundu.cc",
4077 "src/xnnpack/AlignedAllocator.h",
4078 "src/xnnpack/math-stubs.h",
4079 ] + MICROKERNEL_TEST_HDRS,
4080 automatic = False,
4081 deps = MICROKERNEL_TEST_DEPS,
4082)
4083
4084xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004085 name = "f32_roundz_eval",
4086 srcs = [
4087 "eval/f32-roundz.cc",
4088 "src/xnnpack/AlignedAllocator.h",
4089 "src/xnnpack/math-stubs.h",
4090 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004091 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004092 deps = MICROKERNEL_TEST_DEPS,
4093)
4094
Marat Dukhan346a9e52019-11-15 09:06:30 -08004095xnnpack_benchmark(
4096 name = "f32_sigmoid_eval",
4097 srcs = [
4098 "eval/f32-sigmoid.cc",
4099 "src/xnnpack/AlignedAllocator.h",
4100 ] + ACCURACY_EVAL_HDRS,
4101 deps = ACCURACY_EVAL_DEPS,
4102)
4103
Marat Dukhan84000762020-06-29 18:38:43 -07004104xnnpack_benchmark(
4105 name = "f32_sqrt_eval",
4106 srcs = [
4107 "eval/f32-sqrt.cc",
4108 "src/xnnpack/AlignedAllocator.h",
4109 ] + ACCURACY_EVAL_HDRS,
4110 deps = ACCURACY_EVAL_DEPS + [
4111 ":bench_utils",
4112 ],
4113)
4114
Marat Dukhan08c4a432019-10-03 09:29:21 -07004115######################### Unit tests for micro-kernels #########################
4116
4117xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004118 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004119 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004120 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004121 "test/gemm-microkernel-tester.h",
4122 "src/xnnpack/AlignedAllocator.h",
4123 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004124 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004125)
4126
4127xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004128 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004129 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004130 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004131 "test/spmm-microkernel-tester.h",
4132 "src/xnnpack/AlignedAllocator.h",
4133 ] + MICROKERNEL_TEST_HDRS,
4134 deps = MICROKERNEL_TEST_DEPS,
4135)
4136
4137xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004138 name = "f32_argmaxpool_test",
4139 srcs = [
4140 "test/f32-argmaxpool.cc",
4141 "test/argmaxpool-microkernel-tester.h",
4142 "src/xnnpack/AlignedAllocator.h",
4143 ] + MICROKERNEL_TEST_HDRS,
4144 deps = MICROKERNEL_TEST_DEPS,
4145)
4146
4147xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004148 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004149 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004150 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004151 "test/avgpool-microkernel-tester.h",
4152 "src/xnnpack/AlignedAllocator.h",
4153 ] + MICROKERNEL_TEST_HDRS,
4154 deps = MICROKERNEL_TEST_DEPS,
4155)
4156
4157xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004158 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004159 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004160 "test/f32-ibilinear.cc",
4161 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004162 "src/xnnpack/AlignedAllocator.h",
4163 ] + MICROKERNEL_TEST_HDRS,
4164 deps = MICROKERNEL_TEST_DEPS,
4165)
4166
4167xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004168 name = "f16_clamp_test",
4169 srcs = [
4170 "test/f16-clamp.cc",
4171 "test/clamp-microkernel-tester.h",
4172 ] + MICROKERNEL_TEST_HDRS,
4173 deps = MICROKERNEL_TEST_DEPS,
4174)
4175
4176xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004177 name = "f32_clamp_test",
4178 srcs = [
4179 "test/f32-clamp.cc",
4180 "test/clamp-microkernel-tester.h",
4181 ] + MICROKERNEL_TEST_HDRS,
4182 deps = MICROKERNEL_TEST_DEPS,
4183)
4184
4185xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004186 name = "f32_igemm_test",
4187 srcs = [
4188 "test/f32-igemm.cc",
4189 "test/gemm-microkernel-tester.h",
4190 "src/xnnpack/AlignedAllocator.h",
4191 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004192 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004193)
4194
4195xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004196 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004197 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004198 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004199 "test/gemm-microkernel-tester.h",
4200 "src/xnnpack/AlignedAllocator.h",
4201 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004202 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004203)
4204
4205xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004206 name = "f32_igemm_minmax_test",
4207 srcs = [
4208 "test/f32-igemm-minmax.cc",
4209 "test/gemm-microkernel-tester.h",
4210 "src/xnnpack/AlignedAllocator.h",
4211 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004212 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004213)
4214
4215xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004216 name = "f16_igemm_minmax_test",
4217 srcs = [
4218 "test/f16-igemm-minmax.cc",
4219 "test/gemm-microkernel-tester.h",
4220 "src/xnnpack/AlignedAllocator.h",
4221 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004222 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004223)
4224
4225xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004226 name = "f32_conv_hwc_test",
4227 srcs = [
4228 "test/f32-conv-hwc.cc",
4229 "test/conv-hwc-microkernel-tester.h",
4230 "src/xnnpack/AlignedAllocator.h",
4231 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004232 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004233)
4234
4235xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004236 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004237 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004238 "test/f32-conv-hwc2chw.cc",
4239 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004240 "src/xnnpack/AlignedAllocator.h",
4241 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004242 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004243)
4244
4245xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004246 name = "f32_dwconv_test",
4247 srcs = [
4248 "test/f32-dwconv.cc",
4249 "test/dwconv-microkernel-tester.h",
4250 "src/xnnpack/AlignedAllocator.h",
4251 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004252 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004253)
4254
4255xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004256 name = "f16_dwconv_minmax_test",
4257 srcs = [
4258 "test/f16-dwconv-minmax.cc",
4259 "test/dwconv-microkernel-tester.h",
4260 "src/xnnpack/AlignedAllocator.h",
4261 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004262 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004263)
4264
4265xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004266 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004267 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004268 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004269 "test/dwconv-microkernel-tester.h",
