blob: dd9611c009f7d841c4ec4978bfce25feaa7ddd48 [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",
Marat Dukhan1f256fc2020-09-24 21:27:14 -0700113 "src/tables/exp2minus-k-over-64.c",
114 "src/tables/exp2minus-k-over-2048.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800115]
116
Marat Dukhan08c4a432019-10-03 09:29:21 -0700117SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800119 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800120 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700121 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
122 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700123 "src/f32-clamp/gen/scalar-x1.c",
124 "src/f32-clamp/gen/scalar-x2.c",
125 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800126 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700127 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700128 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
129 "src/f32-dwconv-chw/3x3p1-scalar.c",
130 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
131 "src/f32-dwconv-chw/5x5p2-scalar.c",
132 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700133 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700135 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700136 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700137 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700138 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700139 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700140 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700141 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700142 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700143 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700144 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700145 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700146 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700147 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700148 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700149 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700150 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700151 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700152 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700153 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700154 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700155 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700156 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700157 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700158 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
159 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700160 "src/f32-gemm/gen/1x4-scalar.c",
161 "src/f32-gemm/gen/2x4-scalar.c",
162 "src/f32-gemm/gen/4x2-scalar.c",
163 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700164 "src/f32-gemm/gen/1x4-relu-scalar.c",
165 "src/f32-gemm/gen/2x4-relu-scalar.c",
166 "src/f32-gemm/gen/4x2-relu-scalar.c",
167 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700168 "src/f32-gemm/gen/1x4-minmax-scalar.c",
169 "src/f32-gemm/gen/2x4-minmax-scalar.c",
170 "src/f32-gemm/gen/4x2-minmax-scalar.c",
171 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700172 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
173 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
174 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700175 "src/f32-hswish/gen/hswish-scalar-x1.c",
176 "src/f32-hswish/gen/hswish-scalar-x2.c",
177 "src/f32-hswish/gen/hswish-scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700178 "src/f32-ibilinear/gen/scalar-c1.c",
179 "src/f32-ibilinear/gen/scalar-c2.c",
180 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700181 "src/f32-igemm/gen/1x4-scalar.c",
182 "src/f32-igemm/gen/2x4-scalar.c",
183 "src/f32-igemm/gen/4x2-scalar.c",
184 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700185 "src/f32-igemm/gen/1x4-relu-scalar.c",
186 "src/f32-igemm/gen/2x4-relu-scalar.c",
187 "src/f32-igemm/gen/4x2-relu-scalar.c",
188 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700189 "src/f32-igemm/gen/1x4-minmax-scalar.c",
190 "src/f32-igemm/gen/2x4-minmax-scalar.c",
191 "src/f32-igemm/gen/4x2-minmax-scalar.c",
192 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700193 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
194 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
195 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700196 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
197 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
198 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700199 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
200 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
201 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
202 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800203 "src/f32-prelu/gen/scalar-2x1.c",
204 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800205 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
211 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
212 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
213 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
214 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
215 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
216 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700217 "src/f32-rmax/scalar.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700218 "src/f32-relu/gen/scalar-x1.c",
219 "src/f32-relu/gen/scalar-x2.c",
220 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700221 "src/f32-relu/gen/scalar-x8.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800222 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
223 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
224 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
225 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
226 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
227 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
228 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
229 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
230 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700231 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
232 "src/f32-spmm/gen/1x1-minmax-scalar.c",
233 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
234 "src/f32-spmm/gen/2x1-minmax-scalar.c",
235 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
236 "src/f32-spmm/gen/4x1-minmax-scalar.c",
237 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
238 "src/f32-spmm/gen/8x1-minmax-scalar.c",
239 "src/f32-spmm/gen/8x2-minmax-scalar.c",
240 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700241 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
242 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
243 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700244 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700245 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
246 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
247 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700248 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700249 "src/f32-vbinary/gen/vadd-scalar-x1.c",
250 "src/f32-vbinary/gen/vadd-scalar-x2.c",
251 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700252 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700253 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
254 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
255 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700256 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700257 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
258 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
259 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700260 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700261 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
262 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
263 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700264 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700265 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
266 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
267 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700268 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700269 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
270 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
271 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700272 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700273 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
274 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
275 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700276 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700277 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
278 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
279 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700280 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700281 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
282 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
283 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700284 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700285 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
286 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
287 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700288 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800289 "src/f32-vbinary/gen/vmax-scalar-x1.c",
290 "src/f32-vbinary/gen/vmax-scalar-x2.c",
291 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700292 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800293 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
294 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
295 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700296 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800297 "src/f32-vbinary/gen/vmin-scalar-x1.c",
298 "src/f32-vbinary/gen/vmin-scalar-x2.c",
299 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700300 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800301 "src/f32-vbinary/gen/vminc-scalar-x1.c",
302 "src/f32-vbinary/gen/vminc-scalar-x2.c",
303 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700304 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700305 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
306 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
307 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700308 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700309 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
310 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
311 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700312 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700313 "src/f32-vbinary/gen/vmul-scalar-x1.c",
314 "src/f32-vbinary/gen/vmul-scalar-x2.c",
315 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700316 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700317 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
318 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
319 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700320 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700321 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
322 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
323 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700324 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700325 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
326 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
327 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700328 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700329 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
330 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
331 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700332 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700333 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
334 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
335 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700336 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700337 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
338 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
339 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700340 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700341 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
342 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
343 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700344 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700345 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
346 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
347 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700348 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700349 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
350 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
351 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700352 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700353 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
354 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
355 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700356 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700357 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
358 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
359 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700360 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700361 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
362 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
363 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700364 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700365 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
366 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
367 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700368 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700369 "src/f32-vbinary/gen/vsub-scalar-x1.c",
370 "src/f32-vbinary/gen/vsub-scalar-x2.c",
371 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700372 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700373 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
374 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
375 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700376 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700377 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
378 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
379 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700380 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700381 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
382 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
383 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700384 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700385 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
386 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
387 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700388 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
389 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
390 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700391 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
392 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
393 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
394 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
395 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
396 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
397 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
398 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
399 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700400 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
401 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
402 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700403 "src/f32-vunary/gen/vabs-scalar-x1.c",
404 "src/f32-vunary/gen/vabs-scalar-x2.c",
405 "src/f32-vunary/gen/vabs-scalar-x4.c",
406 "src/f32-vunary/gen/vneg-scalar-x1.c",
407 "src/f32-vunary/gen/vneg-scalar-x2.c",
408 "src/f32-vunary/gen/vneg-scalar-x4.c",
409 "src/f32-vunary/gen/vsqr-scalar-x1.c",
410 "src/f32-vunary/gen/vsqr-scalar-x2.c",
411 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800412 "src/math/expminus-scalar-lut2048-p1.c",
413 "src/math/expminus-scalar-lut64-p2.c",
414 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700415 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700416 "src/math/roundne-scalar-nearbyint.c",
417 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700418 "src/math/roundd-scalar-addsub.c",
419 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700420 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700421 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700422 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700423 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700424 "src/math/roundz-scalar-addsub.c",
425 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700426 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700427 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
428 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
429 "src/math/sigmoid-scalar-rr2-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700430 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
431 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700432 "src/qs8-requantization/fp32-scalar-lrintf.c",
433 "src/qs8-requantization/fp32-scalar-magic.c",
434 "src/qs8-requantization/precise-scalar-signed64.c",
435 "src/qs8-requantization/precise-scalar-unsigned32.c",
436 "src/qs8-requantization/precise-scalar-unsigned64.c",
437 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700438 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
439 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
440 "src/qu8-dwconv/up1x9-minmax-scalar.c",
441 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
442 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700443 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
444 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700445 "src/qu8-gemm/2x2-minmax-scalar.c",
446 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700447 "src/qu8-requantization/fp32-scalar-lrintf.c",
448 "src/qu8-requantization/fp32-scalar-magic.c",
449 "src/qu8-requantization/precise-scalar-signed64.c",
450 "src/qu8-requantization/precise-scalar-unsigned32.c",
451 "src/qu8-requantization/precise-scalar-unsigned64.c",
452 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700453 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700454 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700455 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700456 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700457 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700458 "src/x32-fill/scalar-float.c",
459 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700460 "src/x32-packx/x2-scalar.c",
461 "src/x32-packx/x3-scalar.c",
462 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700463 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700464 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700465 "src/x32-unpool/scalar.c",
466 "src/x32-zip/x2-scalar.c",
467 "src/x32-zip/x3-scalar.c",
468 "src/x32-zip/x4-scalar.c",
469 "src/x32-zip/xm-scalar.c",
470 "src/x8-lut/scalar.c",
471 "src/x8-zip/x2-scalar.c",
472 "src/x8-zip/x3-scalar.c",
473 "src/x8-zip/x4-scalar.c",
474 "src/x8-zip/xm-scalar.c",
475]
476
Marat Dukhan436ebe62019-12-04 15:10:12 -0800477WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700478 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
479 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700480 "src/f32-clamp/gen/wasm-x1.c",
481 "src/f32-clamp/gen/wasm-x2.c",
482 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700483 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
484 "src/f32-dwconv/gen/up1x4-wasm.c",
485 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
486 "src/f32-dwconv/gen/up1x9-wasm.c",
487 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
488 "src/f32-dwconv/gen/up1x25-wasm.c",
489 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
490 "src/f32-dwconv/gen/up2x4-wasm.c",
491 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
492 "src/f32-dwconv/gen/up2x9-wasm.c",
493 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
494 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700495 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
496 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
497 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
498 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700499 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
500 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700501 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
502 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
503 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
504 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700505 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
506 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700507 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
508 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700509 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
510 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
511 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700512 "src/f32-gemm/gen/1x4-wasm.c",
513 "src/f32-gemm/gen/2x4-wasm.c",
514 "src/f32-gemm/gen/4x2-wasm.c",
515 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700516 "src/f32-gemm/gen/1x4-relu-wasm.c",
517 "src/f32-gemm/gen/2x4-relu-wasm.c",
518 "src/f32-gemm/gen/4x2-relu-wasm.c",
519 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700520 "src/f32-gemm/gen/1x4-minmax-wasm.c",
521 "src/f32-gemm/gen/2x4-minmax-wasm.c",
522 "src/f32-gemm/gen/4x2-minmax-wasm.c",
523 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700524 "src/f32-hswish/gen/hswish-wasm-x1.c",
525 "src/f32-hswish/gen/hswish-wasm-x2.c",
526 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700527 "src/f32-igemm/gen/1x4-wasm.c",
528 "src/f32-igemm/gen/2x4-wasm.c",
529 "src/f32-igemm/gen/4x2-wasm.c",
530 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700531 "src/f32-igemm/gen/1x4-relu-wasm.c",
532 "src/f32-igemm/gen/2x4-relu-wasm.c",
533 "src/f32-igemm/gen/4x2-relu-wasm.c",
534 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700535 "src/f32-igemm/gen/1x4-minmax-wasm.c",
536 "src/f32-igemm/gen/2x4-minmax-wasm.c",
537 "src/f32-igemm/gen/4x2-minmax-wasm.c",
538 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700539 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
540 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
541 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700542 "src/f32-prelu/gen/wasm-2x1.c",
543 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700544 "src/f32-relu/gen/wasm-x1.c",
545 "src/f32-relu/gen/wasm-x2.c",
546 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700547 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700548 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
549 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
550 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700551 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700552 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
553 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
554 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700555 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700556 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
557 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
558 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700559 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700560 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
561 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
562 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700563 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700564 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
565 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
566 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700567 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700568 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
569 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
570 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700571 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700572 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
573 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
574 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700575 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700576 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
577 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
578 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700579 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800580 "src/f32-vbinary/gen/vmax-wasm-x1.c",
581 "src/f32-vbinary/gen/vmax-wasm-x2.c",
582 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700583 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800584 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
585 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
586 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700587 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800588 "src/f32-vbinary/gen/vmin-wasm-x1.c",
589 "src/f32-vbinary/gen/vmin-wasm-x2.c",
590 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700591 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800592 "src/f32-vbinary/gen/vminc-wasm-x1.c",
593 "src/f32-vbinary/gen/vminc-wasm-x2.c",
594 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700595 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700596 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
597 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
598 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700599 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700600 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
601 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
602 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700603 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700604 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
605 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
606 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700607 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700608 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
609 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
610 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700611 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700612 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
613 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
614 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700615 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700616 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
617 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
618 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700619 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700620 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
621 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
622 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700623 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700624 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
625 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
626 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700627 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700628 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
629 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
630 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700631 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700632 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
633 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
634 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700635 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700636 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
637 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
638 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700639 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700640 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
641 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
642 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700643 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700644 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
645 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
646 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700647 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
648 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
649 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800650]
651
Marat Dukhan290055c2020-06-09 12:24:29 -0700652WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700653 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
654 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
655 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700656 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
657 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
658 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
659 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700660 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
661 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
662 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
663 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700664 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
665 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
666 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
667 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
668 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
669 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700670 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
671 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
672 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
673 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
674 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
675 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
676 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
677 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
678 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
679 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
680 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
681 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
682 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
683 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
684 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
685 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
686 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
687 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
688 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
689 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
690 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
691 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
692 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
693 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700694 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
695 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
696 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700697 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700698 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
699 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
700 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
701 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700702 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
703 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
704 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
705 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
706 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
707 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
708 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
709 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
710 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
711 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
712 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
713 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
714 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
715 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
716 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
717 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
718 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
719 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
720 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
721 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
722 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
723 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
724 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
725 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
726 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
727 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
728 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
729 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
730 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
731 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700732 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
733 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700734 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
735 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
736 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
737 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
738 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
739 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
740 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
741 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
742 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
743 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
744 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
745 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
746 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
747 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
748 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
749 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
750 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
751 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
752 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
753 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
754 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
755 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
756 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
757 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
758 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
759 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
760 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
761 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
762 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
763 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700764 "src/f32-ibilinear/gen/wasmsimd-c4.c",
765 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700766 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
767 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
768 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700769 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700770 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
771 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
772 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
773 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700774 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
775 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
776 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
777 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
778 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
779 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
780 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
781 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
782 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
783 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
784 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
785 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
786 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
787 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
788 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
789 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
790 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
791 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
792 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
793 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
794 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
795 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
796 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
797 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
798 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
799 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
800 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
801 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
802 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
803 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700804 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
805 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700806 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
807 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
808 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
809 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700810 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
811 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700812 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
813 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
814 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700815 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
816 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700817 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
818 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
819 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
820 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700821 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
822 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700823 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
824 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
825 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700826 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
827 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700828 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
829 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
830 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
831 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
832 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
833 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
834 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700835 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
836 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700837 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
838 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
839 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
840 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700841 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
842 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
843 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
844 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
845 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
846 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
847 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
848 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
849 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
850 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
851 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
852 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700853 "src/f32-relu/gen/wasmsimd-x4.c",
854 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700855 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700856 "src/f32-rmax/wasmsimd-arm.c",
857 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700858 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
859 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
860 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
861 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
862 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
863 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700864 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
865 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
866 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
867 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
868 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
869 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700870 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
871 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700872 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700873 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
874 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700875 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700876 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
877 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700878 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700879 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
880 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700881 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700882 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
883 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700884 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700885 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
886 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700887 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700888 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
889 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700890 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700891 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
892 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700893 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700894 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
895 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700896 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700897 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
898 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700899 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700900 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
901 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700902 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700903 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
904 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700905 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700906 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
907 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700908 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700909 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
910 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700911 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700912 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
913 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700914 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700915 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
916 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700917 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700918 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
919 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700920 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700921 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
922 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700923 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700924 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
925 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700926 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700927 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
928 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700929 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700930 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
931 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700932 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700933 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