4270 "src/xnnpack/AlignedAllocator.h",
4271 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004272 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004273)
4274
4275xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004276 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004277 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004278 "test/f32-dwconv-chw.cc",
4279 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004280 "src/xnnpack/AlignedAllocator.h",
4281 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004282 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004283)
4284
4285xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004286 name = "f16_gavgpool_minmax_test",
4287 srcs = [
4288 "test/f16-gavgpool-minmax.cc",
4289 "test/gavgpool-microkernel-tester.h",
4290 "src/xnnpack/AlignedAllocator.h",
4291 ] + MICROKERNEL_TEST_HDRS,
4292 deps = MICROKERNEL_TEST_DEPS,
4293)
4294
4295xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004296 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004297 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004298 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004299 "test/gavgpool-microkernel-tester.h",
4300 "src/xnnpack/AlignedAllocator.h",
4301 ] + MICROKERNEL_TEST_HDRS,
4302 deps = MICROKERNEL_TEST_DEPS,
4303)
4304
4305xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004306 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004307 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004308 "test/f32-gavgpool-cw.cc",
4309 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004310 "src/xnnpack/AlignedAllocator.h",
4311 ] + MICROKERNEL_TEST_HDRS,
4312 deps = MICROKERNEL_TEST_DEPS,
4313)
4314
4315xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004316 name = "f32_gemm_test",
4317 srcs = [
4318 "test/f32-gemm.cc",
4319 "test/gemm-microkernel-tester.h",
4320 "src/xnnpack/AlignedAllocator.h",
4321 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004322 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004323)
4324
4325xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004326 name = "f32_gemm_relu_test",
4327 srcs = [
4328 "test/f32-gemm-relu.cc",
4329 "test/gemm-microkernel-tester.h",
4330 "src/xnnpack/AlignedAllocator.h",
4331 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004332 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004333)
4334
4335xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004336 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004337 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004338 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004339 "test/gemm-microkernel-tester.h",
4340 "src/xnnpack/AlignedAllocator.h",
4341 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004342 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004343)
4344
4345xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004346 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004347 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004348 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004349 "test/gemm-microkernel-tester.h",
4350 "src/xnnpack/AlignedAllocator.h",
4351 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004352 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004353)
4354
4355xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004356 name = "f16_hswish_test",
4357 srcs = [
4358 "test/f16-hswish.cc",
4359 "test/hswish-microkernel-tester.h",
4360 ] + MICROKERNEL_TEST_HDRS,
4361 deps = MICROKERNEL_TEST_DEPS,
4362)
4363
4364xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004365 name = "f32_hswish_test",
4366 srcs = [
4367 "test/f32-hswish.cc",
4368 "test/hswish-microkernel-tester.h",
4369 ] + MICROKERNEL_TEST_HDRS,
4370 deps = MICROKERNEL_TEST_DEPS,
4371)
4372
4373xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004374 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004375 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004376 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004377 "test/maxpool-microkernel-tester.h",
4378 ] + MICROKERNEL_TEST_HDRS,
4379 deps = MICROKERNEL_TEST_DEPS,
4380)
4381
4382xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004383 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004384 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004385 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004386 "test/avgpool-microkernel-tester.h",
4387 "src/xnnpack/AlignedAllocator.h",
4388 ] + MICROKERNEL_TEST_HDRS,
4389 deps = MICROKERNEL_TEST_DEPS,
4390)
4391
4392xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004393 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004394 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004395 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004396 "test/gemm-microkernel-tester.h",
4397 "src/xnnpack/AlignedAllocator.h",
4398 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004399 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004400)
4401
4402xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004403 name = "f16_prelu_test",
4404 srcs = [
4405 "test/f16-prelu.cc",
4406 "test/prelu-microkernel-tester.h",
4407 "src/xnnpack/AlignedAllocator.h",
4408 ] + MICROKERNEL_TEST_HDRS,
4409 deps = MICROKERNEL_TEST_DEPS,
4410)
4411
4412xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004413 name = "f32_prelu_test",
4414 srcs = [
4415 "test/f32-prelu.cc",
4416 "test/prelu-microkernel-tester.h",
4417 "src/xnnpack/AlignedAllocator.h",
4418 ] + MICROKERNEL_TEST_HDRS,
4419 deps = MICROKERNEL_TEST_DEPS,
4420)
4421
4422xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004423 name = "f32_raddexpminusmax_test",
4424 srcs = [
4425 "test/f32-raddexpminusmax.cc",
4426 "test/raddexpminusmax-microkernel-tester.h",
4427 ] + MICROKERNEL_TEST_HDRS,
4428 deps = MICROKERNEL_TEST_DEPS,
4429)
4430
4431xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004432 name = "f32_raddextexp_test",
4433 srcs = [
4434 "test/f32-raddextexp.cc",
4435 "test/raddextexp-microkernel-tester.h",
4436 ] + MICROKERNEL_TEST_HDRS,
4437 deps = MICROKERNEL_TEST_DEPS,
4438)
4439
4440xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004441 name = "f32_raddstoreexpminusmax_test",
4442 srcs = [
4443 "test/f32-raddstoreexpminusmax.cc",
4444 "test/raddstoreexpminusmax-microkernel-tester.h",
4445 ] + MICROKERNEL_TEST_HDRS,
4446 deps = MICROKERNEL_TEST_DEPS,
4447)
4448
4449xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004450 name = "f32_relu_test",
4451 srcs = [
4452 "test/f32-relu.cc",
4453 "test/vunary-microkernel-tester.h",
4454 ] + MICROKERNEL_TEST_HDRS,
4455 deps = MICROKERNEL_TEST_DEPS,
4456)
4457
4458xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004459 name = "f32_rmax_test",