934 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700935 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700936 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
937 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700938 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700939 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
940 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700941 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700942 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
943 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700944 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700945 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
946 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700947 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700948 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
949 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700950 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700951 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
952 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700953 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700954 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
955 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700956 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700957 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
958 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700959 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700960 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
961 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700962 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700963 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
964 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700965 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700966 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
967 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700968 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700969 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
970 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700971 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700972 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
973 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700974 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700975 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
976 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700977 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700978 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
979 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700980 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700981 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
982 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700983 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700984 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
985 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700986 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700987 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
988 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700989 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700990 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
991 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700992 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700993 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
994 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700995 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700996 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
997 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700998 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700999 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1000 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001001 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001002 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1003 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001004 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001005 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1006 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001007 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001008 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1009 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001010 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001011 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1012 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001013 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001014 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1015 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001016 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001017 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1018 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001019 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001020 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1021 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1022 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1023 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001024 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
1025 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
1026 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
1027 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001028 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1029 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1030 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1031 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1032 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1033 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
1034 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1035 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1036 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1037 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1038 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1039 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1040 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1041 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001042 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1043 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001044 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1045 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1046 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1047 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1048 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1049 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001050 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1051 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1052 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanef451802020-08-06 11:53:47 -07001053 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1054 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1055 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001056 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1057 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1058 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001059 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
1060 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1061 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1062 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
1063 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
1064 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
1065 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
1066 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
1067 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1068 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
1069 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1070 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1071 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
1072 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
1073 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001074 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001075 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001076 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1077 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1078 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1079 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1080 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1081 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1082 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1083 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001084 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001085 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001086 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001087 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001088 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001089 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001090 "src/x32-zip/x2-wasmsimd.c",
1091 "src/x32-zip/x3-wasmsimd.c",
1092 "src/x32-zip/x4-wasmsimd.c",
1093 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001094 "src/math/roundne-wasmsimd-addsub.c",
1095 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001096 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001097 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001098 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001099 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001100 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -07001101 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1102 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001103]
1104
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001105PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -07001106 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -07001107 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -07001108 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -07001109 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -07001110 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001111 "src/f32-spmm/gen/4x1-minmax-psimd.c",
1112 "src/f32-spmm/gen/8x1-minmax-psimd.c",
1113 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001114]
1115
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001116PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001117 "src/qs8-requantization/precise-psimd.c",
1118 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001119 "src/qu8-requantization/precise-psimd.c",
1120 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001121]
1122
Marat Dukhan08c4a432019-10-03 09:29:21 -07001123# ISA-specific micro-kernels
1124NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001125 "src/f32-argmaxpool/4x-neon-c4.c",
1126 "src/f32-argmaxpool/9p8x-neon-c4.c",
1127 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001128 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1129 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001130 "src/f32-clamp/gen/neon-x4.c",
1131 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001132 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
1133 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001134 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
1135 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001136 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
1137 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001138 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
1139 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001140 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
1141 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
1142 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1143 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001144 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
1145 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
1146 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
1147 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001148 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1149 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
1150 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
1151 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001152 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001153 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1154 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001155 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1156 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1157 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1158 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1159 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1160 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1161 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1162 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1163 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1164 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1165 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1166 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1167 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1168 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1169 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1170 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
1171 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1172 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1173 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1174 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
1175 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1176 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
1177 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
1178 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1179 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
1180 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1181 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
1182 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
1183 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1184 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1185 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001186 "src/f32-hswish/gen/hswish-neon-x4.c",
1187 "src/f32-hswish/gen/hswish-neon-x8.c",
1188 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001189 "src/f32-ibilinear/gen/neon-c4.c",
1190 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001191 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
1192 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1193 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
1194 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1195 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1196 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1197 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
1198 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
1199 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1200 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
1201 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1202 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1203 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1204 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1205 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1206 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001207 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1208 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1209 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001210 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1211 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001212 "src/f32-prelu/gen/neon-1x4.c",
1213 "src/f32-prelu/gen/neon-1x8.c",
1214 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001215 "src/f32-prelu/gen/neon-2x4.c",
1216 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001217 "src/f32-prelu/gen/neon-2x16.c",
1218 "src/f32-prelu/gen/neon-4x4.c",
1219 "src/f32-prelu/gen/neon-4x8.c",
1220 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001221 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1222 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1223 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1224 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1225 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1226 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1227 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1228 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1229 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1230 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1231 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1232 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1233 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1234 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1235 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1236 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1237 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1238 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1239 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1240 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1241 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1242 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1243 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1244 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001245 "src/f32-relu/gen/neon-x4.c",
1246 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001247 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001248 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001249 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1250 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1251 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1252 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1253 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1254 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1255 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1256 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1257 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1258 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1259 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1260 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1261 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1262 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1263 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1264 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1265 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1266 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001267 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1268 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1269 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1270 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001271 "src/f32-vbinary/gen/vmax-neon-x4.c",
1272 "src/f32-vbinary/gen/vmax-neon-x8.c",
1273 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1274 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1275 "src/f32-vbinary/gen/vmin-neon-x4.c",
1276 "src/f32-vbinary/gen/vmin-neon-x8.c",
1277 "src/f32-vbinary/gen/vminc-neon-x4.c",
1278 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001279 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1280 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1281 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1282 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1283 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1284 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001285 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1286 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1287 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1288 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001289 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1290 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1291 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1292 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001293 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1294 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001295 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1296 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001297 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1298 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1299 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1300 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1301 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1302 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1303 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1304 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001305 "src/f32-vunary/gen/vabs-neon-x4.c",
1306 "src/f32-vunary/gen/vabs-neon-x8.c",
1307 "src/f32-vunary/gen/vneg-neon-x4.c",
1308 "src/f32-vunary/gen/vneg-neon-x8.c",
1309 "src/f32-vunary/gen/vsqr-neon-x4.c",
1310 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001311 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1312 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1313 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1314 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001315 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1316 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1317 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1318 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
1319 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1320 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1321 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1322 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001323 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
1324 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
1325 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1326 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1327 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
1328 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1329 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1330 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001331 "src/qs8-requantization/precise-neon.c",
1332 "src/qs8-requantization/fp32-neon.c",
1333 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001334 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1335 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1336 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1337 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1338 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1339 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1340 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1341 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001342 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1343 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1344 "src/qu8-dwconv/up8x9-minmax-neon.c",
1345 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1346 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1347 "src/qu8-gemm/4x8-minmax-neon.c",
1348 "src/qu8-gemm/8x8-minmax-neon.c",
1349 "src/qu8-igemm/4x8-minmax-neon.c",
1350 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001351 "src/qu8-requantization/precise-neon.c",
1352 "src/qu8-requantization/fp32-neon.c",
1353 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001354 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001355 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001356 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001357 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001358 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001359 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001360 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001361 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001362 "src/x32-zip/x2-neon.c",
1363 "src/x32-zip/x3-neon.c",
1364 "src/x32-zip/x4-neon.c",
1365 "src/x32-zip/xm-neon.c",
1366 "src/x8-zip/x2-neon.c",
1367 "src/x8-zip/x3-neon.c",
1368 "src/x8-zip/x4-neon.c",
1369 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001370 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001371 "src/math/roundd-neon-addsub.c",
1372 "src/math/roundd-neon-cvt.c",
1373 "src/math/roundu-neon-addsub.c",
1374 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001375 "src/math/roundz-neon-addsub.c",
1376 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001377 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001378 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1379 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1380 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1381 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1382 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1383 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001384 "src/math/sqrt-neon-nr1rsqrts.c",
1385 "src/math/sqrt-neon-nr2rsqrts.c",
1386 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001387]
1388
1389NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001390 "src/f32-ibilinear/gen/neonfma-c4.c",
1391 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001392 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1393 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1394 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1395 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1396 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1397 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1398 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1399 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1400 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001401 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1402 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1403 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1404 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001405 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1406 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1407 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1408 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001409 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1410 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1411 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1412 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001413 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1414 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1415 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1416 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1417 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1418 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1419 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1420 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1421 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1422 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1423 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1424 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1425 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1426 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1427 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1428 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1429 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1430 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001431 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1432 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001433 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1434 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1435 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1436 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1437 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1438 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1439 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1440 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1441 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1442 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1443 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1444 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1445 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1446 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1447 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1448 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1449 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1450 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1451 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1452 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1453 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1454 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1455 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1456 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001457 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1458 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1459 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1460 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1461 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1462 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1463 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1464 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1465 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1466 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1467 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1468 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1469 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1470 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1471 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1472 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1473 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1474 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1475 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1476 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1477 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1478 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1479 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1480 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1481 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1482 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1483 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1484 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1485 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1486 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1487 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1488 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1489 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1490 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1491 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1492 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1493 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1494 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1495 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1496 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1497 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1498 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1499 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1500 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1501 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1502 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1503 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1504 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1505 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1506 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1507 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1508 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1509 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1510 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001511 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1512 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001513 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1514 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1515 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1516 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1517 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1518 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1519 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1520 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1521 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1522 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1523 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1524 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1525 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1526 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1527 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1528 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1529 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1530 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1531 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1532 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001533 "src/math/exp-neonfma-lut64-p2.c",
1534 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001535 "src/math/expminus-neonfma-lut2048-p1.c",
1536 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001537 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001538 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1539 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1540 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1541 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1542 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1543 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1544 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1545 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1546 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1547 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1548 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1549 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1550 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1551 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1552 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1553 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1554 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1555 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001556 "src/math/sqrt-neonfma-nr1fma.c",
1557 "src/math/sqrt-neonfma-nr2fma.c",
1558 "src/math/sqrt-neonfma-nr3fma.c",
1559 "src/math/sqrt-neonfma-nr2fma1adj.c",
1560 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001561]
1562
1563AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001564 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1565 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1566 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1567 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1568 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1569 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001570 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1571 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1572 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1573 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1574 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1575 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1576 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1577 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1578 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1579 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1580 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1581 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1582 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1583 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1584 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1585 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1586 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1587 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1588 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1589 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001590 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1591 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001592 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1593 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001594 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1595 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001596 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1597 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001598 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1599 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1600 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1601 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1602 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001603 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1604 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1605 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1606 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1607 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1608 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1609 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1610 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1611 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1612 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1613 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1614 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1615 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1616 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1617 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1618 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1619 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1620 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001621 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1622 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1623 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1624 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1625 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1626 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1627 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1628 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1629 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1630 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1631 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1632 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1633 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1634 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1635 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1636 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1637 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1638 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001639 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1640 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001641 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1642 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1643 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1644 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1645 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1646 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001647]
1648
Marat Dukhan8853b822020-05-07 12:19:01 -07001649NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001650 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1651 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1652 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1653 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1654 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1655 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1656 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1657 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001658 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001659 "src/math/roundd-neonv8.c",
1660 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001661 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001662]
1663
Marat Dukhan08c4a432019-10-03 09:29:21 -07001664AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001665 "src/f16-clamp/gen/neonfp16arith-x16.c",
1666 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001667 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1668 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1669 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1670 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1671 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1672 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1673 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1674 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1675 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1676 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1677 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1678 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001679 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1680 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001681 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1682 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001683 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001684 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001685 "src/f16-relu/gen/neonfp16arith-x16.c",
1686 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001687 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1688 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1689 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1690 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1691 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1692 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1693 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1694 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1695 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1696 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1697 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1698 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1699 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1700 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1701 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1702 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1703 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1704 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1705 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1706 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1707 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1708 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1709 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1710 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1711 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1712 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1713 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1714 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001715 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1716 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001717 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1718 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1719 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1720 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1721 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1722 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1723 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1724 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001725 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1726 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1727 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1728 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001729 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1730 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1731 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1732 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1733 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1734 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1735 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1736 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1737 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1738 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1739 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1740 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001741 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1742 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1743 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1744 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1745 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1746 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1747 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1748 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001749]
1750
Benoit Jacoba9644732020-08-13 12:48:55 -07001751NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07001752 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001753 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
1754 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001755 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
1756 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001757 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001758 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001759 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001760 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001761 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001762 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
1763 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001764 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
1765 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001766 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001767 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001768 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001769 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001770]
1771
Marat Dukhan08c4a432019-10-03 09:29:21 -07001772SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001773 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1774 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001775 "src/f32-clamp/gen/sse-x4.c",
1776 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001777 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1778 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001779 "src/f32-dwconv-chw/3x3p1-sse.c",
1780 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001781 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1782 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1783 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1784 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1785 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1786 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1787 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1788 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1789 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1790 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1791 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1792 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001793 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001794 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1795 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001796 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1797 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1798 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1799 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1800 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1801 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1802 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1803 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1804 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1805 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1806 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1807 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1808 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001809 "src/f32-hswish/gen/hswish-sse-x4.c",
1810 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001811 "src/f32-ibilinear/gen/sse-c4.c",
1812 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001813 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1814 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1815 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1816 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1817 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1818 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1819 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001820 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1821 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1822 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001823 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001824 "src/f32-prelu/gen/sse-2x4.c",
1825 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001826 "src/f32-relu/gen/sse-x4.c",
1827 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001828 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001829 "src/f32-spmm/gen/4x1-minmax-sse.c",
1830 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001831 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001832 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1833 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1834 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1835 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1836 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1837 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1838 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1839 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001840 "src/f32-vbinary/gen/vmax-sse-x4.c",
1841 "src/f32-vbinary/gen/vmax-sse-x8.c",
1842 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1843 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1844 "src/f32-vbinary/gen/vmin-sse-x4.c",
1845 "src/f32-vbinary/gen/vmin-sse-x8.c",
1846 "src/f32-vbinary/gen/vminc-sse-x4.c",
1847 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001848 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1849 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1850 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1851 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1852 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1853 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1854 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1855 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001856 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1857 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1858 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1859 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001860 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1861 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1862 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1863 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001864 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1865 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001866 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1867 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001868 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1869 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001870 "src/f32-vunary/gen/vabs-sse-x4.c",
1871 "src/f32-vunary/gen/vabs-sse-x8.c",
1872 "src/f32-vunary/gen/vneg-sse-x4.c",
1873 "src/f32-vunary/gen/vneg-sse-x8.c",
1874 "src/f32-vunary/gen/vsqr-sse-x4.c",
1875 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001876 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001877 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001878 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001879 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001880 "src/math/roundd-sse-addsub.c",
1881 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001882 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001883 "src/math/sqrt-sse-hh1mac.c",
1884 "src/math/sqrt-sse-nr1mac.c",
1885 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001886]
1887
1888SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001889 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1890 "src/f32-argmaxpool/4x-sse2-c4.c",
1891 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001892 "src/f32-prelu/gen/sse2-2x4.c",
1893 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001894 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1895 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1896 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1897 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1898 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1899 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1900 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1901 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1902 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1903 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1904 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1905 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07001906 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
1907 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
1908 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
1909 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
1910 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
1911 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001912 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001913 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001914 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001915 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001916 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1917 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001918 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1919 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001920 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1921 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1922 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1923 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1924 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1925 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1926 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1927 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001928 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1929 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1930 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001931 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1932 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1933 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001934 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1935 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1936 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001937 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1938 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001939 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1940 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001941 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1942 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001943 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1944 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001945 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001946 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1947 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001948 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001949 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1950 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1951 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001952 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1953 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001954 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1955 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001956 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1957 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001958 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001959 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1960 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001961 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001962 "src/qs8-requantization/precise-sse2.c",
1963 "src/qs8-requantization/fp32-sse2.c",
1964 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07001965 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
1966 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
1967 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
1968 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07001969 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
1970 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
1971 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
1972 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001973 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1974 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1975 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1976 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1977 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1978 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1979 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1980 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001981 "src/qu8-requantization/precise-sse2.c",
1982 "src/qu8-requantization/fp32-sse2.c",
1983 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001984 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001985 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001986 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001987 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001988 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001989 "src/x32-zip/x2-sse2.c",
1990 "src/x32-zip/x3-sse2.c",
1991 "src/x32-zip/x4-sse2.c",
1992 "src/x32-zip/xm-sse2.c",
1993 "src/x8-zip/x2-sse2.c",
1994 "src/x8-zip/x3-sse2.c",
1995 "src/x8-zip/x4-sse2.c",
1996 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001997 "src/math/exp-sse2-p5.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07001998 "src/math/exp-sse2-lut64-p2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001999 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07002000 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002001 "src/math/roundd-sse2-cvt.c",
2002 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002003 "src/math/roundz-sse2-cvt.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -07002004 "src/math/sigmoid-sse2-rr2-p5-div.c",
2005 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002006]
2007
2008SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002009 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2010 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2011 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07002012 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2013 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2014 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002015 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2016 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2017 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002018 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
2019 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002020 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
2021 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002022 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
2023 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002024 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
2025 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002026 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002027 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
2028 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002029 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002030 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
2031 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
2032 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002033 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
2034 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002035 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
2036 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002037 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
2038 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002039 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002040 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
2041 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002042 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002043 "src/qs8-requantization/precise-ssse3.c",
2044 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002045 "src/qu8-requantization/precise-ssse3.c",
2046 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002047]
2048
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002049SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002050 "src/f32-prelu/gen/sse41-2x4.c",
2051 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002052 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2053 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2054 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2055 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2056 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2057 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002058 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002059 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002060 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002061 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002062 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2063 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002064 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2065 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002066 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2067 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2068 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2069 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2070 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2071 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2072 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2073 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002074 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2075 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2076 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07002077 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2078 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2079 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002080 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2081 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2082 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002083 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
2084 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002085 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
2086 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002087 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
2088 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002089 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
2090 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002091 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002092 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
2093 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002094 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002095 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
2096 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
2097 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002098 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
2099 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002100 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
2101 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002102 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
2103 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002104 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002105 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
2106 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002107 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002108 "src/qs8-requantization/fp32-sse4.c",
2109 "src/qs8-requantization/precise-sse4.c",
2110 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002111 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2112 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2113 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2114 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002115 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2116 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2117 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2118 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002119 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2120 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2121 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2122 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002123 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2124 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2125 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2126 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002127 "src/qu8-requantization/precise-sse4.c",
2128 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002129 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002130 "src/math/roundd-sse41.c",
2131 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002132 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002133]
2134
Marat Dukhan08c4a432019-10-03 09:29:21 -07002135AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002136 "src/f32-clamp/gen/avx-x8.c",
2137 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002138 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2139 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2140 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2141 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
2142 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2143 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2144 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2145 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
2146 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2147 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
2148 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2149 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
2150 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2151 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2152 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2153 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2154 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
2155 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2156 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2157 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2158 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2159 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
2160 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
2161 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
2162 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2163 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2164 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2165 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
2166 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
2167 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002168 "src/f32-hswish/gen/hswish-avx-x8.c",
2169 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002170 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
2171 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
2172 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
2173 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2174 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
2175 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2176 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
2177 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
2178 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002179 "src/f32-prelu/gen/avx-2x8.c",
2180 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002181 "src/f32-relu/gen/avx-x16.c",
2182 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002183 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002184 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2185 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2186 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2187 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2188 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2189 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2190 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2191 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002192 "src/f32-vbinary/gen/vmax-avx-x8.c",
2193 "src/f32-vbinary/gen/vmax-avx-x16.c",
2194 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2195 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2196 "src/f32-vbinary/gen/vmin-avx-x8.c",
2197 "src/f32-vbinary/gen/vmin-avx-x16.c",
2198 "src/f32-vbinary/gen/vminc-avx-x8.c",
2199 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002200 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2201 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2202 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2203 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2204 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2205 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2206 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2207 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002208 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2209 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2210 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2211 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002212 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2213 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2214 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2215 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002216 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2217 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002218 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002219 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2220 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2221 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2222 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2223 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2224 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2225 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2226 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002227 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2228 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002229 "src/f32-vunary/gen/vabs-avx-x8.c",
2230 "src/f32-vunary/gen/vabs-avx-x16.c",
2231 "src/f32-vunary/gen/vneg-avx-x8.c",
2232 "src/f32-vunary/gen/vneg-avx-x16.c",
2233 "src/f32-vunary/gen/vsqr-avx-x8.c",
2234 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhand08abcb2020-09-21 00:26:25 -07002235 "src/math/exp-avx-p5.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002236 "src/math/sigmoid-avx-rr2-p5-div.c",
2237 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002238]
2239
Marat Dukhan1566fee2020-08-02 21:55:41 -07002240XOP_UKERNELS = [
2241 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
2242 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2243 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
2244 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002245 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
2246 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002247 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
2248 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002249 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002250 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
2251 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002252 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002253 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
2254 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
2255 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002256 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
2257 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2258 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
2259 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
2260 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
2261 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002262 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002263 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
2264 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002265 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002266 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2267 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2268 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2269 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2270 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2271 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2272 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2273 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002274]
2275
Marat Dukhanfda12b82019-11-21 12:27:59 -08002276FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002277 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2278 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2279 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2280 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
2281 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2282 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2283 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2284 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
2285 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2286 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
2287 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2288 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
2289 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2290 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2291 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2292 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2293 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2294 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
2295 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2296 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2297 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2298 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2299 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2300 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2301 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2302 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2303 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
2304 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
2305 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
2306 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2307 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2308 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
2309 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2310 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2311 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2312 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2313 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2314 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
2315 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
2316 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002317 "src/f32-hswish/gen/hswish-fma3-x8.c",
2318 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002319 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
2320 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2321 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2322 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2323 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2324 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2325 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2326 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2327 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2328 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2329 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2330 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2331 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2332 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002333 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2334 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2335 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2336 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2337 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2338 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2339 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2340 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002341 "src/math/sqrt-fma3-nr1fma.c",
2342 "src/math/sqrt-fma3-nr2fma.c",
2343 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002344]
2345
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002346AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002347 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2348 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2349 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2350 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2351 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2352 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2353 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2354 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2355 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2356 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2357 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2358 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2359 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2360 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2361 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2362 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2363 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2364 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2365 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2366 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2367 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2368 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2369 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2370 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2371 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2372 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2373 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2374 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2375 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2376 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2377 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2378 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2379 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2380 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2381 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2382 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002383 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2384 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2385 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2386 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2387 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2388 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2389 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2390 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2391 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2392 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2393 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2394 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2395 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2396 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2397 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2398 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2399 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2400 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2401 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2402 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2403 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2404 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2405 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2406 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2407 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2408 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2409 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2410 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2411 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2412 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002413 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2414 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2415 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2416 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2417 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2418 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2419 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2420 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2421 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2422 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2423 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2424 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2425 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2426 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2427 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2428 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2429 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2430 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2431 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2432 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2433 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2434 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2435 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2436 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan40135522020-08-07 01:21:00 -07002437 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
2438 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002439 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2440 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2441 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2442 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002443 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2444 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2445 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002446 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2447 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2448 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002449 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2450 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2451 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002452 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2453 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2454 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2455 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2456 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2457 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2458 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2459 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002460 "src/math/exp-avx2-p5.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07002461 "src/math/exp-avx2-lut8-p3-perm.c",
2462 "src/math/exp-avx2-lut8-p4-perm.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002463 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002464 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002465 "src/math/sigmoid-avx2-rr1-p5-div.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002466 "src/math/sigmoid-avx2-rr2-p5-div.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002467 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002468 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002469 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002470 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002471]
2472
Marat Dukhan08c4a432019-10-03 09:29:21 -07002473AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002474 "src/f32-clamp/gen/avx512f-x16.c",
2475 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002476 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2477 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2478 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2479 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2480 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2481 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2482 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2483 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2484 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2485 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2486 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2487 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2488 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2489 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2490 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2491 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2492 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2493 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2494 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2495 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2496 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2497 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2498 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2499 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002500 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2501 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002502 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2503 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2504 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2505 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2506 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2507 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002508 "src/f32-prelu/gen/avx512f-2x16.c",
2509 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002510 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2511 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2512 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2513 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2514 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2515 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2516 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2517 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2518 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2519 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2520 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2521 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2522 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2523 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2524 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2525 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2526 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2527 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2528 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2529 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2530 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2531 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2532 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2533 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2534 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2535 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2536 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2537 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2538 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2539 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2540 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2541 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2542 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2543 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2544 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2545 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002546 "src/f32-relu/gen/avx512f-x16.c",
2547 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002548 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002549 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2550 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2551 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2552 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2553 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2554 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2555 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2556 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2557 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2558 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2559 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2560 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2561 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2562 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2563 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2564 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