4460 srcs = [
4461 "test/f32-rmax.cc",
4462 "test/rmax-microkernel-tester.h",
4463 ] + MICROKERNEL_TEST_HDRS,
4464 deps = MICROKERNEL_TEST_DEPS,
4465)
4466
4467xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004468 name = "f32_sigmoid_test",
4469 srcs = [
4470 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004471 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004472 ] + MICROKERNEL_TEST_HDRS,
4473 deps = MICROKERNEL_TEST_DEPS,
4474)
4475
4476xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004477 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004478 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004479 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004480 "test/spmm-microkernel-tester.h",
4481 "src/xnnpack/AlignedAllocator.h",
4482 ] + MICROKERNEL_TEST_HDRS,
4483 deps = MICROKERNEL_TEST_DEPS,
4484)
4485
4486xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004487 name = "f16_vadd_minmax_test",
4488 srcs = [
4489 "test/f16-vadd-minmax.cc",
4490 "test/vbinary-microkernel-tester.h",
4491 ] + MICROKERNEL_TEST_HDRS,
4492 deps = MICROKERNEL_TEST_DEPS,
4493)
4494
4495xnnpack_unit_test(
4496 name = "f16_vaddc_minmax_test",
4497 srcs = [
4498 "test/f16-vaddc-minmax.cc",
4499 "test/vbinaryc-microkernel-tester.h",
4500 ] + MICROKERNEL_TEST_HDRS,
4501 deps = MICROKERNEL_TEST_DEPS,
4502)
4503
4504xnnpack_unit_test(
4505 name = "f16_vdiv_minmax_test",
4506 srcs = [
4507 "test/f16-vdiv-minmax.cc",
4508 "test/vbinary-microkernel-tester.h",
4509 ] + MICROKERNEL_TEST_HDRS,
4510 deps = MICROKERNEL_TEST_DEPS,
4511)
4512
4513xnnpack_unit_test(
4514 name = "f16_vdivc_minmax_test",
4515 srcs = [
4516 "test/f16-vdivc-minmax.cc",
4517 "test/vbinaryc-microkernel-tester.h",
4518 ] + MICROKERNEL_TEST_HDRS,
4519 deps = MICROKERNEL_TEST_DEPS,
4520)
4521
4522xnnpack_unit_test(
4523 name = "f16_vrdivc_minmax_test",
4524 srcs = [
4525 "test/f16-vrdivc-minmax.cc",
4526 "test/vbinaryc-microkernel-tester.h",
4527 ] + MICROKERNEL_TEST_HDRS,
4528 deps = MICROKERNEL_TEST_DEPS,
4529)
4530
4531xnnpack_unit_test(
4532 name = "f16_vmax_test",
4533 srcs = [
4534 "test/f16-vmax.cc",
4535 "test/vbinary-microkernel-tester.h",
4536 ] + MICROKERNEL_TEST_HDRS,
4537 deps = MICROKERNEL_TEST_DEPS,
4538)
4539
4540xnnpack_unit_test(
4541 name = "f16_vmaxc_test",
4542 srcs = [
4543 "test/f16-vmaxc.cc",
4544 "test/vbinaryc-microkernel-tester.h",
4545 ] + MICROKERNEL_TEST_HDRS,
4546 deps = MICROKERNEL_TEST_DEPS,
4547)
4548
4549xnnpack_unit_test(
4550 name = "f16_vmin_test",
4551 srcs = [
4552 "test/f16-vmin.cc",
4553 "test/vbinary-microkernel-tester.h",
4554 ] + MICROKERNEL_TEST_HDRS,
4555 deps = MICROKERNEL_TEST_DEPS,
4556)
4557
4558xnnpack_unit_test(
4559 name = "f16_vminc_test",
4560 srcs = [
4561 "test/f16-vminc.cc",
4562 "test/vbinaryc-microkernel-tester.h",
4563 ] + MICROKERNEL_TEST_HDRS,
4564 deps = MICROKERNEL_TEST_DEPS,
4565)
4566
4567xnnpack_unit_test(
4568 name = "f16_vmul_minmax_test",
4569 srcs = [
4570 "test/f16-vmul-minmax.cc",
4571 "test/vbinary-microkernel-tester.h",
4572 ] + MICROKERNEL_TEST_HDRS,
4573 deps = MICROKERNEL_TEST_DEPS,
4574)
4575
4576xnnpack_unit_test(
4577 name = "f16_vmulc_minmax_test",
4578 srcs = [
4579 "test/f16-vmulc-minmax.cc",
4580 "test/vbinaryc-microkernel-tester.h",
4581 ] + MICROKERNEL_TEST_HDRS,
4582 deps = MICROKERNEL_TEST_DEPS,
4583)
4584
4585xnnpack_unit_test(
4586 name = "f16_vsub_minmax_test",
4587 srcs = [
4588 "test/f16-vsub-minmax.cc",
4589 "test/vbinary-microkernel-tester.h",
4590 ] + MICROKERNEL_TEST_HDRS,
4591 deps = MICROKERNEL_TEST_DEPS,
4592)
4593
4594xnnpack_unit_test(
4595 name = "f16_vsubc_minmax_test",
4596 srcs = [
4597 "test/f16-vsubc-minmax.cc",
4598 "test/vbinaryc-microkernel-tester.h",
4599 ] + MICROKERNEL_TEST_HDRS,
4600 deps = MICROKERNEL_TEST_DEPS,
4601)
4602
4603xnnpack_unit_test(
4604 name = "f16_vrsubc_minmax_test",
4605 srcs = [
4606 "test/f16-vrsubc-minmax.cc",
4607 "test/vbinaryc-microkernel-tester.h",
4608 ] + MICROKERNEL_TEST_HDRS,
4609 deps = MICROKERNEL_TEST_DEPS,
4610)
4611
4612xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004613 name = "f32_vabs_test",
4614 srcs = [
4615 "test/f32-vabs.cc",
4616 "test/vunary-microkernel-tester.h",
4617 ] + MICROKERNEL_TEST_HDRS,
4618 deps = MICROKERNEL_TEST_DEPS,
4619)
4620
4621xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004622 name = "f32_vadd_test",
4623 srcs = [
4624 "test/f32-vadd.cc",
4625 "test/vbinary-microkernel-tester.h",
4626 ] + MICROKERNEL_TEST_HDRS,
4627 deps = MICROKERNEL_TEST_DEPS,
4628)
4629
4630xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004631 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004632 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004633 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004634 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004635 ] + MICROKERNEL_TEST_HDRS,
4636 deps = MICROKERNEL_TEST_DEPS,
4637)
4638
4639xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004640 name = "f32_vaddc_test",
4641 srcs = [
4642 "test/f32-vaddc.cc",
4643 "test/vbinaryc-microkernel-tester.h",
4644 ] + MICROKERNEL_TEST_HDRS,
4645 deps = MICROKERNEL_TEST_DEPS,
4646)
4647
4648xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004649 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004650 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004651 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004652 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004653 ] + MICROKERNEL_TEST_HDRS,
4654 deps = MICROKERNEL_TEST_DEPS,
4655)
4656
4657xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004658 name = "f32_vdiv_test",
4659 srcs = [
4660 "test/f32-vdiv.cc",
4661 "test/vbinary-microkernel-tester.h",
4662 ] + MICROKERNEL_TEST_HDRS,
4663 deps = MICROKERNEL_TEST_DEPS,
4664)
4665
4666xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004667 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004668 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004669 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004670 "test/vbinary-microkernel-tester.h",
4671 ] + MICROKERNEL_TEST_HDRS,
4672 deps = MICROKERNEL_TEST_DEPS,
4673)
4674
4675xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004676 name = "f32_vdivc_test",
4677 srcs = [
4678 "test/f32-vdivc.cc",
4679 "test/vbinaryc-microkernel-tester.h",
4680 ] + MICROKERNEL_TEST_HDRS,
4681 deps = MICROKERNEL_TEST_DEPS,
4682)
4683
4684xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004685 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004686 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004687 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004688 "test/vbinaryc-microkernel-tester.h",
4689 ] + MICROKERNEL_TEST_HDRS,
4690 deps = MICROKERNEL_TEST_DEPS,
4691)
4692