2565 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2566 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2567 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2568 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2569 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2570 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2571 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2572 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2573 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2574 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2575 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2576 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2577 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2578 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2579 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2580 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
2581 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2582 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2583 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2584 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2585 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2586 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2587 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2588 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2589 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2590 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2591 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2592 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2593 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2594 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2595 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2596 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002597 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2598 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2599 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2600 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2601 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2602 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2603 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2604 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002605 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2606 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2607 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2608 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2609 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2610 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2611 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2612 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002613 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2614 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2615 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2616 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2617 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2618 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2619 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2620 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002621 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2622 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2623 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2624 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002625 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2626 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2627 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2628 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002629 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2630 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002631 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002632 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2633 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2634 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2635 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2636 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2637 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2638 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2639 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2640 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2641 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2642 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2643 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2644 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2645 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2646 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2647 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2648 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2649 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2650 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2651 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2652 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2653 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2654 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2655 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002656 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2657 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2658 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2659 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2660 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2661 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2662 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2663 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002664 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2665 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2666 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2667 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2668 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2669 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2670 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2671 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002672 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2673 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2674 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2675 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2676 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2677 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002678 "src/math/exp-avx512f-p5-scalef.c",
2679 "src/math/exp-avx512f-p5.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07002680 "src/math/exp-avx512f-lut16-p3-perm.c",
Marat Dukhan80a8ac52020-09-21 17:00:27 -07002681 "src/math/exp-avx512f-lut16-p3-perm-scalef.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07002682 "src/math/exp-avx512f-lut32-p2-perm2.c",
Marat Dukhan80a8ac52020-09-21 17:00:27 -07002683 "src/math/exp-avx512f-lut32-p2-perm2-scalef.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002684 "src/math/extexp-avx512f-p5.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002685 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002686 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002687 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002688 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002689 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002690 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002691 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
2692 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
2693 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
2694 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2695 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2696 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2697 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
2698 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
2699 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
2700 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2701 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
2702 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002703 "src/math/sqrt-avx512f-nr1fma.c",
2704 "src/math/sqrt-avx512f-nr2fma.c",
2705 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002706]
2707
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002708AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002709 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2710 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002711 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2712 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2713 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2714 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2715 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2716 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2717 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2718 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2719]
2720
Frank Barchardbcedc082020-08-17 18:00:51 -07002721WASM32_ASM_UKERNELS = [
2722 "src/f32-relu/wasm_shr_x1.S",
2723 "src/f32-relu/wasm_shr_x2.S",
2724 "src/f32-relu/wasm_shr_x4.S",
2725]
2726
Marat Dukhan08c4a432019-10-03 09:29:21 -07002727AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002728 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002729 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002730 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2731 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002732 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002733 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002734 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002735 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002736 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2737 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002738 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2739 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2740 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2741 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002742]
2743
2744AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002745 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002746 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2747 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002748 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002749 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2750 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002751 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2752 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2753 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002754 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002755 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2756 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2757 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002758 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002759 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2760 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002761 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002762 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2763 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2764 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2765 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2766 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2767 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2768 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2769 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2770 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2771 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2772 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2773 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2774 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2775 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2776 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2777 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2778 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2779 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002780 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2781 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002782 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002783 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2784 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2785 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2786 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2787 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2788 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2789 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2790 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2791 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2792 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2793 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2794 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2795 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2796 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2797 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2798 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2799 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2800 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002801 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2802 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2803 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2804 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2805 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2806 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2807 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2808 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2809 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2810 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2811 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2812 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2813 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2814 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2815 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2816 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2817 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2818 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard31bb45b2020-10-06 00:26:33 -07002819 "src/qs8-gemm/4x16c4-minmax-neondot.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002820]
2821
Marat Dukhan1b354632020-03-23 12:50:22 -07002822INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002823 "src/xnnpack/argmaxpool.h",
2824 "src/xnnpack/avgpool.h",
2825 "src/xnnpack/clamp.h",
2826 "src/xnnpack/common.h",
2827 "src/xnnpack/conv.h",
2828 "src/xnnpack/dwconv.h",
2829 "src/xnnpack/gavgpool.h",
2830 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002831 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002832 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002833 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002834 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002835 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002836 "src/xnnpack/lut.h",
2837 "src/xnnpack/math.h",
2838 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002839 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002840 "src/xnnpack/packx.h",
2841 "src/xnnpack/pad.h",
2842 "src/xnnpack/params.h",
2843 "src/xnnpack/pavgpool.h",
2844 "src/xnnpack/ppmm.h",
2845 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002846 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002847 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002848 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002849 "src/xnnpack/rmax.h",
2850 "src/xnnpack/scalar-utils.h",
2851 "src/xnnpack/spmm.h",
2852 "src/xnnpack/unpool.h",
2853 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002854 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002855 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002856 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002857 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002858 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002859 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002860 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002861]
2862
2863INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002864 "include/xnnpack.h",
2865 "src/xnnpack/allocator.h",
2866 "src/xnnpack/compute.h",
2867 "src/xnnpack/im2col.h",
2868 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002869 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002870 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002871 "src/xnnpack/operator.h",
2872 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002873 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002874 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002875 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002876 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002877]
2878
Marat Dukhan1b354632020-03-23 12:50:22 -07002879ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002880 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002881]
2882
Marat Dukhan1b354632020-03-23 12:50:22 -07002883MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002884 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002885 "include/xnnpack.h",
2886]
2887
Marat Dukhan1b354632020-03-23 12:50:22 -07002888MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002889 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002890 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002891 "src/xnnpack/requantization.h",
2892 "include/xnnpack.h",
2893]
2894
2895OPERATOR_TEST_PARAMS_HDRS = [
2896 "src/xnnpack/params.h",
2897 "src/xnnpack/common.h",
2898]
2899
2900WEIGHTS_PACK_HDRS = [
2901 "src/xnnpack/pack.h",
2902 "src/xnnpack/operator.h",
2903 "src/xnnpack/compute.h",
2904]
2905
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002906LOGGING_COPTS = select({
2907 # No logging in optimized mode
2908 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2909 # Full logging in debug mode
2910 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2911 # Error-only logging in default (fastbuild) mode
2912 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2913})
2914
Marat Dukhan3b59de22020-06-03 20:15:19 -07002915LOGGING_SRCS = select({
2916 # No logging in optimized mode
2917 ":optimized_build": [],
2918 "//conditions:default": [
2919 "src/operator-strings.c",
2920 "src/subgraph-strings.c",
2921 ],
2922})
2923
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002924LOGGING_HDRS = [
2925 "src/xnnpack/log.h",
2926]
2927
Marat Dukhan08c4a432019-10-03 09:29:21 -07002928xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002929 name = "tables",
2930 srcs = TABLE_SRCS,
2931 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002932 gcc_copts = xnnpack_gcc_std_copts(),
2933 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002934)
2935
2936xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002937 name = "scalar_ukernels",
2938 srcs = SCALAR_UKERNELS,
2939 hdrs = INTERNAL_HDRS,
2940 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002941 gcc_copts = xnnpack_gcc_std_copts(),
2942 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002943 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002944 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002945 "@FP16",
2946 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002947 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002948 ],
2949)
2950
2951xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002952 name = "scalar_ukernels_test_mode",
2953 srcs = SCALAR_UKERNELS,
2954 hdrs = INTERNAL_HDRS,
2955 aarch32_copts = ["-marm"],
2956 copts = [
2957 "-UNDEBUG",
2958 "-DXNN_TEST_MODE=1",
2959 ],
2960 gcc_copts = xnnpack_gcc_std_copts(),
2961 msvc_copts = xnnpack_msvc_std_copts(),
2962 deps = [
2963 ":tables",
2964 "@FP16",
2965 "@FXdiv",
2966 "@pthreadpool",
2967 ],
2968)
2969
2970xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002971 name = "wasm_ukernels",
2972 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002973 gcc_copts = xnnpack_gcc_std_copts(),
2974 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002975 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002976 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002977 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002978 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002979 "@FP16",
2980 "@FXdiv",
2981 "@pthreadpool",
2982 ],
2983)
2984
2985xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002986 name = "wasm_ukernels_test_mode",
2987 hdrs = INTERNAL_HDRS,
2988 copts = [
2989 "-UNDEBUG",
2990 "-DXNN_TEST_MODE=1",
2991 ],
2992 gcc_copts = xnnpack_gcc_std_copts(),
2993 msvc_copts = xnnpack_msvc_std_copts(),
2994 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002995 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002996 deps = [
2997 ":tables",
2998 "@FP16",
2999 "@FXdiv",
3000 "@pthreadpool",
3001 ],
3002)
3003
3004xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003005 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003006 hdrs = INTERNAL_HDRS,
3007 aarch32_copts = [
3008 "-marm",
3009 "-mfpu=neon",
3010 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003011 gcc_copts = xnnpack_gcc_std_copts(),
3012 gcc_x86_copts = ["-msse2"],
3013 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003014 optimized_copts = [
3015 "-O3",
3016 "-ffast-math",
3017 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003018 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003019 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003020 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003021 "@FP16",
3022 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003023 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003024 ],
3025)
3026
3027xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003028 name = "psimd_fastmath_ukernels_test_mode",
3029 hdrs = INTERNAL_HDRS,
3030 aarch32_copts = [
3031 "-marm",
3032 "-mfpu=neon",
3033 ],
3034 copts = [
3035 "-UNDEBUG",
3036 "-DXNN_TEST_MODE=1",
3037 ],
3038 gcc_copts = xnnpack_gcc_std_copts(),
3039 gcc_x86_copts = ["-msse2"],
3040 msvc_copts = xnnpack_msvc_std_copts(),
3041 optimized_copts = [
3042 "-O3",
3043 "-ffast-math",
3044 ],
3045 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
3046 deps = [
3047 ":tables",
3048 "@FP16",
3049 "@psimd",
3050 "@pthreadpool",
3051 ],
3052)
3053
3054xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003055 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003056 hdrs = INTERNAL_HDRS,
3057 aarch32_copts = [
3058 "-marm",
3059 "-mfpu=neon",
3060 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003061 gcc_copts = xnnpack_gcc_std_copts(),
3062 gcc_x86_copts = ["-msse2"],
3063 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003064 optimized_copts = [
3065 "-O3",
3066 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003067 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003068 deps = [
3069 ":tables",
3070 "@FP16",
3071 "@psimd",
3072 "@pthreadpool",
3073 ],
3074)
3075
3076xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003077 name = "psimd_accmath_ukernels_test_mode",
3078 hdrs = INTERNAL_HDRS,
3079 aarch32_copts = [
3080 "-marm",
3081 "-mfpu=neon",
3082 ],
3083 copts = [
3084 "-UNDEBUG",
3085 "-DXNN_TEST_MODE=1",
3086 ],
3087 gcc_copts = xnnpack_gcc_std_copts(),
3088 gcc_x86_copts = ["-msse2"],
3089 msvc_copts = xnnpack_msvc_std_copts(),
3090 optimized_copts = [
3091 "-O3",
3092 ],
3093 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3094 deps = [
3095 ":tables",
3096 "@FP16",
3097 "@psimd",
3098 "@pthreadpool",
3099 ],
3100)
3101
3102xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003103 name = "neon_ukernels",
3104 hdrs = INTERNAL_HDRS,
3105 aarch32_copts = [
3106 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003107 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003108 "-mfpu=neon",
3109 ],
3110 aarch32_srcs = NEON_UKERNELS,
3111 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003112 gcc_copts = xnnpack_gcc_std_copts(),
3113 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003114 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003115 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003116 "@FP16",
3117 "@pthreadpool",
3118 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003119)
3120
3121xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003122 name = "neon_ukernels_test_mode",
3123 hdrs = INTERNAL_HDRS,
3124 aarch32_copts = [
3125 "-marm",
3126 "-march=armv7-a",
3127 "-mfpu=neon",
3128 ],
3129 aarch32_srcs = NEON_UKERNELS,
3130 aarch64_srcs = NEON_UKERNELS,
3131 copts = [
3132 "-UNDEBUG",
3133 "-DXNN_TEST_MODE=1",
3134 ],
3135 gcc_copts = xnnpack_gcc_std_copts(),
3136 msvc_copts = xnnpack_msvc_std_copts(),
3137 deps = [
3138 ":tables",
3139 "@FP16",
3140 "@pthreadpool",
3141 ],
3142)
3143
3144xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003145 name = "neonfma_ukernels",
3146 hdrs = INTERNAL_HDRS,
3147 aarch32_copts = [
3148 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003149 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003150 "-mfpu=neon-vfpv4",
3151 ],
3152 aarch32_srcs = NEONFMA_UKERNELS,
3153 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003154 apple_aarch32_copts = [
3155 "-mcpu=swift",
3156 "-mtune=generic",
3157 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003158 gcc_copts = xnnpack_gcc_std_copts(),
3159 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003160 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003161 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003162 "@FP16",
3163 "@pthreadpool",
3164 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003165)
3166
3167xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003168 name = "neonfma_ukernels_test_mode",
3169 hdrs = INTERNAL_HDRS,
3170 aarch32_copts = [
3171 "-marm",
3172 "-march=armv7-a",
3173 "-mfpu=neon-vfpv4",
3174 ],
3175 aarch32_srcs = NEONFMA_UKERNELS,
3176 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003177 apple_aarch32_copts = [
3178 "-mcpu=swift",
3179 "-mtune=generic",
3180 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003181 copts = [
3182 "-UNDEBUG",
3183 "-DXNN_TEST_MODE=1",
3184 ],
3185 gcc_copts = xnnpack_gcc_std_copts(),
3186 msvc_copts = xnnpack_msvc_std_copts(),
3187 deps = [
3188 ":tables",
3189 "@FP16",
3190 "@pthreadpool",
3191 ],
3192)
3193
3194xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003195 name = "neonv8_ukernels",
3196 hdrs = INTERNAL_HDRS,
3197 aarch32_copts = [
3198 "-marm",
3199 "-march=armv8-a",
3200 "-mfpu=neon-fp-armv8",
3201 ],
3202 aarch32_srcs = NEONV8_UKERNELS,
3203 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003204 apple_aarch32_copts = [
3205 "-mcpu=cyclone",
3206 "-mtune=generic",
3207 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003208 gcc_copts = xnnpack_gcc_std_copts(),
3209 msvc_copts = xnnpack_msvc_std_copts(),
3210 deps = [
3211 ":tables",
3212 "@FP16",
3213 "@pthreadpool",
3214 ],
3215)
3216
3217xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003218 name = "neonv8_ukernels_test_mode",
3219 hdrs = INTERNAL_HDRS,
3220 aarch32_copts = [
3221 "-marm",
3222 "-march=armv8-a",
3223 "-mfpu=neon-fp-armv8",
3224 ],
3225 aarch32_srcs = NEONV8_UKERNELS,
3226 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003227 apple_aarch32_copts = [
3228 "-mcpu=cyclone",
3229 "-mtune=generic",
3230 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003231 copts = [
3232 "-UNDEBUG",
3233 "-DXNN_TEST_MODE=1",
3234 ],
3235 gcc_copts = xnnpack_gcc_std_copts(),
3236 msvc_copts = xnnpack_msvc_std_copts(),
3237 deps = [
3238 ":tables",
3239 "@FP16",
3240 "@pthreadpool",
3241 ],
3242)
3243
3244xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003245 name = "neonfp16arith_ukernels",
3246 hdrs = INTERNAL_HDRS,
3247 aarch64_copts = ["-march=armv8.2-a+fp16"],
3248 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003249 gcc_copts = xnnpack_gcc_std_copts(),
3250 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003251 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003252 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003253 "@FP16",
3254 "@pthreadpool",
3255 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003256)
3257
3258xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003259 name = "neonfp16arith_ukernels_test_mode",
3260 hdrs = INTERNAL_HDRS,
3261 aarch64_copts = ["-march=armv8.2-a+fp16"],
3262 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3263 copts = [
3264 "-UNDEBUG",
3265 "-DXNN_TEST_MODE=1",
3266 ],
3267 gcc_copts = xnnpack_gcc_std_copts(),
3268 msvc_copts = xnnpack_msvc_std_copts(),
3269 deps = [
3270 ":tables",
3271 "@FP16",
3272 "@pthreadpool",
3273 ],
3274)
3275
3276xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003277 name = "neondot_ukernels",
3278 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003279 aarch32_copts = [
3280 "-marm",
3281 "-march=armv8.2-a+dotprod",
3282 "-mfpu=neon-fp-armv8",
3283 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003284 aarch32_srcs = NEONDOT_UKERNELS,
3285 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3286 aarch64_srcs = NEONDOT_UKERNELS,
3287 gcc_copts = xnnpack_gcc_std_copts(),
3288 msvc_copts = xnnpack_msvc_std_copts(),
3289 deps = [
3290 ":tables",
3291 "@FP16",
3292 "@pthreadpool",
3293 ],
3294)
3295
3296xnnpack_cc_library(
3297 name = "neondot_ukernels_test_mode",
3298 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003299 aarch32_copts = [
3300 "-marm",
3301 "-march=armv8.2-a+dotprod",
3302 "-mfpu=neon-fp-armv8",
3303 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003304 aarch32_srcs = NEONDOT_UKERNELS,
3305 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3306 aarch64_srcs = NEONDOT_UKERNELS,
3307 copts = [
3308 "-UNDEBUG",
3309 "-DXNN_TEST_MODE=1",
3310 ],
3311 gcc_copts = xnnpack_gcc_std_copts(),
3312 msvc_copts = xnnpack_msvc_std_copts(),
3313 deps = [
3314 ":tables",
3315 "@FP16",
3316 "@pthreadpool",
3317 ],
3318)
3319
3320xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003321 name = "sse2_ukernels",
3322 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003323 gcc_copts = xnnpack_gcc_std_copts(),
3324 gcc_x86_copts = ["-msse2"],
3325 msvc_copts = xnnpack_msvc_std_copts(),
3326 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003327 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003328 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003329 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003330 "@FP16",
3331 "@pthreadpool",
3332 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003333)
3334
3335xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003336 name = "sse2_ukernels_test_mode",
3337 hdrs = INTERNAL_HDRS,
3338 copts = [
3339 "-UNDEBUG",
3340 "-DXNN_TEST_MODE=1",
3341 ],
3342 gcc_copts = xnnpack_gcc_std_copts(),
3343 gcc_x86_copts = ["-msse2"],
3344 msvc_copts = xnnpack_msvc_std_copts(),
3345 msvc_x86_32_copts = ["/arch:SSE2"],
3346 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3347 deps = [
3348 ":tables",
3349 "@FP16",
3350 "@pthreadpool",
3351 ],
3352)
3353
3354xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003355 name = "ssse3_ukernels",
3356 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003357 gcc_copts = xnnpack_gcc_std_copts(),
3358 gcc_x86_copts = ["-mssse3"],
3359 msvc_copts = xnnpack_msvc_std_copts(),
3360 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003361 x86_srcs = SSSE3_UKERNELS,
3362 deps = [
3363 ":tables",
3364 "@FP16",
3365 "@pthreadpool",
3366 ],
3367)
3368
3369xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003370 name = "ssse3_ukernels_test_mode",
3371 hdrs = INTERNAL_HDRS,
3372 copts = [
3373 "-UNDEBUG",
3374 "-DXNN_TEST_MODE=1",
3375 ],
3376 gcc_copts = xnnpack_gcc_std_copts(),
3377 gcc_x86_copts = ["-mssse3"],
3378 msvc_copts = xnnpack_msvc_std_copts(),
3379 msvc_x86_32_copts = ["/arch:SSE2"],
3380 x86_srcs = SSSE3_UKERNELS,
3381 deps = [
3382 ":tables",
3383 "@FP16",
3384 "@pthreadpool",
3385 ],
3386)
3387
3388xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003389 name = "sse41_ukernels",
3390 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003391 gcc_copts = xnnpack_gcc_std_copts(),
3392 gcc_x86_copts = ["-msse4.1"],
3393 msvc_copts = xnnpack_msvc_std_copts(),
3394 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003395 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003396 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003397 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003398 "@FP16",
3399 "@pthreadpool",
3400 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003401)
3402
3403xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003404 name = "sse41_ukernels_test_mode",
3405 hdrs = INTERNAL_HDRS,
3406 copts = [
3407 "-UNDEBUG",
3408 "-DXNN_TEST_MODE=1",
3409 ],
3410 gcc_copts = xnnpack_gcc_std_copts(),
3411 gcc_x86_copts = ["-msse4.1"],
3412 msvc_copts = xnnpack_msvc_std_copts(),
3413 msvc_x86_32_copts = ["/arch:SSE2"],
3414 x86_srcs = SSE41_UKERNELS,
3415 deps = [
3416 ":tables",
3417 "@FP16",
3418 "@pthreadpool",
3419 ],
3420)
3421
3422xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003423 name = "avx_ukernels",
3424 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003425 gcc_copts = xnnpack_gcc_std_copts(),
3426 gcc_x86_copts = ["-mavx"],
3427 msvc_copts = xnnpack_msvc_std_copts(),
3428 msvc_x86_32_copts = ["/arch:AVX"],
3429 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003430 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003431 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003432 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003433 "@FP16",
3434 "@pthreadpool",
3435 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003436)
3437
3438xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003439 name = "avx_ukernels_test_mode",
3440 hdrs = INTERNAL_HDRS,
3441 copts = [
3442 "-UNDEBUG",
3443 "-DXNN_TEST_MODE=1",
3444 ],
3445 gcc_copts = xnnpack_gcc_std_copts(),
3446 gcc_x86_copts = ["-mavx"],
3447 msvc_copts = xnnpack_msvc_std_copts(),
3448 msvc_x86_32_copts = ["/arch:AVX"],
3449 msvc_x86_64_copts = ["/arch:AVX"],
3450 x86_srcs = AVX_UKERNELS,
3451 deps = [
3452 ":tables",
3453 "@FP16",
3454 "@pthreadpool",
3455 ],
3456)
3457
3458xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003459 name = "xop_ukernels",
3460 hdrs = INTERNAL_HDRS,
3461 gcc_copts = xnnpack_gcc_std_copts(),
3462 gcc_x86_copts = ["-mxop"],
3463 msvc_copts = xnnpack_msvc_std_copts(),
3464 msvc_x86_32_copts = ["/arch:AVX"],
3465 msvc_x86_64_copts = ["/arch:AVX"],
3466 x86_srcs = XOP_UKERNELS,
3467 deps = [
3468 ":tables",
3469 "@FP16",
3470 "@pthreadpool",
3471 ],
3472)
3473
3474xnnpack_cc_library(
3475 name = "xop_ukernels_test_mode",
3476 hdrs = INTERNAL_HDRS,
3477 copts = [
3478 "-UNDEBUG",
3479 "-DXNN_TEST_MODE=1",
3480 ],
3481 gcc_copts = xnnpack_gcc_std_copts(),
3482 gcc_x86_copts = ["-mxop"],
3483 msvc_copts = xnnpack_msvc_std_copts(),
3484 msvc_x86_32_copts = ["/arch:AVX"],
3485 msvc_x86_64_copts = ["/arch:AVX"],
3486 x86_srcs = XOP_UKERNELS,
3487 deps = [
3488 ":tables",
3489 "@FP16",
3490 "@pthreadpool",
3491 ],
3492)
3493
3494xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003495 name = "fma3_ukernels",
3496 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003497 gcc_copts = xnnpack_gcc_std_copts(),
3498 gcc_x86_copts = ["-mfma"],
3499 msvc_copts = xnnpack_msvc_std_copts(),
3500 msvc_x86_32_copts = ["/arch:AVX"],
3501 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003502 x86_srcs = FMA3_UKERNELS,
3503 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003504 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003505 "@FP16",
3506 "@pthreadpool",
3507 ],
3508)
3509
3510xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003511 name = "fma3_ukernels_test_mode",
3512 hdrs = INTERNAL_HDRS,
3513 copts = [
3514 "-UNDEBUG",
3515 "-DXNN_TEST_MODE=1",
3516 ],
3517 gcc_copts = xnnpack_gcc_std_copts(),
3518 gcc_x86_copts = ["-mfma"],
3519 msvc_copts = xnnpack_msvc_std_copts(),
3520 msvc_x86_32_copts = ["/arch:AVX"],
3521 msvc_x86_64_copts = ["/arch:AVX"],
3522 x86_srcs = FMA3_UKERNELS,
3523 deps = [
3524 ":tables",
3525 "@FP16",
3526 "@pthreadpool",
3527 ],
3528)
3529
3530xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003531 name = "avx2_ukernels",
3532 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003533 gcc_copts = xnnpack_gcc_std_copts(),
3534 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003535 "-mfma",
3536 "-mavx2",
3537 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003538 msvc_copts = xnnpack_msvc_std_copts(),
3539 msvc_x86_32_copts = ["/arch:AVX2"],
3540 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003541 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003542 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003543 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003544 "@FP16",
3545 "@pthreadpool",
3546 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003547)
3548
3549xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003550 name = "avx2_ukernels_test_mode",
3551 hdrs = INTERNAL_HDRS,
3552 copts = [
3553 "-UNDEBUG",
3554 "-DXNN_TEST_MODE=1",
3555 ],
3556 gcc_copts = xnnpack_gcc_std_copts(),
3557 gcc_x86_copts = [
3558 "-mfma",
3559 "-mavx2",
3560 ],
3561 msvc_copts = xnnpack_msvc_std_copts(),
3562 msvc_x86_32_copts = ["/arch:AVX2"],
3563 msvc_x86_64_copts = ["/arch:AVX2"],
3564 x86_srcs = AVX2_UKERNELS,
3565 deps = [
3566 ":tables",
3567 "@FP16",
3568 "@pthreadpool",
3569 ],
3570)
3571
3572xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003573 name = "avx512f_ukernels",
3574 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003575 gcc_copts = xnnpack_gcc_std_copts(),
3576 gcc_x86_copts = ["-mavx512f"],
3577 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3578 msvc_copts = xnnpack_msvc_std_copts(),
3579 msvc_x86_32_copts = ["/arch:AVX512"],
3580 msvc_x86_64_copts = ["/arch:AVX512"],
3581 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003582 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003583 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003584 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003585 "@FP16",
3586 "@pthreadpool",
3587 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003588)
3589
3590xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003591 name = "avx512f_ukernels_test_mode",
3592 hdrs = INTERNAL_HDRS,
3593 copts = [
3594 "-UNDEBUG",
3595 "-DXNN_TEST_MODE=1",
3596 ],
3597 gcc_copts = xnnpack_gcc_std_copts(),
3598 gcc_x86_copts = ["-mavx512f"],
3599 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3600 msvc_copts = xnnpack_msvc_std_copts(),
3601 msvc_x86_32_copts = ["/arch:AVX512"],
3602 msvc_x86_64_copts = ["/arch:AVX512"],
3603 msys_copts = ["-fno-asynchronous-unwind-tables"],
3604 x86_srcs = AVX512F_UKERNELS,
3605 deps = [
3606 ":tables",
3607 "@FP16",
3608 "@pthreadpool",
3609 ],
3610)
3611
3612xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003613 name = "avx512skx_ukernels",
3614 hdrs = INTERNAL_HDRS,
3615 gcc_copts = xnnpack_gcc_std_copts(),
3616 gcc_x86_copts = [
3617 "-mavx512f",
3618 "-mavx512cd",
3619 "-mavx512bw",
3620 "-mavx512dq",
3621 "-mavx512vl",
3622 ],
3623 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3624 msvc_copts = xnnpack_msvc_std_copts(),
3625 msvc_x86_32_copts = ["/arch:AVX512"],
3626 msvc_x86_64_copts = ["/arch:AVX512"],
3627 msys_copts = ["-fno-asynchronous-unwind-tables"],
3628 x86_srcs = AVX512SKX_UKERNELS,
3629 deps = [
3630 ":tables",
3631 "@FP16",
3632 "@pthreadpool",
3633 ],
3634)
3635
3636xnnpack_cc_library(
3637 name = "avx512skx_ukernels_test_mode",
3638 hdrs = INTERNAL_HDRS,
3639 copts = [
3640 "-UNDEBUG",
3641 "-DXNN_TEST_MODE=1",
3642 ],
3643 gcc_copts = xnnpack_gcc_std_copts(),
3644 gcc_x86_copts = [
3645 "-mavx512f",
3646 "-mavx512cd",
3647 "-mavx512bw",
3648 "-mavx512dq",
3649 "-mavx512vl",
3650 ],
3651 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3652 msvc_copts = xnnpack_msvc_std_copts(),
3653 msvc_x86_32_copts = ["/arch:AVX512"],
3654 msvc_x86_64_copts = ["/arch:AVX512"],
3655 msys_copts = ["-fno-asynchronous-unwind-tables"],
3656 x86_srcs = AVX512SKX_UKERNELS,
3657 deps = [
3658 ":tables",
3659 "@FP16",
3660 "@pthreadpool",
3661 ],
3662)
3663
3664xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003665 name = "asm_ukernels",
3666 hdrs = ["src/xnnpack/assembly.h"],
3667 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07003668 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003669 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003670 wasm_srcs = WASM32_ASM_UKERNELS,
3671 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003672)
3673
Marat Dukhan3b59de22020-06-03 20:15:19 -07003674xnnpack_cc_library(
3675 name = "logging_utils",
3676 srcs = LOGGING_SRCS,
3677 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3678 copts = LOGGING_COPTS + [
3679 "-Isrc",
3680 "-Iinclude",
3681 ] + select({
3682 ":debug_build": [],
3683 "//conditions:default": xnnpack_min_size_copts(),
3684 }),
3685 gcc_copts = xnnpack_gcc_std_copts(),
3686 msvc_copts = xnnpack_msvc_std_copts(),
3687 visibility = xnnpack_visibility(),
3688 deps = [
3689 "@FP16",
3690 "@clog",
3691 "@pthreadpool",
3692 ],
3693)
3694
Marat Dukhan08c4a432019-10-03 09:29:21 -07003695xnnpack_aggregate_library(
3696 name = "ukernels",
3697 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003698 ":neon_ukernels",
3699 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003700 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003701 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003702 ":asm_ukernels",
3703 ],
3704 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003705 ":neon_ukernels",
3706 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003707 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003708 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003709 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003710 ":asm_ukernels",
3711 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003712 generic_deps = [
3713 ":scalar_ukernels",
3714 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003715 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003716 ":psimd_fastmath_ukernels",
3717 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003718 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003719 wasm_deps = [
3720 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003721 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003722 ],
3723 wasmsimd_deps = [
3724 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003725 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003726 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003727 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003728 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003729 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003730 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003731 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003732 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003733 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003734 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003735 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003736 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003737 ],
3738)
3739
Marat Dukhan33fcf782020-05-24 14:27:15 -07003740xnnpack_aggregate_library(
3741 name = "ukernels_test_mode",
3742 aarch32_deps = [
3743 ":neon_ukernels_test_mode",
3744 ":neonfma_ukernels_test_mode",
3745 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003746 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003747 ":asm_ukernels",
3748 ],
3749 aarch64_deps = [
3750 ":neon_ukernels_test_mode",
3751 ":neonfma_ukernels_test_mode",
3752 ":neonv8_ukernels_test_mode",
3753 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003754 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003755 ":asm_ukernels",
3756 ],
3757 generic_deps = [
3758 ":scalar_ukernels_test_mode",
3759 ],
3760 psimd_deps = [
3761 ":psimd_fastmath_ukernels_test_mode",
3762 ":psimd_accmath_ukernels_test_mode",
3763 ],
3764 wasm_deps = [
3765 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003766 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003767 ],
3768 wasmsimd_deps = [
3769 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003770 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003771 ],
3772 x86_deps = [
3773 ":sse2_ukernels_test_mode",
3774 ":ssse3_ukernels_test_mode",
3775 ":sse41_ukernels_test_mode",
3776 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003777 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003778 ":fma3_ukernels_test_mode",
3779 ":avx2_ukernels_test_mode",
3780 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003781 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003782 ],
3783)
3784
Marat Dukhan86df3842020-06-17 14:03:31 -07003785# This target must be compatible with Android NDK r17 and compile with Clang 6.