4693xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004694 name = "f32_vrdivc_test",
4695 srcs = [
4696 "test/f32-vrdivc.cc",
4697 "test/vbinaryc-microkernel-tester.h",
4698 ] + MICROKERNEL_TEST_HDRS,
4699 deps = MICROKERNEL_TEST_DEPS,
4700)
4701
4702xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004703 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004704 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004705 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004706 "test/vbinaryc-microkernel-tester.h",
4707 ] + MICROKERNEL_TEST_HDRS,
4708 deps = MICROKERNEL_TEST_DEPS,
4709)
4710
4711xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004712 name = "f32_vmax_test",
4713 srcs = [
4714 "test/f32-vmax.cc",
4715 "test/vbinary-microkernel-tester.h",
4716 ] + MICROKERNEL_TEST_HDRS,
4717 deps = MICROKERNEL_TEST_DEPS,
4718)
4719
4720xnnpack_unit_test(
4721 name = "f32_vmaxc_test",
4722 srcs = [
4723 "test/f32-vmaxc.cc",
4724 "test/vbinaryc-microkernel-tester.h",
4725 ] + MICROKERNEL_TEST_HDRS,
4726 deps = MICROKERNEL_TEST_DEPS,
4727)
4728
4729xnnpack_unit_test(
4730 name = "f32_vmin_test",
4731 srcs = [
4732 "test/f32-vmin.cc",
4733 "test/vbinary-microkernel-tester.h",
4734 ] + MICROKERNEL_TEST_HDRS,
4735 deps = MICROKERNEL_TEST_DEPS,
4736)
4737
4738xnnpack_unit_test(
4739 name = "f32_vminc_test",
4740 srcs = [
4741 "test/f32-vminc.cc",
4742 "test/vbinaryc-microkernel-tester.h",
4743 ] + MICROKERNEL_TEST_HDRS,
4744 deps = MICROKERNEL_TEST_DEPS,
4745)
4746
4747xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004748 name = "f32_vmul_test",
4749 srcs = [
4750 "test/f32-vmul.cc",
4751 "test/vbinary-microkernel-tester.h",
4752 ] + MICROKERNEL_TEST_HDRS,
4753 deps = MICROKERNEL_TEST_DEPS,
4754)
4755
4756xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004757 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004758 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004759 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004760 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004761 ] + MICROKERNEL_TEST_HDRS,
4762 deps = MICROKERNEL_TEST_DEPS,
4763)
4764
4765xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004766 name = "f32_vmulc_test",
4767 srcs = [
4768 "test/f32-vmulc.cc",
4769 "test/vbinaryc-microkernel-tester.h",
4770 ] + MICROKERNEL_TEST_HDRS,
4771 deps = MICROKERNEL_TEST_DEPS,
4772)
4773
4774xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004775 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004776 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004777 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004778 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004779 ] + MICROKERNEL_TEST_HDRS,
4780 deps = MICROKERNEL_TEST_DEPS,
4781)
4782
4783xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07004784 name = "f16_vmulcaddc_minmax_test",
4785 srcs = [
4786 "test/f16-vmulcaddc-minmax.cc",
4787 "test/vmulcaddc-microkernel-tester.h",
4788 "src/xnnpack/AlignedAllocator.h",
4789 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004790 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07004791)
4792
4793xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004794 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004795 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004796 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004797 "test/vmulcaddc-microkernel-tester.h",
4798 "src/xnnpack/AlignedAllocator.h",
4799 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004800 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004801)
4802
4803xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07004804 name = "f32_vlrelu_test",
4805 srcs = [
4806 "test/f32-vlrelu.cc",
4807 "test/vunary-microkernel-tester.h",
4808 ] + MICROKERNEL_TEST_HDRS,
4809 deps = MICROKERNEL_TEST_DEPS,
4810)
4811
4812xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004813 name = "f32_vneg_test",
4814 srcs = [
4815 "test/f32-vneg.cc",
4816 "test/vunary-microkernel-tester.h",
4817 ] + MICROKERNEL_TEST_HDRS,
4818 deps = MICROKERNEL_TEST_DEPS,
4819)
4820
4821xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07004822 name = "f32_vrndne_test",
4823 srcs = [
4824 "test/f32-vrndne.cc",
4825 "test/vunary-microkernel-tester.h",
4826 ] + MICROKERNEL_TEST_HDRS,
4827 deps = MICROKERNEL_TEST_DEPS,
4828)
4829
4830xnnpack_unit_test(
4831 name = "f32_vrndz_test",
4832 srcs = [
4833 "test/f32-vrndz.cc",
4834 "test/vunary-microkernel-tester.h",
4835 ] + MICROKERNEL_TEST_HDRS,
4836 deps = MICROKERNEL_TEST_DEPS,
4837)
4838
4839xnnpack_unit_test(
4840 name = "f32_vrndu_test",
4841 srcs = [
4842 "test/f32-vrndu.cc",
4843 "test/vunary-microkernel-tester.h",
4844 ] + MICROKERNEL_TEST_HDRS,
4845 deps = MICROKERNEL_TEST_DEPS,
4846)
4847
4848xnnpack_unit_test(
4849 name = "f32_vrndd_test",
4850 srcs = [
4851 "test/f32-vrndd.cc",
4852 "test/vunary-microkernel-tester.h",
4853 ] + MICROKERNEL_TEST_HDRS,
4854 deps = MICROKERNEL_TEST_DEPS,
4855)
4856
4857xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07004858 name = "f32_vscale_test",
4859 srcs = [
4860 "test/f32-vscale.cc",
4861 "test/vscale-microkernel-tester.h",
4862 ] + MICROKERNEL_TEST_HDRS,
4863 deps = MICROKERNEL_TEST_DEPS,
4864)
4865
4866xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004867 name = "f32_vscaleexpminusmax_test",
4868 srcs = [
4869 "test/f32-vscaleexpminusmax.cc",
4870 "test/vscaleexpminusmax-microkernel-tester.h",
4871 ] + MICROKERNEL_TEST_HDRS,
4872 deps = MICROKERNEL_TEST_DEPS,
4873)
4874
4875xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004876 name = "f32_vscaleextexp_test",
4877 srcs = [
4878 "test/f32-vscaleextexp.cc",
4879 "test/vscaleextexp-microkernel-tester.h",
4880 ] + MICROKERNEL_TEST_HDRS,
4881 deps = MICROKERNEL_TEST_DEPS,
4882)
4883
4884xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004885 name = "f32_vsqr_test",
4886 srcs = [
4887 "test/f32-vsqr.cc",
4888 "test/vunary-microkernel-tester.h",
4889 ] + MICROKERNEL_TEST_HDRS,
4890 deps = MICROKERNEL_TEST_DEPS,
4891)
4892
4893xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07004894 name = "f32_vsqrdiff_test",
4895 srcs = [
4896 "test/f32-vsqrdiff.cc",
4897 "test/vbinary-microkernel-tester.h",
4898 ] + MICROKERNEL_TEST_HDRS,
4899 deps = MICROKERNEL_TEST_DEPS,
4900)
4901
4902xnnpack_unit_test(
4903 name = "f32_vsqrdiffc_test",
4904 srcs = [
4905 "test/f32-vsqrdiffc.cc",
4906 "test/vbinaryc-microkernel-tester.h",
4907 ] + MICROKERNEL_TEST_HDRS,
4908 deps = MICROKERNEL_TEST_DEPS,
4909)
4910
4911xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004912 name = "f32_vsqrt_test",
4913 srcs = [
4914 "test/f32-vsqrt.cc",
4915 "test/vunary-microkernel-tester.h",
4916 ] + MICROKERNEL_TEST_HDRS,
4917 deps = MICROKERNEL_TEST_DEPS,
4918)
4919
4920xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004921 name = "f32_vsub_test",
4922 srcs = [
4923 "test/f32-vsub.cc",
4924 "test/vbinary-microkernel-tester.h",
4925 ] + MICROKERNEL_TEST_HDRS,
4926 deps = MICROKERNEL_TEST_DEPS,
4927)
4928
4929xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004930 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07004931 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004932 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004933 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004934 ] + MICROKERNEL_TEST_HDRS,
4935 deps = MICROKERNEL_TEST_DEPS,
4936)
4937
4938xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004939 name = "f32_vsubc_test",
4940 srcs = [
4941 "test/f32-vsubc.cc",
4942 "test/vbinaryc-microkernel-tester.h",
4943 ] + MICROKERNEL_TEST_HDRS,
4944 deps = MICROKERNEL_TEST_DEPS,
4945)
4946
4947xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004948 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004949 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004950 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004951 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004952 ] + MICROKERNEL_TEST_HDRS,
4953 deps = MICROKERNEL_TEST_DEPS,
4954)
4955
4956xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004957 name = "f32_vrsubc_test",
4958 srcs = [
4959 "test/f32-vrsubc.cc",
4960 "test/vbinaryc-microkernel-tester.h",
4961 ] + MICROKERNEL_TEST_HDRS,
4962 deps = MICROKERNEL_TEST_DEPS,
4963)
4964
4965xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004966 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004967 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004968 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004969 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07004970 ] + MICROKERNEL_TEST_HDRS,
4971 deps = MICROKERNEL_TEST_DEPS,
4972)
4973
4974xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004975 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004976 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004977 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004978 "test/avgpool-microkernel-tester.h",
4979 "src/xnnpack/AlignedAllocator.h",
4980 ] + MICROKERNEL_TEST_HDRS,
4981 deps = MICROKERNEL_TEST_DEPS,
4982)
4983
4984xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004985 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004986 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004987 "test/qu8-igemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004988 "test/gemm-microkernel-tester.h",
4989 "src/xnnpack/AlignedAllocator.h",
4990 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004991 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004992)
4993
4994xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004995 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004996 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07004997 "test/qu8-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004998 "test/dwconv-microkernel-tester.h",
4999 "src/xnnpack/AlignedAllocator.h",
5000 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005001 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005002)
5003
5004xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005005 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005006 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005007 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005008 "test/gavgpool-microkernel-tester.h",
5009 "src/xnnpack/AlignedAllocator.h",
5010 ] + MICROKERNEL_TEST_HDRS,
5011 deps = MICROKERNEL_TEST_DEPS,
5012)
5013
5014xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005015 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005016 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005017 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005018 "test/gemm-microkernel-tester.h",
5019 "src/xnnpack/AlignedAllocator.h",
5020 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005021 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005022)
5023
5024xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005025 name = "qu8_requantization_test",
5026 srcs = [
5027 "src/xnnpack/requantization-stubs.h",
5028 "test/qu8-requantization.cc",
5029 "test/requantization-tester.h",
5030 ] + MICROKERNEL_TEST_HDRS,
5031 deps = MICROKERNEL_TEST_DEPS,
5032)
5033
5034xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005035 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005036 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005037 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005038 "test/vadd-microkernel-tester.h",
5039 ] + MICROKERNEL_TEST_HDRS,
5040 deps = MICROKERNEL_TEST_DEPS,
5041)
5042
5043xnnpack_unit_test(
5044 name = "u8_clamp_test",
5045 srcs = [
5046 "test/u8-clamp.cc",
5047 "test/clamp-microkernel-tester.h",
5048 ] + MICROKERNEL_TEST_HDRS,
5049 deps = MICROKERNEL_TEST_DEPS,
5050)
5051
5052xnnpack_unit_test(
5053 name = "u8_lut32norm_test",
5054 srcs = [
5055 "test/u8-lut32norm.cc",
5056 "test/lut-norm-microkernel-tester.h",
5057 ] + MICROKERNEL_TEST_HDRS,
5058 deps = MICROKERNEL_TEST_DEPS,
5059)
5060
5061xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005062 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005063 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005064 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005065 "test/maxpool-microkernel-tester.h",
5066 ] + MICROKERNEL_TEST_HDRS,
5067 deps = MICROKERNEL_TEST_DEPS,
5068)
5069
5070xnnpack_unit_test(
5071 name = "u8_rmax_test",
5072 srcs = [
5073 "test/u8-rmax.cc",
5074 "test/rmax-microkernel-tester.h",
5075 ] + MICROKERNEL_TEST_HDRS,
5076 deps = MICROKERNEL_TEST_DEPS,
5077)
5078
5079xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005080 name = "x32_fill_test",
5081 srcs = [
5082 "test/x32-fill.cc",
5083 "test/fill-microkernel-tester.h",
5084 ] + MICROKERNEL_TEST_HDRS,
5085 deps = MICROKERNEL_TEST_DEPS,
5086)
5087
5088xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005089 name = "x32_packx_test",
5090 srcs = [
5091 "test/x32-packx.cc",
5092 "test/pack-microkernel-tester.h",
5093 "src/xnnpack/AlignedAllocator.h",
5094 ] + MICROKERNEL_TEST_HDRS,
5095 deps = MICROKERNEL_TEST_DEPS,
5096)
5097
5098xnnpack_unit_test(
5099 name = "x32_pad_test",
5100 srcs = [
5101 "test/x32-pad.cc",
5102 "test/pad-microkernel-tester.h",
5103 ] + MICROKERNEL_TEST_HDRS,
5104 deps = MICROKERNEL_TEST_DEPS,
5105)
5106
5107xnnpack_unit_test(
5108 name = "x32_unpool_test",
5109 srcs = [
5110 "test/x32-unpool.cc",
5111 "test/unpool-microkernel-tester.h",
5112 ] + MICROKERNEL_TEST_HDRS,
5113 deps = MICROKERNEL_TEST_DEPS,
5114)
5115
5116xnnpack_unit_test(
5117 name = "x32_zip_test",
5118 srcs = [
5119 "test/x32-zip.cc",
5120 "test/zip-microkernel-tester.h",
5121 ] + MICROKERNEL_TEST_HDRS,
5122 deps = MICROKERNEL_TEST_DEPS,
5123)
5124
5125xnnpack_unit_test(