3786# It is similar to :ukernels target, but excludes microkernels which require
3787# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003788xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003789 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003790 aarch32_deps = [
3791 ":neon_ukernels",
3792 ":neonfma_ukernels",
3793 ":neonv8_ukernels",
3794 ":asm_ukernels",
3795 ],
3796 aarch64_deps = [
3797 ":neon_ukernels",
3798 ":neonfma_ukernels",
3799 ":neonv8_ukernels",
3800 ":asm_ukernels",
3801 ],
3802 generic_deps = [
3803 ":scalar_ukernels",
3804 ],
3805 psimd_deps = [
3806 ":psimd_fastmath_ukernels",
3807 ":psimd_accmath_ukernels",
3808 ],
3809 wasm_deps = [
3810 ":wasm_ukernels",
3811 ],
3812 wasmsimd_deps = [
3813 ":wasm_ukernels",
3814 ],
3815 x86_deps = [
3816 ":sse2_ukernels",
3817 ":ssse3_ukernels",
3818 ":sse41_ukernels",
3819 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003820 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003821 ":fma3_ukernels",
3822 ":avx2_ukernels",
3823 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003824 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003825 ],
3826)
3827
Marat Dukhan08c4a432019-10-03 09:29:21 -07003828xnnpack_cc_library(
3829 name = "im2col",
3830 srcs = ["src/im2col.c"],
3831 hdrs = [
3832 "src/xnnpack/common.h",
3833 "src/xnnpack/im2col.h",
3834 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003835 gcc_copts = xnnpack_gcc_std_copts(),
3836 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003837)
3838
3839xnnpack_cc_library(
3840 name = "indirection",
3841 srcs = ["src/indirection.c"],
3842 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003843 gcc_copts = xnnpack_gcc_std_copts(),
3844 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003845 deps = [
3846 "@FP16",
3847 "@FXdiv",
3848 "@pthreadpool",
3849 ],
3850)
3851
3852xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003853 name = "indirection_test_mode",
3854 srcs = ["src/indirection.c"],
3855 hdrs = INTERNAL_HDRS,
3856 copts = [
3857 "-UNDEBUG",
3858 "-DXNN_TEST_MODE=1",
3859 ],
3860 gcc_copts = xnnpack_gcc_std_copts(),
3861 msvc_copts = xnnpack_msvc_std_copts(),
3862 deps = [
3863 "@FP16",
3864 "@FXdiv",
3865 "@pthreadpool",
3866 ],
3867)
3868
3869xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003870 name = "packing",
3871 srcs = ["src/packing.c"],
3872 hdrs = INTERNAL_HDRS,
3873 gcc_copts = xnnpack_gcc_std_copts(),
3874 msvc_copts = xnnpack_msvc_std_copts(),
3875 deps = [
3876 "@FP16",
3877 "@FXdiv",
3878 "@pthreadpool",
3879 ],
3880)
3881
3882xnnpack_cc_library(
3883 name = "packing_test_mode",
3884 srcs = ["src/packing.c"],
3885 hdrs = INTERNAL_HDRS,
3886 copts = [
3887 "-UNDEBUG",
3888 "-DXNN_TEST_MODE=1",
3889 ],
3890 gcc_copts = xnnpack_gcc_std_copts(),
3891 msvc_copts = xnnpack_msvc_std_copts(),
3892 deps = [
3893 "@FP16",
3894 "@FXdiv",
3895 "@pthreadpool",
3896 ],
3897)
3898
3899xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003900 name = "operator_run",
3901 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003902 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003903 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003904 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3905 "//conditions:default": [],
3906 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003907 gcc_copts = xnnpack_gcc_std_copts(),
3908 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003909 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003910 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003911 "@FP16",
3912 "@FXdiv",
3913 "@clog",
3914 "@pthreadpool",
3915 ],
3916)
3917
Chao Mei6ddfc602020-05-13 22:29:36 -07003918xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003919 name = "operator_run_test_mode",
3920 srcs = ["src/operator-run.c"],
3921 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3922 copts = LOGGING_COPTS + [
3923 "-UNDEBUG",
3924 "-DXNN_TEST_MODE=1",
3925 ] + select({
3926 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3927 "//conditions:default": [],
3928 }),
3929 gcc_copts = xnnpack_gcc_std_copts(),
3930 msvc_copts = xnnpack_msvc_std_copts(),
3931 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003932 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003933 "@FP16",
3934 "@FXdiv",
3935 "@clog",
3936 "@pthreadpool",
3937 ],
3938)
3939
3940xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003941 name = "memory_planner",
3942 srcs = ["src/memory-planner.c"],
3943 hdrs = INTERNAL_HDRS,
3944 defines = select({
3945 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3946 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3947 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3948 }),
3949 gcc_copts = xnnpack_gcc_std_copts(),
3950 msvc_copts = xnnpack_msvc_std_copts(),
3951 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003952 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003953 "@pthreadpool",
3954 ],
3955)
3956
Marat Dukhan33fcf782020-05-24 14:27:15 -07003957xnnpack_cc_library(
3958 name = "memory_planner_test_mode",
3959 srcs = ["src/memory-planner.c"],
3960 hdrs = INTERNAL_HDRS,
3961 copts = [
3962 "-UNDEBUG",
3963 "-DXNN_TEST_MODE=1",
3964 ],
3965 defines = select({
3966 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3967 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3968 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3969 }),
3970 gcc_copts = xnnpack_gcc_std_copts(),
3971 msvc_copts = xnnpack_msvc_std_copts(),
3972 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003973 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003974 "@pthreadpool",
3975 ],
3976)
3977
Marat Dukhan08c4a432019-10-03 09:29:21 -07003978cc_library(
3979 name = "enable_assembly",
3980 defines = select({
3981 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3982 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003983 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003984 }),
3985)
3986
Marat Dukhan9de90e02020-06-18 16:04:12 -07003987cc_library(
3988 name = "enable_sparse",
3989 defines = select({
3990 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3991 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003992 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003993 }),
3994)
3995
Marat Dukhancf056b22019-10-07 10:26:29 -07003996xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003997 name = "operators",
3998 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003999 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004000 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004001 ],
4002 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004003 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004004 "-Isrc",
4005 "-Iinclude",
4006 ] + select({
4007 ":debug_build": [],
4008 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004009 }) + select({
4010 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4011 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004012 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004013 gcc_copts = xnnpack_gcc_std_copts(),
4014 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004015 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004016 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004017 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004018 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004019 "@FP16",
4020 "@FXdiv",
4021 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004022 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004023 ],
4024)
4025
Marat Dukhan10a38082020-04-17 03:58:35 -07004026xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004027 name = "operators_test_mode",
4028 srcs = OPERATOR_SRCS + [
4029 "src/memory.c",
4030 "src/operator-delete.c",
4031 ],
4032 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4033 copts = LOGGING_COPTS + [
4034 "-Isrc",
4035 "-Iinclude",
4036 "-UNDEBUG",
4037 "-DXNN_TEST_MODE=1",
4038 ] + select({
4039 ":debug_build": [],
4040 "//conditions:default": xnnpack_min_size_copts(),
4041 }) + select({
4042 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4043 "//conditions:default": [],
4044 }),
4045 gcc_copts = xnnpack_gcc_std_copts(),
4046 msvc_copts = xnnpack_msvc_std_copts(),
4047 deps = [
4048 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004049 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004050 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004051 "@FP16",
4052 "@FXdiv",
4053 "@clog",
4054 "@pthreadpool",
4055 ],
4056)
4057
4058xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004059 name = "XNNPACK",
4060 srcs = [
4061 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004062 "src/runtime.c",
4063 "src/subgraph.c",
4064 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004065 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004066 hdrs = ["include/xnnpack.h"],
4067 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004068 "-Isrc",
4069 "-Iinclude",
4070 ] + select({
4071 ":debug_build": [],
4072 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004073 }) + select({
4074 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4075 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004076 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004077 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004078 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004079 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004080 visibility = xnnpack_visibility(),
4081 deps = [
4082 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004083 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004084 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004085 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004086 ":operator_run",
4087 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004088 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004089 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004090 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004091 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004092 ] + select({
4093 ":emscripten": [],
4094 "//conditions:default": ["@cpuinfo"],
4095 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004096)
4097
Marat Dukhan10a38082020-04-17 03:58:35 -07004098xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004099 name = "XNNPACK_test_mode",
4100 srcs = [
4101 "src/init.c",
4102 "src/runtime.c",
4103 "src/subgraph.c",
4104 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004105 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004106 hdrs = ["include/xnnpack.h"],
4107 copts = LOGGING_COPTS + [
4108 "-Isrc",
4109 "-Iinclude",
4110 "-UNDEBUG",
4111 "-DXNN_TEST_MODE=1",
4112 ] + select({
4113 ":debug_build": [],
4114 "//conditions:default": xnnpack_min_size_copts(),
4115 }) + select({
4116 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4117 "//conditions:default": [],
4118 }),
4119 gcc_copts = xnnpack_gcc_std_copts(),
4120 includes = ["include"],
4121 msvc_copts = xnnpack_msvc_std_copts(),
4122 visibility = xnnpack_visibility(),
4123 deps = [
4124 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004125 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004126 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004127 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004128 ":operator_run_test_mode",
4129 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004130 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004131 "@clog",
4132 "@FP16",
4133 "@pthreadpool",
4134 ] + select({
4135 ":emscripten": [],
4136 "//conditions:default": ["@cpuinfo"],
4137 }),
4138)
4139
4140xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004141 name = "xnnpack_f32",
4142 srcs = [
4143 "src/init.c",
4144 "src/runtime.c",
4145 "src/subgraph.c",
4146 "src/tensor.c",
4147 ] + SUBGRAPH_SRCS,
4148 hdrs = ["include/xnnpack.h"],
4149 copts = LOGGING_COPTS + [
4150 "-Isrc",
4151 "-Iinclude",
4152 ] + select({
4153 ":debug_build": [],
4154 "//conditions:default": xnnpack_min_size_copts(),
4155 }) + select({
4156 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4157 "//conditions:default": [],
4158 }),
4159 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004160 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004161 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004162 "XNN_NO_U8_OPERATORS",
4163 "XNN_NO_X8_OPERATORS",
4164 "XNN_NO_F16_OPERATORS",
4165 "XNN_NO_X16_OPERATORS",
4166 ],
4167 gcc_copts = xnnpack_gcc_std_copts(),
4168 includes = ["include"],
4169 msvc_copts = xnnpack_msvc_std_copts(),
4170 visibility = xnnpack_visibility(),
4171 deps = [
4172 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004173 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004174 ":logging_utils",
4175 ":memory_planner",
4176 ":operator_run",
4177 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004178 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004179 "@clog",
4180 "@FP16",
4181 "@pthreadpool",
4182 ] + select({
4183 ":emscripten": [],
4184 "//conditions:default": ["@cpuinfo"],
4185 }),
4186)
4187
4188xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004189 name = "xnnpack_operators_nhwc_f32",
4190 srcs = [
4191 "src/init.c",
4192 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004193 hdrs = ["include/xnnpack.h"],
4194 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004195 "-Isrc",
4196 "-Iinclude",
4197 ] + select({
4198 ":debug_build": [],
4199 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004200 }) + select({
4201 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4202 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004203 }),
4204 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004205 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004206 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004207 "XNN_NO_U8_OPERATORS",
4208 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004209 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004210 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004211 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004212 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004213 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004214 visibility = xnnpack_visibility(),
4215 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004216 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004217 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004218 ":operator_run",
4219 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004220 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004221 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004222 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004223 ] + select({
4224 ":emscripten": [],
4225 "//conditions:default": ["@cpuinfo"],
4226 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004227)
4228
Marat Dukhancf056b22019-10-07 10:26:29 -07004229xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004230 name = "bench_utils",
4231 srcs = ["bench/utils.cc"],
4232 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004233 deps = [
4234 "@com_google_benchmark//:benchmark",
4235 "@cpuinfo",
4236 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004237)
4238
Frank Barchard7e955972019-10-11 10:34:25 -07004239######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004240
4241xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004242 name = "qs8_gemm_bench",
4243 srcs = [
4244 "bench/gemm.h",
4245 "bench/qs8-gemm.cc",
4246 "src/xnnpack/AlignedAllocator.h",
4247 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
4248 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
4249)
4250
4251xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004252 name = "qs8_requantization_bench",
4253 srcs = [
4254 "bench/qs8-requantization.cc",
4255 "src/xnnpack/requantization-stubs.h",
4256 "src/xnnpack/AlignedAllocator.h",
4257 ] + MICROKERNEL_BENCHMARK_HDRS,
4258 deps = MICROKERNEL_BENCHMARK_DEPS,
4259)
4260
4261xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004262 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004263 srcs = [
4264 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004265 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004266 "src/xnnpack/AlignedAllocator.h",
4267 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004268 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004269 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004270)
4271
4272xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004273 name = "qu8_requantization_bench",
4274 srcs = [
4275 "bench/qu8-requantization.cc",
4276 "src/xnnpack/requantization-stubs.h",
4277 "src/xnnpack/AlignedAllocator.h",
4278 ] + MICROKERNEL_BENCHMARK_HDRS,
4279 deps = MICROKERNEL_BENCHMARK_DEPS,
4280)
4281
4282xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004283 name = "f16_igemm_bench",
4284 srcs = [
4285 "bench/f16-igemm.cc",
4286 "bench/conv.h",
4287 "bench/google/conv.h",
4288 "src/xnnpack/AlignedAllocator.h",
4289 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004290 deps = MICROKERNEL_BENCHMARK_DEPS + [
4291 ":indirection",
4292 ":packing",
4293 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004294)
4295
4296xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004297 name = "f16_gemm_bench",
4298 srcs = [
4299 "bench/f16-gemm.cc",
4300 "bench/gemm.h",
4301 "src/xnnpack/AlignedAllocator.h",
4302 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004303 deps = MICROKERNEL_BENCHMARK_DEPS + [
4304 ":packing",
4305 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004306)
4307