5126 name = "x8_lut_test",
5127 srcs = [
5128 "test/x8-lut.cc",
5129 "test/lut-microkernel-tester.h",
5130 ] + MICROKERNEL_TEST_HDRS,
5131 deps = MICROKERNEL_TEST_DEPS,
5132)
5133
5134xnnpack_unit_test(
5135 name = "x8_zip_test",
5136 srcs = [
5137 "test/x8-zip.cc",
5138 "test/zip-microkernel-tester.h",
5139 ] + MICROKERNEL_TEST_HDRS,
5140 deps = MICROKERNEL_TEST_DEPS,
5141)
5142
Marat Dukhan20c3b922020-03-10 03:45:06 -07005143########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005144
5145xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005146 name = "operator_size_test",
5147 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005148 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005149)
5150
Marat Dukhan20c3b922020-03-10 03:45:06 -07005151xnnpack_binary(
5152 name = "subgraph_size_test",
5153 srcs = ["test/subgraph-size.c"],
5154 deps = [":XNNPACK"],
5155)
5156
5157########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005158
5159xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005160 name = "abs_nc_test",
5161 srcs = [
5162 "test/abs-nc.cc",
5163 "test/abs-operator-tester.h",
5164 ],
5165 deps = OPERATOR_TEST_DEPS,
5166)
5167
5168xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005169 name = "add_nd_test",
5170 srcs = [
5171 "test/add-nd.cc",
5172 "test/binary-elementwise-operator-tester.h",
5173 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005174 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005175)
5176
5177xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005178 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005179 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005180 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005181 "test/argmax-pooling-operator-tester.h",
5182 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005183 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005184)
5185
5186xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005187 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005188 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005189 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005190 "test/average-pooling-operator-tester.h",
5191 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005192 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005193)
5194
5195xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005196 name = "bankers_rounding_nc_test",
5197 srcs = [
5198 "test/bankers-rounding-nc.cc",
5199 "test/bankers-rounding-operator-tester.h",
5200 ],
5201 deps = OPERATOR_TEST_DEPS,
5202)
5203
5204xnnpack_unit_test(
5205 name = "ceiling_nc_test",
5206 srcs = [
5207 "test/ceiling-nc.cc",
5208 "test/ceiling-operator-tester.h",
5209 ],
5210 deps = OPERATOR_TEST_DEPS,
5211)
5212
5213xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005214 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005215 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005216 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005217 "test/channel-shuffle-operator-tester.h",
5218 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005219 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005220)
5221
5222xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005223 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005224 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005225 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005226 "test/clamp-operator-tester.h",
5227 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005228 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005229)
5230
5231xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005232 name = "constant_pad_nd_test",
5233 srcs = [
5234 "test/constant-pad-nd.cc",
5235 "test/constant-pad-operator-tester.h",
5236 ],
5237 deps = OPERATOR_TEST_DEPS,
5238)
5239
5240xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005241 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005242 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005243 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005244 "test/convolution-operator-tester.h",
5245 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005246 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005247)
5248
5249xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005250 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005251 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005252 "test/convolution-nchw.cc",
5253 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005254 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005255 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005256)
5257
5258xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005259 name = "copy_nc_test",
5260 srcs = [
5261 "test/copy-nc.cc",
5262 "test/copy-operator-tester.h",
5263 ],
5264 deps = OPERATOR_TEST_DEPS,
5265)
5266
5267xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005268 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005269 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005270 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005271 "test/deconvolution-operator-tester.h",
5272 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005273 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005274)
5275
5276xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005277 name = "divide_nd_test",
5278 srcs = [
5279 "test/binary-elementwise-operator-tester.h",
5280 "test/divide-nd.cc",
5281 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005282 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005283)
5284
5285xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005286 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005287 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005288 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005289 "test/fully-connected-operator-tester.h",
5290 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005291 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005292)
5293
5294xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005295 name = "floor_nc_test",
5296 srcs = [
5297 "test/floor-nc.cc",
5298 "test/floor-operator-tester.h",
5299 ],
5300 deps = OPERATOR_TEST_DEPS,
5301)
5302
5303xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005304 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005305 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005306 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005307 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005308 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005309 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005310)
5311
5312xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005313 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005314 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005315 "test/global-average-pooling-ncw.cc",