4308xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004309 name = "f16_spmm_bench",
4310 srcs = [
4311 "bench/f16-spmm.cc",
4312 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004313 "src/xnnpack/AlignedAllocator.h",
4314 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004315 deps = MICROKERNEL_BENCHMARK_DEPS,
4316)
4317
4318xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004319 name = "f32_igemm_bench",
4320 srcs = [
4321 "bench/f32-igemm.cc",
4322 "bench/conv.h",
4323 "src/xnnpack/AlignedAllocator.h",
4324 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004325 deps = MICROKERNEL_BENCHMARK_DEPS + [
4326 ":indirection",
4327 ":packing",
4328 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004329)
4330
4331xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004332 name = "f16_relu_bench",
4333 srcs = [
4334 "bench/f16-relu.cc",
4335 "src/xnnpack/AlignedAllocator.h",
4336 ] + MICROKERNEL_BENCHMARK_HDRS,
4337 deps = MICROKERNEL_BENCHMARK_DEPS,
4338)
4339
4340xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004341 name = "f32_conv_hwc_bench",
4342 srcs = [
4343 "bench/f32-conv-hwc.cc",
4344 "bench/dconv.h",
4345 "src/xnnpack/AlignedAllocator.h",
4346 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004347 deps = MICROKERNEL_BENCHMARK_DEPS + [
4348 ":packing",
4349 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004350)
4351
4352xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004353 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004354 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004355 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004356 "bench/dconv.h",
4357 "src/xnnpack/AlignedAllocator.h",
4358 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004359 deps = MICROKERNEL_BENCHMARK_DEPS + [
4360 ":packing",
4361 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004362)
4363
4364xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004365 name = "f16_dwconv_bench",
4366 srcs = [
4367 "bench/f16-dwconv.cc",
4368 "bench/dwconv.h",
4369 "bench/google/dwconv.h",
4370 "src/xnnpack/AlignedAllocator.h",
4371 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004372 deps = MICROKERNEL_BENCHMARK_DEPS + [
4373 ":indirection",
4374 ":packing",
4375 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004376)
4377
4378xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004379 name = "f32_dwconv_bench",
4380 srcs = [
4381 "bench/f32-dwconv.cc",
4382 "bench/dwconv.h",
4383 "src/xnnpack/AlignedAllocator.h",
4384 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004385 deps = MICROKERNEL_BENCHMARK_DEPS + [
4386 ":indirection",
4387 ":packing",
4388 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004389)
4390
4391xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004392 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004393 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004394 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004395 "bench/dwconv.h",
4396 "src/xnnpack/AlignedAllocator.h",
4397 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004398 deps = MICROKERNEL_BENCHMARK_DEPS + [
4399 ":indirection",
4400 ":packing",
4401 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004402)
4403
4404xnnpack_benchmark(
4405 name = "f32_gemm_bench",
4406 srcs = [
4407 "bench/f32-gemm.cc",
4408 "bench/gemm.h",
4409 "src/xnnpack/AlignedAllocator.h",
4410 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004411 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004412 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004413)
4414
4415xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004416 name = "f32_hswish_bench",
4417 srcs = [
4418 "bench/f32-hswish.cc",
4419 "src/xnnpack/AlignedAllocator.h",
4420 ] + MICROKERNEL_BENCHMARK_HDRS,
4421 deps = MICROKERNEL_BENCHMARK_DEPS,
4422)
4423
4424xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004425 name = "f32_raddexpminusmax_bench",
4426 srcs = [
4427 "bench/f32-raddexpminusmax.cc",
4428 "src/xnnpack/AlignedAllocator.h",
4429 ] + MICROKERNEL_BENCHMARK_HDRS,
4430 deps = MICROKERNEL_BENCHMARK_DEPS,
4431)
4432
4433xnnpack_benchmark(
4434 name = "f32_raddextexp_bench",
4435 srcs = [
4436 "bench/f32-raddextexp.cc",
4437 "src/xnnpack/AlignedAllocator.h",
4438 ] + MICROKERNEL_BENCHMARK_HDRS,
4439 deps = MICROKERNEL_BENCHMARK_DEPS,
4440)
4441
4442xnnpack_benchmark(
4443 name = "f32_raddstoreexpminusmax_bench",
4444 srcs = [
4445 "bench/f32-raddstoreexpminusmax.cc",
4446 "src/xnnpack/AlignedAllocator.h",
4447 ] + MICROKERNEL_BENCHMARK_HDRS,
4448 deps = MICROKERNEL_BENCHMARK_DEPS,
4449)
4450
4451xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004452 name = "f32_relu_bench",
4453 srcs = [
4454 "bench/f32-relu.cc",
4455 "src/xnnpack/AlignedAllocator.h",
4456 ] + MICROKERNEL_BENCHMARK_HDRS,
4457 deps = MICROKERNEL_BENCHMARK_DEPS,
4458)
4459
4460xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004461 name = "f32_rmax_bench",
4462 srcs = [
4463 "bench/f32-rmax.cc",
4464 "src/xnnpack/AlignedAllocator.h",
4465 ] + MICROKERNEL_BENCHMARK_HDRS,
4466 deps = MICROKERNEL_BENCHMARK_DEPS,
4467)
4468
4469xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004470 name = "f32_sigmoid_bench",
4471 srcs = [
4472 "bench/f32-sigmoid.cc",
4473 "src/xnnpack/AlignedAllocator.h",
4474 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004475 deps = MICROKERNEL_BENCHMARK_DEPS,
4476)
4477
4478xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004479 name = "f32_spmm_bench",
4480 srcs = [
4481 "bench/f32-spmm.cc",
4482 "bench/gemm.h",
4483 "src/xnnpack/AlignedAllocator.h",
4484 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004485 deps = MICROKERNEL_BENCHMARK_DEPS,
4486)
4487
4488xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004489 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004490 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004491 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004492 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004493 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004494 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004495)
4496
4497xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004498 name = "f32_vscaleexpminusmax_bench",
4499 srcs = [
4500 "bench/f32-vscaleexpminusmax.cc",
4501 "src/xnnpack/AlignedAllocator.h",
4502 ] + MICROKERNEL_BENCHMARK_HDRS,
4503 deps = MICROKERNEL_BENCHMARK_DEPS,
4504)
4505
4506xnnpack_benchmark(
4507 name = "f32_vscaleextexp_bench",
4508 srcs = [
4509 "bench/f32-vscaleextexp.cc",
4510 "src/xnnpack/AlignedAllocator.h",
4511 ] + MICROKERNEL_BENCHMARK_HDRS,
4512 deps = MICROKERNEL_BENCHMARK_DEPS,
4513)
4514
4515xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004516 name = "f32_vsqrt_bench",
4517 srcs = [
4518 "bench/f32-vsqrt.cc",
4519 "src/xnnpack/AlignedAllocator.h",
4520 ] + MICROKERNEL_BENCHMARK_HDRS,
4521 deps = MICROKERNEL_BENCHMARK_DEPS,
4522)
4523
4524xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004525 name = "f32_im2col_gemm_bench",
4526 srcs = [
4527 "bench/f32-im2col-gemm.cc",
4528 "bench/conv.h",
4529 "src/xnnpack/AlignedAllocator.h",
4530 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004531 deps = MICROKERNEL_BENCHMARK_DEPS + [
4532 ":im2col",
4533 ":packing",
4534 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004535)
4536
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004537xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004538 name = "rounding_bench",
4539 srcs = [
4540 "bench/rounding.cc",
4541 "src/xnnpack/math-stubs.h",
4542 "src/xnnpack/AlignedAllocator.h",
4543 ] + MICROKERNEL_BENCHMARK_HDRS,
4544 deps = MICROKERNEL_BENCHMARK_DEPS,
4545)
4546
Marat Dukhan08c4a432019-10-03 09:29:21 -07004547########################### Benchmarks for operators ###########################
4548
4549xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004550 name = "average_pooling_bench",
4551 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004552 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004553 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004554 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004555)
4556
4557xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004558 name = "bankers_rounding_bench",
4559 srcs = ["bench/bankers-rounding.cc"],
4560 copts = xnnpack_optional_tflite_copts(),
4561 tags = ["nowin32"],
4562 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4563)
4564
4565xnnpack_benchmark(
4566 name = "ceiling_bench",
4567 srcs = ["bench/ceiling.cc"],
4568 copts = xnnpack_optional_tflite_copts(),
4569 tags = ["nowin32"],
4570 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4571)
4572
4573xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004574 name = "channel_shuffle_bench",
4575 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004576 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004577)
4578
4579xnnpack_benchmark(
4580 name = "convolution_bench",
4581 srcs = ["bench/convolution.cc"],
4582 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004583 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004584 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004585)
4586
4587xnnpack_benchmark(
4588 name = "deconvolution_bench",
4589 srcs = ["bench/deconvolution.cc"],
4590 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004591 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004592 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004593)
4594
4595xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004596 name = "floor_bench",
4597 srcs = ["bench/floor.cc"],
4598 copts = xnnpack_optional_tflite_copts(),
4599 tags = ["nowin32"],
4600 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4601)
4602
4603xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004604 name = "global_average_pooling_bench",
4605 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004606 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004607)
4608
4609xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004610 name = "hardswish_bench",
4611 srcs = ["bench/hardswish.cc"],
4612 copts = xnnpack_optional_tflite_copts(),
4613 tags = ["nowin32"],
4614 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4615)
4616
4617xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004618 name = "max_pooling_bench",
4619 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004620 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004621)
4622
4623xnnpack_benchmark(
4624 name = "sigmoid_bench",
4625 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004626 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004627 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004628 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004629)
4630
4631xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004632 name = "prelu_bench",
4633 srcs = ["bench/prelu.cc"],
4634 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004635 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004636 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004637)
4638
4639xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004640 name = "softmax_bench",
4641 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004642 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004643 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004644 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004645)
4646
Marat Dukhan87727142020-06-24 15:24:10 -07004647xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004648 name = "square_root_bench",
4649 srcs = ["bench/square-root.cc"],
4650 copts = xnnpack_optional_tflite_copts(),
4651 tags = ["nowin32"],
4652 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4653)
4654
4655xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004656 name = "truncation_bench",
4657 srcs = ["bench/truncation.cc"],
4658 deps = OPERATOR_BENCHMARK_DEPS,
4659)
4660
Marat Dukhanc068bb62019-10-04 13:24:39 -07004661############################# End-to-end benchmarks ############################
4662
4663cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004664 name = "fp32_mobilenet_v1",
4665 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004666 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004667 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004668 linkstatic = True,
4669 deps = [
4670 ":XNNPACK",
4671 "@pthreadpool",
4672 ],
4673)
4674
4675cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004676 name = "fp16_mobilenet_v1",
4677 srcs = ["models/fp16-mobilenet-v1.cc"],
4678 hdrs = ["models/models.h"],
4679 copts = xnnpack_std_cxxopts(),
4680 linkstatic = True,
4681 deps = [
4682 ":XNNPACK",
4683 "@FP16",
4684 "@pthreadpool",
4685 ],
4686)
4687
4688cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004689 name = "qs8_mobilenet_v1",
4690 srcs = ["models/qs8-mobilenet-v1.cc"],
4691 hdrs = ["models/models.h"],
4692 copts = xnnpack_std_cxxopts(),
4693 linkstatic = True,
4694 deps = [
4695 ":XNNPACK",
4696 "@pthreadpool",
4697 ],
4698)
4699
4700cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004701 name = "qs8_mobilenet_v2",
4702 srcs = ["models/qs8-mobilenet-v2.cc"],
4703 hdrs = ["models/models.h"],
4704 copts = xnnpack_std_cxxopts(),
4705 linkstatic = True,
4706 deps = [
4707 ":XNNPACK",
4708 "@pthreadpool",
4709 ],
4710)
4711
4712cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004713 name = "fp32_mobilenet_v2",
4714 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004715 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004716 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004717 linkstatic = True,
4718 deps = [
4719 ":XNNPACK",
4720 "@pthreadpool",
4721 ],
4722)
4723
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004724cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004725 name = "fp16_mobilenet_v2",
4726 srcs = ["models/fp16-mobilenet-v2.cc"],
4727 hdrs = ["models/models.h"],
4728 copts = xnnpack_std_cxxopts(),
4729 linkstatic = True,
4730 deps = [
4731 ":XNNPACK",
4732 "@FP16",
4733 "@pthreadpool",
4734 ],
4735)
4736
4737cc_library(
4738 name = "fp32_mobilenet_v3_large",
4739 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004740 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004741 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004742 linkstatic = True,
4743 deps = [
4744 ":XNNPACK",
4745 "@pthreadpool",
4746 ],
4747)
4748
4749cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004750 name = "fp16_mobilenet_v3_large",
4751 srcs = ["models/fp16-mobilenet-v3-large.cc"],
4752 hdrs = ["models/models.h"],
4753 copts = xnnpack_std_cxxopts(),
4754 linkstatic = True,
4755 deps = [
4756 ":XNNPACK",
4757 "@FP16",
4758 "@pthreadpool",
4759 ],
4760)
4761
4762cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004763 name = "fp32_mobilenet_v3_small",
4764 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004765 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004766 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004767 linkstatic = True,
4768 deps = [
4769 ":XNNPACK",
4770 "@pthreadpool",
4771 ],
4772)
4773
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004774cc_library(
4775 name = "fp16_mobilenet_v3_small",
4776 srcs = ["models/fp16-mobilenet-v3-small.cc"],
4777 hdrs = ["models/models.h"],
4778 copts = xnnpack_std_cxxopts(),
4779 linkstatic = True,
4780 deps = [
4781 ":XNNPACK",
4782 "@FP16",
4783 "@pthreadpool",
4784 ],
4785)
4786
Marat Dukhanc068bb62019-10-04 13:24:39 -07004787xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004788 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004789 srcs = [
4790 "bench/f32-dwconv-e2e.cc",
4791 "bench/end2end.h",
4792 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004793 deps = MICROKERNEL_BENCHMARK_DEPS + [
4794 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004795 ":fp32_mobilenet_v1",
4796 ":fp32_mobilenet_v2",
4797 ":fp32_mobilenet_v3_large",
4798 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004799 ],
4800)
4801
4802xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004803 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004804 srcs = [
4805 "bench/f32-gemm-e2e.cc",
4806 "bench/end2end.h",
4807 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004808 deps = MICROKERNEL_BENCHMARK_DEPS + [
4809 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004810 ":fp32_mobilenet_v1",
4811 ":fp32_mobilenet_v2",
4812 ":fp32_mobilenet_v3_large",
4813 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004814 ],
4815)
4816
4817xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004818 name = "end2end_bench",
4819 srcs = ["bench/end2end.cc"],
4820 deps = [
4821 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004822 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004823 ":fp16_mobilenet_v1",
4824 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004825 ":fp16_mobilenet_v3_large",
4826 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004827 ":fp32_mobilenet_v1",
4828 ":fp32_mobilenet_v2",
4829 ":fp32_mobilenet_v3_large",
4830 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004831 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07004832 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004833 "@pthreadpool",
4834 ],
4835)
4836
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004837#################### Accuracy evaluation for math functions ####################
4838
4839xnnpack_benchmark(
4840 name = "f32_exp_eval",
4841 srcs = [
4842 "eval/f32-exp.cc",
4843 "src/xnnpack/AlignedAllocator.h",
4844 ] + ACCURACY_EVAL_HDRS,
4845 deps = ACCURACY_EVAL_DEPS,
4846)
4847