5316 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005317 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005318 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005319)
5320
5321xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005322 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005323 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005324 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005325 "test/hardswish-operator-tester.h",
5326 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005327 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005328)
5329
5330xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005331 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005332 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005333 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005334 "test/leaky-relu-operator-tester.h",
5335 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005336 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005337)
5338
5339xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005340 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005341 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005342 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005343 "test/max-pooling-operator-tester.h",
5344 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005345 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005346)
5347
5348xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005349 name = "maximum_nd_test",
5350 srcs = [
5351 "test/binary-elementwise-operator-tester.h",
5352 "test/maximum-nd.cc",
5353 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005354 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005355)
5356
5357xnnpack_unit_test(
5358 name = "minimum_nd_test",
5359 srcs = [
5360 "test/binary-elementwise-operator-tester.h",
5361 "test/minimum-nd.cc",
5362 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005363 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005364)
5365
5366xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005367 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005368 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005369 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005370 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005371 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005372 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005373)
5374
5375xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005376 name = "negate_nc_test",
5377 srcs = [
5378 "test/negate-nc.cc",
5379 "test/negate-operator-tester.h",
5380 ],
5381 deps = OPERATOR_TEST_DEPS,
5382)
5383
5384xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005385 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005386 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005387 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005388 "test/prelu-operator-tester.h",
5389 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005390 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005391)
5392
5393xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005394 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005395 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005396 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005397 "test/resize-bilinear-operator-tester.h",
5398 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005399 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005400)
5401
5402xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005403 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005404 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005405 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005406 "test/sigmoid-operator-tester.h",
5407 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005408 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005409)
5410
5411xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005412 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005413 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005414 "test/softmax-nc.cc",
5415 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005416 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005417 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005418)
5419
5420xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005421 name = "square_nc_test",
5422 srcs = [
5423 "test/square-nc.cc",
5424 "test/square-operator-tester.h",
5425 ],
5426 deps = OPERATOR_TEST_DEPS,
5427)
5428
5429xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005430 name = "square_root_nc_test",
5431 srcs = [
5432 "test/square-root-nc.cc",
5433 "test/square-root-operator-tester.h",
5434 ],
5435 deps = OPERATOR_TEST_DEPS,
5436)
5437
5438xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005439 name = "squared_difference_nd_test",
5440 srcs = [
5441 "test/binary-elementwise-operator-tester.h",
5442 "test/squared-difference-nd.cc",
5443 ],
5444 deps = OPERATOR_TEST_DEPS,
5445)
5446
5447xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005448 name = "subtract_nd_test",
5449 srcs = [
5450 "test/binary-elementwise-operator-tester.h",
5451 "test/subtract-nd.cc",
5452 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005453 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005454)
5455
5456xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005457 name = "truncation_nc_test",
5458 srcs = [
5459 "test/truncation-nc.cc",
5460 "test/truncation-operator-tester.h",
5461 ],
5462 deps = OPERATOR_TEST_DEPS,
5463)
5464
5465xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005466 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005467 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005468 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005469 "test/unpooling-operator-tester.h",
5470 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005471 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005472)
5473
Chao Mei6ddfc602020-05-13 22:29:36 -07005474############################### Misc unit tests ###############################
5475
5476xnnpack_unit_test(
5477 name = "memory_planner_test",
5478 srcs = [
5479 "test/memory-planner-test.cc",
5480 ],
5481 deps = [
5482 ":XNNPACK",
5483 ":memory_planner",
5484 ],
5485)
5486
Marat Dukhan08c4a432019-10-03 09:29:21 -07005487############################# Build configurations #############################
5488
Marat Dukhanb8642352019-10-30 15:43:02 -07005489# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005490config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005491 name = "xnn_enable_assembly_explicit_true",
5492 define_values = {"xnn_enable_assembly": "true"},
5493)
5494
5495# Disables usage of assembly kernels.