Marat Dukhan515c9772019-10-17 18:07:57 -07004848xnnpack_benchmark(
4849 name = "f32_expminus_eval",
4850 srcs = [
4851 "eval/f32-expminus.cc",
4852 "src/xnnpack/AlignedAllocator.h",
4853 ] + ACCURACY_EVAL_HDRS,
4854 deps = ACCURACY_EVAL_DEPS,
4855)
4856
Marat Dukhan98ba4412019-10-23 02:14:28 -07004857xnnpack_benchmark(
4858 name = "f32_extexp_eval",
4859 srcs = [
4860 "eval/f32-extexp.cc",
4861 "src/xnnpack/AlignedAllocator.h",
4862 ] + ACCURACY_EVAL_HDRS,
4863 deps = ACCURACY_EVAL_DEPS,
4864)
4865
Marat Dukhan8853b822020-05-07 12:19:01 -07004866xnnpack_unit_test(
4867 name = "f32_roundne_eval",
4868 srcs = [
4869 "eval/f32-roundne.cc",
4870 "src/xnnpack/AlignedAllocator.h",
4871 "src/xnnpack/math-stubs.h",
4872 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004873 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004874 deps = MICROKERNEL_TEST_DEPS,
4875)
4876
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004877xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004878 name = "f32_roundd_eval",
4879 srcs = [
4880 "eval/f32-roundd.cc",
4881 "src/xnnpack/AlignedAllocator.h",
4882 "src/xnnpack/math-stubs.h",
4883 ] + MICROKERNEL_TEST_HDRS,
4884 automatic = False,
4885 deps = MICROKERNEL_TEST_DEPS,
4886)
4887
4888xnnpack_unit_test(
4889 name = "f32_roundu_eval",
4890 srcs = [
4891 "eval/f32-roundu.cc",
4892 "src/xnnpack/AlignedAllocator.h",
4893 "src/xnnpack/math-stubs.h",
4894 ] + MICROKERNEL_TEST_HDRS,
4895 automatic = False,
4896 deps = MICROKERNEL_TEST_DEPS,
4897)
4898
4899xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004900 name = "f32_roundz_eval",
4901 srcs = [
4902 "eval/f32-roundz.cc",
4903 "src/xnnpack/AlignedAllocator.h",
4904 "src/xnnpack/math-stubs.h",
4905 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004906 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004907 deps = MICROKERNEL_TEST_DEPS,
4908)
4909
Marat Dukhan346a9e52019-11-15 09:06:30 -08004910xnnpack_benchmark(
4911 name = "f32_sigmoid_eval",
4912 srcs = [
4913 "eval/f32-sigmoid.cc",
4914 "src/xnnpack/AlignedAllocator.h",
4915 ] + ACCURACY_EVAL_HDRS,
4916 deps = ACCURACY_EVAL_DEPS,
4917)
4918
Marat Dukhan84000762020-06-29 18:38:43 -07004919xnnpack_benchmark(
4920 name = "f32_sqrt_eval",
4921 srcs = [
4922 "eval/f32-sqrt.cc",
4923 "src/xnnpack/AlignedAllocator.h",
4924 ] + ACCURACY_EVAL_HDRS,
4925 deps = ACCURACY_EVAL_DEPS + [
4926 ":bench_utils",
4927 ],
4928)
4929
Marat Dukhan08c4a432019-10-03 09:29:21 -07004930######################### Unit tests for micro-kernels #########################
4931
4932xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004933 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004934 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004935 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004936 "test/gemm-microkernel-tester.h",
4937 "src/xnnpack/AlignedAllocator.h",
4938 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004939 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004940)
4941
4942xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004943 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004944 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004945 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004946 "test/spmm-microkernel-tester.h",
4947 "src/xnnpack/AlignedAllocator.h",
4948 ] + MICROKERNEL_TEST_HDRS,
4949 deps = MICROKERNEL_TEST_DEPS,
4950)
4951
4952xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004953 name = "f32_argmaxpool_test",
4954 srcs = [
4955 "test/f32-argmaxpool.cc",
4956 "test/argmaxpool-microkernel-tester.h",
4957 "src/xnnpack/AlignedAllocator.h",
4958 ] + MICROKERNEL_TEST_HDRS,
4959 deps = MICROKERNEL_TEST_DEPS,
4960)
4961
4962xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004963 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004964 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004965 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004966 "test/avgpool-microkernel-tester.h",
4967 "src/xnnpack/AlignedAllocator.h",
4968 ] + MICROKERNEL_TEST_HDRS,
4969 deps = MICROKERNEL_TEST_DEPS,
4970)
4971
4972xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004973 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004974 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004975 "test/f32-ibilinear.cc",
4976 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004977 "src/xnnpack/AlignedAllocator.h",
4978 ] + MICROKERNEL_TEST_HDRS,
4979 deps = MICROKERNEL_TEST_DEPS,
4980)
4981
4982xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004983 name = "f16_clamp_test",
4984 srcs = [
4985 "test/f16-clamp.cc",
4986 "test/clamp-microkernel-tester.h",
4987 ] + MICROKERNEL_TEST_HDRS,
4988 deps = MICROKERNEL_TEST_DEPS,
4989)
4990
4991xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004992 name = "f32_clamp_test",
4993 srcs = [
4994 "test/f32-clamp.cc",
4995 "test/clamp-microkernel-tester.h",
4996 ] + MICROKERNEL_TEST_HDRS,
4997 deps = MICROKERNEL_TEST_DEPS,
4998)
4999
5000xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005001 name = "f32_igemm_test",
5002 srcs = [
5003 "test/f32-igemm.cc",
5004 "test/gemm-microkernel-tester.h",
5005 "src/xnnpack/AlignedAllocator.h",
5006 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005007 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005008)
5009
5010xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005011 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005012 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005013 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005014 "test/gemm-microkernel-tester.h",
5015 "src/xnnpack/AlignedAllocator.h",
5016 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005017 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005018)
5019
5020xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005021 name = "f32_igemm_minmax_test",
5022 srcs = [
5023 "test/f32-igemm-minmax.cc",
5024 "test/gemm-microkernel-tester.h",
5025 "src/xnnpack/AlignedAllocator.h",
5026 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005027 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005028)
5029
5030xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005031 name = "f16_igemm_minmax_test",
5032 srcs = [
5033 "test/f16-igemm-minmax.cc",
5034 "test/gemm-microkernel-tester.h",
5035 "src/xnnpack/AlignedAllocator.h",
5036 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005037 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005038)
5039
5040xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005041 name = "f32_conv_hwc_test",
5042 srcs = [
5043 "test/f32-conv-hwc.cc",
5044 "test/conv-hwc-microkernel-tester.h",
5045 "src/xnnpack/AlignedAllocator.h",
5046 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005047 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005048)
5049
5050xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005051 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005052 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005053 "test/f32-conv-hwc2chw.cc",
5054 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005055 "src/xnnpack/AlignedAllocator.h",
5056 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005057 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005058)
5059
5060xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005061 name = "f32_dwconv_test",
5062 srcs = [
5063 "test/f32-dwconv.cc",
5064 "test/dwconv-microkernel-tester.h",
5065 "src/xnnpack/AlignedAllocator.h",
5066 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005067 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005068)
5069
5070xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005071 name = "f16_dwconv_minmax_test",
5072 srcs = [
5073 "test/f16-dwconv-minmax.cc",
5074 "test/dwconv-microkernel-tester.h",
5075 "src/xnnpack/AlignedAllocator.h",
5076 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005077 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005078)
5079
5080xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005081 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005082 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005083 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005084 "test/dwconv-microkernel-tester.h",
5085 "src/xnnpack/AlignedAllocator.h",
5086 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005087 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005088)
5089
5090xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005091 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005092 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005093 "test/f32-dwconv-chw.cc",
5094 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005095 "src/xnnpack/AlignedAllocator.h",
5096 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005097 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005098)
5099
5100xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005101 name = "f16_gavgpool_minmax_test",
5102 srcs = [
5103 "test/f16-gavgpool-minmax.cc",
5104 "test/gavgpool-microkernel-tester.h",
5105 "src/xnnpack/AlignedAllocator.h",
5106 ] + MICROKERNEL_TEST_HDRS,
5107 deps = MICROKERNEL_TEST_DEPS,
5108)
5109
5110xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005111 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005112 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005113 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005114 "test/gavgpool-microkernel-tester.h",
5115 "src/xnnpack/AlignedAllocator.h",
5116 ] + MICROKERNEL_TEST_HDRS,
5117 deps = MICROKERNEL_TEST_DEPS,
5118)
5119
5120xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005121 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005122 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005123 "test/f32-gavgpool-cw.cc",
5124 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005125 "src/xnnpack/AlignedAllocator.h",
5126 ] + MICROKERNEL_TEST_HDRS,
5127 deps = MICROKERNEL_TEST_DEPS,
5128)
5129
5130xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005131 name = "f32_gemm_test",
5132 srcs = [
5133 "test/f32-gemm.cc",
5134 "test/gemm-microkernel-tester.h",
5135 "src/xnnpack/AlignedAllocator.h",
5136 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005137 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005138)
5139
5140xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005141 name = "f32_gemm_relu_test",
5142 srcs = [
5143 "test/f32-gemm-relu.cc",
5144 "test/gemm-microkernel-tester.h",
5145 "src/xnnpack/AlignedAllocator.h",
5146 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005147 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005148)
5149
5150xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005151 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005152 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005153 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005154 "test/gemm-microkernel-tester.h",
5155 "src/xnnpack/AlignedAllocator.h",
5156 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005157 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005158)
5159
5160xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005161 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005162 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005163 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005164 "test/gemm-microkernel-tester.h",
5165 "src/xnnpack/AlignedAllocator.h",
5166 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005167 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005168)
5169
5170xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005171 name = "f16_hswish_test",
5172 srcs = [
5173 "test/f16-hswish.cc",
5174 "test/hswish-microkernel-tester.h",
5175 ] + MICROKERNEL_TEST_HDRS,
5176 deps = MICROKERNEL_TEST_DEPS,
5177)
5178
5179xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005180 name = "f32_hswish_test",
5181 srcs = [
5182 "test/f32-hswish.cc",
5183 "test/hswish-microkernel-tester.h",
5184 ] + MICROKERNEL_TEST_HDRS,
5185 deps = MICROKERNEL_TEST_DEPS,
5186)
5187
5188xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005189 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005190 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005191 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005192 "test/maxpool-microkernel-tester.h",
5193 ] + MICROKERNEL_TEST_HDRS,
5194 deps = MICROKERNEL_TEST_DEPS,
5195)
5196
5197xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005198 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005199 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005200 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005201 "test/avgpool-microkernel-tester.h",
5202 "src/xnnpack/AlignedAllocator.h",
5203 ] + MICROKERNEL_TEST_HDRS,
5204 deps = MICROKERNEL_TEST_DEPS,
5205)
5206
5207xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005208 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005209 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005210 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005211 "test/gemm-microkernel-tester.h",
5212 "src/xnnpack/AlignedAllocator.h",
5213 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005214 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005215)
5216
5217xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005218 name = "f16_prelu_test",
5219 srcs = [
5220 "test/f16-prelu.cc",
5221 "test/prelu-microkernel-tester.h",
5222 "src/xnnpack/AlignedAllocator.h",
5223 ] + MICROKERNEL_TEST_HDRS,
5224 deps = MICROKERNEL_TEST_DEPS,
5225)
5226
5227xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005228 name = "f32_prelu_test",
5229 srcs = [
5230 "test/f32-prelu.cc",
5231 "test/prelu-microkernel-tester.h",
5232 "src/xnnpack/AlignedAllocator.h",
5233 ] + MICROKERNEL_TEST_HDRS,
5234 deps = MICROKERNEL_TEST_DEPS,
5235)
5236
5237xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005238 name = "f32_raddexpminusmax_test",
5239 srcs = [
5240 "test/f32-raddexpminusmax.cc",
5241 "test/raddexpminusmax-microkernel-tester.h",
5242 ] + MICROKERNEL_TEST_HDRS,
5243 deps = MICROKERNEL_TEST_DEPS,
5244)
5245
5246xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005247 name = "f32_raddextexp_test",
5248 srcs = [
5249 "test/f32-raddextexp.cc",
5250 "test/raddextexp-microkernel-tester.h",
5251 ] + MICROKERNEL_TEST_HDRS,
5252 deps = MICROKERNEL_TEST_DEPS,
5253)
5254
5255xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005256 name = "f32_raddstoreexpminusmax_test",
5257 srcs = [
5258 "test/f32-raddstoreexpminusmax.cc",
5259 "test/raddstoreexpminusmax-microkernel-tester.h",
5260 ] + MICROKERNEL_TEST_HDRS,
5261 deps = MICROKERNEL_TEST_DEPS,
5262)
5263
5264xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005265 name = "f32_relu_test",
5266 srcs = [
5267 "test/f32-relu.cc",
5268 "test/vunary-microkernel-tester.h",
5269 ] + MICROKERNEL_TEST_HDRS,
5270 deps = MICROKERNEL_TEST_DEPS,
5271)
5272
5273xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005274 name = "f32_rmax_test",
5275 srcs = [
5276 "test/f32-rmax.cc",
5277 "test/rmax-microkernel-tester.h",
5278 ] + MICROKERNEL_TEST_HDRS,
5279 deps = MICROKERNEL_TEST_DEPS,
5280)
5281
5282xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005283 name = "f32_sigmoid_test",
5284 srcs = [
5285 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005286 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005287 ] + MICROKERNEL_TEST_HDRS,
5288 deps = MICROKERNEL_TEST_DEPS,
5289)
5290
5291xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005292 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005293 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005294 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005295 "test/spmm-microkernel-tester.h",
5296 "src/xnnpack/AlignedAllocator.h",
5297 ] + MICROKERNEL_TEST_HDRS,
5298 deps = MICROKERNEL_TEST_DEPS,
5299)
5300
5301xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005302 name = "f16_vadd_minmax_test",
5303 srcs = [
5304 "test/f16-vadd-minmax.cc",
5305 "test/vbinary-microkernel-tester.h",
5306 ] + MICROKERNEL_TEST_HDRS,
5307 deps = MICROKERNEL_TEST_DEPS,
5308)
5309
5310xnnpack_unit_test(
5311 name = "f16_vaddc_minmax_test",
5312 srcs = [
5313 "test/f16-vaddc-minmax.cc",
5314 "test/vbinaryc-microkernel-tester.h",
5315 ] + MICROKERNEL_TEST_HDRS,
5316 deps = MICROKERNEL_TEST_DEPS,
5317)
5318
5319xnnpack_unit_test(
5320 name = "f16_vdiv_minmax_test",
5321 srcs = [
5322 "test/f16-vdiv-minmax.cc",
5323 "test/vbinary-microkernel-tester.h",
5324 ] + MICROKERNEL_TEST_HDRS,
5325 deps = MICROKERNEL_TEST_DEPS,
5326)
5327
5328xnnpack_unit_test(
5329 name = "f16_vdivc_minmax_test",
5330 srcs = [
5331 "test/f16-vdivc-minmax.cc",
5332 "test/vbinaryc-microkernel-tester.h",
5333 ] + MICROKERNEL_TEST_HDRS,
5334 deps = MICROKERNEL_TEST_DEPS,
5335)
5336
5337xnnpack_unit_test(
5338 name = "f16_vrdivc_minmax_test",
5339 srcs = [
5340 "test/f16-vrdivc-minmax.cc",
5341 "test/vbinaryc-microkernel-tester.h",
5342 ] + MICROKERNEL_TEST_HDRS,
5343 deps = MICROKERNEL_TEST_DEPS,
5344)
5345
5346xnnpack_unit_test(
5347 name = "f16_vmax_test",
5348 srcs = [
5349 "test/f16-vmax.cc",
5350 "test/vbinary-microkernel-tester.h",
5351 ] + MICROKERNEL_TEST_HDRS,
5352 deps = MICROKERNEL_TEST_DEPS,
5353)
5354
5355xnnpack_unit_test(
5356 name = "f16_vmaxc_test",
5357 srcs = [
5358 "test/f16-vmaxc.cc",
5359 "test/vbinaryc-microkernel-tester.h",
5360 ] + MICROKERNEL_TEST_HDRS,
5361 deps = MICROKERNEL_TEST_DEPS,
5362)
5363
5364xnnpack_unit_test(
5365 name = "f16_vmin_test",
5366 srcs = [
5367 "test/f16-vmin.cc",
5368 "test/vbinary-microkernel-tester.h",
5369 ] + MICROKERNEL_TEST_HDRS,