5496config_setting(
5497 name = "xnn_enable_assembly_explicit_false",
5498 define_values = {"xnn_enable_assembly": "false"},
5499)
5500
Marat Dukhan9de90e02020-06-18 16:04:12 -07005501# Enables usage of sparse inference.
5502config_setting(
5503 name = "xnn_enable_sparse_explicit_true",
5504 define_values = {"xnn_enable_sparse": "true"},
5505)
5506
5507# Disables usage of sparse inference.
5508config_setting(
5509 name = "xnn_enable_sparse_explicit_false",
5510 define_values = {"xnn_enable_sparse": "false"},
5511)
5512
Marat Dukhan05702cf2020-03-26 15:41:33 -07005513# Disables usage of HMP-aware optimizations.
5514config_setting(
5515 name = "xnn_enable_hmp_explicit_false",
5516 define_values = {"xnn_enable_hmp": "false"},
5517)
5518
Chao Mei6ddfc602020-05-13 22:29:36 -07005519# Enable usage of optimized memory allocation
5520config_setting(
5521 name = "xnn_enable_memopt_explicit_true",
5522 define_values = {"xnnpack_enable_memopt": "true"},
5523)
5524
5525# Disable usage of optimized memory allocation
5526config_setting(
5527 name = "xnn_enable_memopt_explicit_false",
5528 define_values = {"xnnpack_enable_memopt": "false"},
5529)
5530
Marat Dukhanb8642352019-10-30 15:43:02 -07005531# Builds with -c dbg
5532config_setting(
5533 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005534 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005535 "compilation_mode": "dbg",
5536 },
5537)
5538
5539# Builds with -c opt
5540config_setting(
5541 name = "optimized_build",
5542 values = {
5543 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005544 },
5545)
5546
5547config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005548 name = "linux_k8",
5549 values = {"cpu": "k8"},
5550)
5551
5552config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005553 name = "linux_arm",
5554 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005555)
5556
5557config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005558 name = "linux_armeabi",
5559 values = {"cpu": "armeabi"},
5560)
5561
5562config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005563 name = "linux_armhf",
5564 values = {"cpu": "armhf"},
5565)
5566
5567config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005568 name = "linux_armv7a",
5569 values = {"cpu": "armv7a"},
5570)
5571
5572config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005573 name = "linux_aarch64",
5574 values = {"cpu": "aarch64"},
5575)
5576
5577config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005578 name = "android",
5579 values = {"crosstool_top": "//external:android/crosstool"},
5580)
5581
5582config_setting(
5583 name = "android_armv7",
5584 values = {
5585 "crosstool_top": "//external:android/crosstool",
5586 "cpu": "armeabi-v7a",
5587 },
5588)
5589
5590config_setting(
5591 name = "android_arm64",
5592 values = {
5593 "crosstool_top": "//external:android/crosstool",
5594 "cpu": "arm64-v8a",
5595 },
5596)
5597
5598config_setting(
5599 name = "android_x86",
5600 values = {
5601 "crosstool_top": "//external:android/crosstool",
5602 "cpu": "x86",
5603 },
5604)
5605
5606config_setting(
5607 name = "android_x86_64",
5608 values = {
5609 "crosstool_top": "//external:android/crosstool",
5610 "cpu": "x86_64",
5611 },
5612)
5613
5614config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005615 name = "windows_x86_64",
5616 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005617)
5618
5619config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005620 name = "windows_x86_64_clang",
5621 values = {
5622 "compiler": "clang-cl",
5623 "cpu": "x64_windows",
5624 },
5625)
5626
5627config_setting(
5628 name = "windows_x86_64_mingw",
5629 values = {
5630 "compiler": "mingw-gcc",
5631 "cpu": "x64_windows",
5632 },
5633)
5634
5635config_setting(
5636 name = "windows_x86_64_msys",
5637 values = {
5638 "compiler": "msys-gcc",
5639 "cpu": "x64_windows",
5640 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005641)
5642
5643config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005644 name = "macos_x86_64",
5645 values = {
5646 "apple_platform_type": "macos",
5647 "cpu": "darwin",
5648 },
5649)
5650
5651config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005652 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005653 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005654)
5655
5656config_setting(
5657 name = "emscripten_wasm",
5658 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005659 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005660 "cpu": "wasm",
5661 },
5662)
5663
5664config_setting(
5665 name = "emscripten_wasmsimd",
5666 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005667 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005668 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005669 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005670 },
5671)
5672
5673config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005674 name = "ios_armv7",
5675 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005676 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005677 "cpu": "ios_armv7",
5678 },
5679)
5680
5681config_setting(
5682 name = "ios_arm64",
5683 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005684 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005685 "cpu": "ios_arm64",
5686 },
5687)
5688
5689config_setting(
5690 name = "ios_arm64e",
5691 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005692 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005693 "cpu": "ios_arm64e",
5694 },
5695)
5696
5697config_setting(
5698 name = "ios_x86",
5699 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005700 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005701 "cpu": "ios_i386",
5702 },
5703)
5704
5705config_setting(
5706 name = "ios_x86_64",
5707 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005708 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005709 "cpu": "ios_x86_64",
5710 },
5711)
5712
5713config_setting(
5714 name = "watchos_armv7k",
5715 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005716 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005717 "cpu": "watchos_armv7k",
5718 },
5719)
5720
5721config_setting(
5722 name = "watchos_arm64_32",
5723 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005724 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005725 "cpu": "watchos_arm64_32",
5726 },
5727)
5728
5729config_setting(
5730 name = "watchos_x86",
5731 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005732 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005733 "cpu": "watchos_i386",
5734 },
5735)
5736
5737config_setting(
5738 name = "watchos_x86_64",
5739 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005740 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005741 "cpu": "watchos_x86_64",
5742 },
5743)
5744
5745config_setting(
5746 name = "tvos_arm64",
5747 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005748 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005749 "cpu": "tvos_arm64",
5750 },
5751)
5752
5753config_setting(
5754 name = "tvos_x86_64",
5755 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08005756 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08005757 "cpu": "tvos_x86_64",
5758 },
5759)