5370 deps = MICROKERNEL_TEST_DEPS,
5371)
5372
5373xnnpack_unit_test(
5374 name = "f16_vminc_test",
5375 srcs = [
5376 "test/f16-vminc.cc",
5377 "test/vbinaryc-microkernel-tester.h",
5378 ] + MICROKERNEL_TEST_HDRS,
5379 deps = MICROKERNEL_TEST_DEPS,
5380)
5381
5382xnnpack_unit_test(
5383 name = "f16_vmul_minmax_test",
5384 srcs = [
5385 "test/f16-vmul-minmax.cc",
5386 "test/vbinary-microkernel-tester.h",
5387 ] + MICROKERNEL_TEST_HDRS,
5388 deps = MICROKERNEL_TEST_DEPS,
5389)
5390
5391xnnpack_unit_test(
5392 name = "f16_vmulc_minmax_test",
5393 srcs = [
5394 "test/f16-vmulc-minmax.cc",
5395 "test/vbinaryc-microkernel-tester.h",
5396 ] + MICROKERNEL_TEST_HDRS,
5397 deps = MICROKERNEL_TEST_DEPS,
5398)
5399
5400xnnpack_unit_test(
5401 name = "f16_vsub_minmax_test",
5402 srcs = [
5403 "test/f16-vsub-minmax.cc",
5404 "test/vbinary-microkernel-tester.h",
5405 ] + MICROKERNEL_TEST_HDRS,
5406 deps = MICROKERNEL_TEST_DEPS,
5407)
5408
5409xnnpack_unit_test(
5410 name = "f16_vsubc_minmax_test",
5411 srcs = [
5412 "test/f16-vsubc-minmax.cc",
5413 "test/vbinaryc-microkernel-tester.h",
5414 ] + MICROKERNEL_TEST_HDRS,
5415 deps = MICROKERNEL_TEST_DEPS,
5416)
5417
5418xnnpack_unit_test(
5419 name = "f16_vrsubc_minmax_test",
5420 srcs = [
5421 "test/f16-vrsubc-minmax.cc",
5422 "test/vbinaryc-microkernel-tester.h",
5423 ] + MICROKERNEL_TEST_HDRS,
5424 deps = MICROKERNEL_TEST_DEPS,
5425)
5426
5427xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005428 name = "f32_vabs_test",
5429 srcs = [
5430 "test/f32-vabs.cc",
5431 "test/vunary-microkernel-tester.h",
5432 ] + MICROKERNEL_TEST_HDRS,
5433 deps = MICROKERNEL_TEST_DEPS,
5434)
5435
5436xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005437 name = "f32_vadd_test",
5438 srcs = [
5439 "test/f32-vadd.cc",
5440 "test/vbinary-microkernel-tester.h",
5441 ] + MICROKERNEL_TEST_HDRS,
5442 deps = MICROKERNEL_TEST_DEPS,
5443)
5444
5445xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005446 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005447 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005448 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005449 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005450 ] + MICROKERNEL_TEST_HDRS,
5451 deps = MICROKERNEL_TEST_DEPS,
5452)
5453
5454xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005455 name = "f32_vadd_relu_test",
5456 srcs = [
5457 "test/f32-vadd-relu.cc",
5458 "test/vbinary-microkernel-tester.h",
5459 ] + MICROKERNEL_TEST_HDRS,
5460 deps = MICROKERNEL_TEST_DEPS,
5461)
5462
5463xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005464 name = "f32_vaddc_test",
5465 srcs = [
5466 "test/f32-vaddc.cc",
5467 "test/vbinaryc-microkernel-tester.h",
5468 ] + MICROKERNEL_TEST_HDRS,
5469 deps = MICROKERNEL_TEST_DEPS,
5470)
5471
5472xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005473 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005474 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005475 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005476 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005477 ] + MICROKERNEL_TEST_HDRS,
5478 deps = MICROKERNEL_TEST_DEPS,
5479)
5480
5481xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005482 name = "f32_vaddc_relu_test",
5483 srcs = [
5484 "test/f32-vaddc-relu.cc",
5485 "test/vbinaryc-microkernel-tester.h",
5486 ] + MICROKERNEL_TEST_HDRS,
5487 deps = MICROKERNEL_TEST_DEPS,
5488)
5489
5490xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005491 name = "f32_vdiv_test",
5492 srcs = [
5493 "test/f32-vdiv.cc",
5494 "test/vbinary-microkernel-tester.h",
5495 ] + MICROKERNEL_TEST_HDRS,
5496 deps = MICROKERNEL_TEST_DEPS,
5497)
5498
5499xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005500 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005501 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005502 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005503 "test/vbinary-microkernel-tester.h",
5504 ] + MICROKERNEL_TEST_HDRS,
5505 deps = MICROKERNEL_TEST_DEPS,
5506)
5507
5508xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005509 name = "f32_vdiv_relu_test",
5510 srcs = [
5511 "test/f32-vdiv-relu.cc",
5512 "test/vbinary-microkernel-tester.h",
5513 ] + MICROKERNEL_TEST_HDRS,
5514 deps = MICROKERNEL_TEST_DEPS,
5515)
5516
5517xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005518 name = "f32_vdivc_test",
5519 srcs = [
5520 "test/f32-vdivc.cc",
5521 "test/vbinaryc-microkernel-tester.h",
5522 ] + MICROKERNEL_TEST_HDRS,
5523 deps = MICROKERNEL_TEST_DEPS,
5524)
5525
5526xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005527 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005528 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005529 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005530 "test/vbinaryc-microkernel-tester.h",
5531 ] + MICROKERNEL_TEST_HDRS,
5532 deps = MICROKERNEL_TEST_DEPS,
5533)
5534
5535xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005536 name = "f32_vdivc_relu_test",
5537 srcs = [
5538 "test/f32-vdivc-relu.cc",
5539 "test/vbinaryc-microkernel-tester.h",
5540 ] + MICROKERNEL_TEST_HDRS,
5541 deps = MICROKERNEL_TEST_DEPS,
5542)
5543
5544xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005545 name = "f32_vrdivc_test",
5546 srcs = [
5547 "test/f32-vrdivc.cc",
5548 "test/vbinaryc-microkernel-tester.h",
5549 ] + MICROKERNEL_TEST_HDRS,
5550 deps = MICROKERNEL_TEST_DEPS,
5551)
5552
5553xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005554 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005555 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005556 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005557 "test/vbinaryc-microkernel-tester.h",
5558 ] + MICROKERNEL_TEST_HDRS,
5559 deps = MICROKERNEL_TEST_DEPS,
5560)
5561
5562xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005563 name = "f32_vrdivc_relu_test",
5564 srcs = [
5565 "test/f32-vrdivc-relu.cc",
5566 "test/vbinaryc-microkernel-tester.h",
5567 ] + MICROKERNEL_TEST_HDRS,
5568 deps = MICROKERNEL_TEST_DEPS,
5569)
5570
5571xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005572 name = "f32_vmax_test",
5573 srcs = [
5574 "test/f32-vmax.cc",
5575 "test/vbinary-microkernel-tester.h",
5576 ] + MICROKERNEL_TEST_HDRS,
5577 deps = MICROKERNEL_TEST_DEPS,
5578)
5579
5580xnnpack_unit_test(
5581 name = "f32_vmaxc_test",
5582 srcs = [
5583 "test/f32-vmaxc.cc",
5584 "test/vbinaryc-microkernel-tester.h",
5585 ] + MICROKERNEL_TEST_HDRS,
5586 deps = MICROKERNEL_TEST_DEPS,
5587)
5588
5589xnnpack_unit_test(
5590 name = "f32_vmin_test",
5591 srcs = [
5592 "test/f32-vmin.cc",
5593 "test/vbinary-microkernel-tester.h",
5594 ] + MICROKERNEL_TEST_HDRS,
5595 deps = MICROKERNEL_TEST_DEPS,
5596)
5597
5598xnnpack_unit_test(
5599 name = "f32_vminc_test",
5600 srcs = [
5601 "test/f32-vminc.cc",
5602 "test/vbinaryc-microkernel-tester.h",
5603 ] + MICROKERNEL_TEST_HDRS,
5604 deps = MICROKERNEL_TEST_DEPS,
5605)
5606
5607xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005608 name = "f32_vmul_test",
5609 srcs = [
5610 "test/f32-vmul.cc",
5611 "test/vbinary-microkernel-tester.h",
5612 ] + MICROKERNEL_TEST_HDRS,
5613 deps = MICROKERNEL_TEST_DEPS,
5614)
5615
5616xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005617 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005618 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005619 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005620 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005621 ] + MICROKERNEL_TEST_HDRS,
5622 deps = MICROKERNEL_TEST_DEPS,
5623)
5624
5625xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005626 name = "f32_vmul_relu_test",
5627 srcs = [
5628 "test/f32-vmul-relu.cc",
5629 "test/vbinary-microkernel-tester.h",
5630 ] + MICROKERNEL_TEST_HDRS,
5631 deps = MICROKERNEL_TEST_DEPS,
5632)
5633
5634xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005635 name = "f32_vmulc_test",
5636 srcs = [
5637 "test/f32-vmulc.cc",
5638 "test/vbinaryc-microkernel-tester.h",
5639 ] + MICROKERNEL_TEST_HDRS,
5640 deps = MICROKERNEL_TEST_DEPS,
5641)
5642
5643xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005644 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005645 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005646 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005647 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005648 ] + MICROKERNEL_TEST_HDRS,
5649 deps = MICROKERNEL_TEST_DEPS,
5650)
5651
5652xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005653 name = "f32_vmulc_relu_test",
5654 srcs = [
5655 "test/f32-vmulc-relu.cc",
5656 "test/vbinaryc-microkernel-tester.h",
5657 ] + MICROKERNEL_TEST_HDRS,
5658 deps = MICROKERNEL_TEST_DEPS,
5659)
5660
5661xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005662 name = "f16_vmulcaddc_minmax_test",
5663 srcs = [
5664 "test/f16-vmulcaddc-minmax.cc",
5665 "test/vmulcaddc-microkernel-tester.h",
5666 "src/xnnpack/AlignedAllocator.h",
5667 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005668 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005669)
5670
5671xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005672 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005673 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005674 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005675 "test/vmulcaddc-microkernel-tester.h",
5676 "src/xnnpack/AlignedAllocator.h",
5677 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005678 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005679)
5680
5681xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005682 name = "f32_vlrelu_test",
5683 srcs = [
5684 "test/f32-vlrelu.cc",
5685 "test/vunary-microkernel-tester.h",
5686 ] + MICROKERNEL_TEST_HDRS,
5687 deps = MICROKERNEL_TEST_DEPS,
5688)
5689
5690xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005691 name = "f32_vneg_test",
5692 srcs = [
5693 "test/f32-vneg.cc",
5694 "test/vunary-microkernel-tester.h",
5695 ] + MICROKERNEL_TEST_HDRS,
5696 deps = MICROKERNEL_TEST_DEPS,
5697)
5698
5699xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005700 name = "f32_vrndne_test",
5701 srcs = [
5702 "test/f32-vrndne.cc",
5703 "test/vunary-microkernel-tester.h",
5704 ] + MICROKERNEL_TEST_HDRS,
5705 deps = MICROKERNEL_TEST_DEPS,
5706)
5707
5708xnnpack_unit_test(
5709 name = "f32_vrndz_test",
5710 srcs = [
5711 "test/f32-vrndz.cc",
5712 "test/vunary-microkernel-tester.h",
5713 ] + MICROKERNEL_TEST_HDRS,
5714 deps = MICROKERNEL_TEST_DEPS,
5715)
5716
5717xnnpack_unit_test(
5718 name = "f32_vrndu_test",
5719 srcs = [
5720 "test/f32-vrndu.cc",
5721 "test/vunary-microkernel-tester.h",
5722 ] + MICROKERNEL_TEST_HDRS,
5723 deps = MICROKERNEL_TEST_DEPS,
5724)
5725
5726xnnpack_unit_test(
5727 name = "f32_vrndd_test",
5728 srcs = [
5729 "test/f32-vrndd.cc",
5730 "test/vunary-microkernel-tester.h",
5731 ] + MICROKERNEL_TEST_HDRS,
5732 deps = MICROKERNEL_TEST_DEPS,
5733)
5734
5735xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005736 name = "f32_vscale_test",
5737 srcs = [
5738 "test/f32-vscale.cc",
5739 "test/vscale-microkernel-tester.h",
5740 ] + MICROKERNEL_TEST_HDRS,
5741 deps = MICROKERNEL_TEST_DEPS,
5742)
5743
5744xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005745 name = "f32_vscaleexpminusmax_test",
5746 srcs = [
5747 "test/f32-vscaleexpminusmax.cc",
5748 "test/vscaleexpminusmax-microkernel-tester.h",
5749 ] + MICROKERNEL_TEST_HDRS,
5750 deps = MICROKERNEL_TEST_DEPS,
5751)
5752
5753xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005754 name = "f32_vscaleextexp_test",
5755 srcs = [
5756 "test/f32-vscaleextexp.cc",
5757 "test/vscaleextexp-microkernel-tester.h",
5758 ] + MICROKERNEL_TEST_HDRS,
5759 deps = MICROKERNEL_TEST_DEPS,
5760)
5761
5762xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005763 name = "f32_vsqr_test",
5764 srcs = [
5765 "test/f32-vsqr.cc",
5766 "test/vunary-microkernel-tester.h",
5767 ] + MICROKERNEL_TEST_HDRS,
5768 deps = MICROKERNEL_TEST_DEPS,
5769)
5770
5771xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005772 name = "f32_vsqrdiff_test",
5773 srcs = [
5774 "test/f32-vsqrdiff.cc",
5775 "test/vbinary-microkernel-tester.h",
5776 ] + MICROKERNEL_TEST_HDRS,
5777 deps = MICROKERNEL_TEST_DEPS,
5778)
5779
5780xnnpack_unit_test(
5781 name = "f32_vsqrdiffc_test",
5782 srcs = [
5783 "test/f32-vsqrdiffc.cc",
5784 "test/vbinaryc-microkernel-tester.h",
5785 ] + MICROKERNEL_TEST_HDRS,
5786 deps = MICROKERNEL_TEST_DEPS,
5787)
5788
5789xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005790 name = "f32_vsqrt_test",
5791 srcs = [
5792 "test/f32-vsqrt.cc",
5793 "test/vunary-microkernel-tester.h",
5794 ] + MICROKERNEL_TEST_HDRS,
5795 deps = MICROKERNEL_TEST_DEPS,
5796)
5797
5798xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005799 name = "f32_vsub_test",
5800 srcs = [
5801 "test/f32-vsub.cc",
5802 "test/vbinary-microkernel-tester.h",
5803 ] + MICROKERNEL_TEST_HDRS,
5804 deps = MICROKERNEL_TEST_DEPS,
5805)
5806
5807xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005808 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005809 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005810 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005811 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005812 ] + MICROKERNEL_TEST_HDRS,
5813 deps = MICROKERNEL_TEST_DEPS,
5814)
5815
5816xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005817 name = "f32_vsub_relu_test",
5818 srcs = [
5819 "test/f32-vsub-relu.cc",
5820 "test/vbinary-microkernel-tester.h",
5821 ] + MICROKERNEL_TEST_HDRS,
5822 deps = MICROKERNEL_TEST_DEPS,
5823)
5824
5825xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005826 name = "f32_vsubc_test",
5827 srcs = [
5828 "test/f32-vsubc.cc",
5829 "test/vbinaryc-microkernel-tester.h",
5830 ] + MICROKERNEL_TEST_HDRS,
5831 deps = MICROKERNEL_TEST_DEPS,
5832)
5833
5834xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005835 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005836 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005837 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005838 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005839 ] + MICROKERNEL_TEST_HDRS,
5840 deps = MICROKERNEL_TEST_DEPS,
5841)
5842
5843xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005844 name = "f32_vsubc_relu_test",
5845 srcs = [
5846 "test/f32-vsubc-relu.cc",
5847 "test/vbinaryc-microkernel-tester.h",
5848 ] + MICROKERNEL_TEST_HDRS,
5849 deps = MICROKERNEL_TEST_DEPS,
5850)
5851
5852xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005853 name = "f32_vrsubc_test",
5854 srcs = [
5855 "test/f32-vrsubc.cc",
5856 "test/vbinaryc-microkernel-tester.h",
5857 ] + MICROKERNEL_TEST_HDRS,
5858 deps = MICROKERNEL_TEST_DEPS,
5859)
5860
5861xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005862 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005863 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005864 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005865 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005866 ] + MICROKERNEL_TEST_HDRS,
5867 deps = MICROKERNEL_TEST_DEPS,
5868)
5869
5870xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005871 name = "f32_vrsubc_relu_test",
5872 srcs = [
5873 "test/f32-vrsubc-relu.cc",
5874 "test/vbinaryc-microkernel-tester.h",
5875 ] + MICROKERNEL_TEST_HDRS,
5876 deps = MICROKERNEL_TEST_DEPS,
5877)
5878
5879xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005880 name = "qs8_dwconv_minmax_test",
5881 srcs = [
5882 "test/qs8-dwconv-minmax.cc",
5883 "test/dwconv-microkernel-tester.h",
5884 "src/xnnpack/AlignedAllocator.h",
5885 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5886 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5887)
5888
5889xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005890 name = "qs8_gavgpool_minmax_test",
5891 srcs = [
5892 "test/qs8-gavgpool-minmax.cc",
5893 "test/gavgpool-microkernel-tester.h",
5894 "src/xnnpack/AlignedAllocator.h",
5895 ] + MICROKERNEL_TEST_HDRS,
5896 deps = MICROKERNEL_TEST_DEPS,
5897)
5898
5899xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005900 name = "qs8_gemm_minmax_test",
5901 srcs = [
5902 "test/qs8-gemm-minmax.cc",
5903 "test/gemm-microkernel-tester.h",
5904 "src/xnnpack/AlignedAllocator.h",
5905 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5906 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5907)
5908
5909xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005910 name = "qs8_igemm_minmax_test",
5911 srcs = [
5912 "test/qs8-igemm-minmax.cc",
5913 "test/gemm-microkernel-tester.h",
5914 "src/xnnpack/AlignedAllocator.h",
5915 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5916 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5917)
5918
5919xnnpack_unit_test(
5920 name = "qs8_requantization_test",
5921 srcs = [
5922 "src/xnnpack/requantization-stubs.h",
5923 "test/qs8-requantization.cc",
5924 "test/requantization-tester.h",
5925 ] + MICROKERNEL_TEST_HDRS,
5926 deps = MICROKERNEL_TEST_DEPS,
5927)
5928
5929xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07005930 name = "qs8_vadd_minmax_test",
5931 srcs = [
5932 "test/qs8-vadd-minmax.cc",
5933 "test/vadd-microkernel-tester.h",
5934 ] + MICROKERNEL_TEST_HDRS,
5935 deps = MICROKERNEL_TEST_DEPS,
5936)
5937
5938xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07005939 name = "qs8_vaddc_minmax_test",
5940 srcs = [
5941 "test/qs8-vaddc-minmax.cc",
5942 "test/vaddc-microkernel-tester.h",
5943 ] + MICROKERNEL_TEST_HDRS,
5944 deps = MICROKERNEL_TEST_DEPS,
5945)
5946
5947xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005948 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005949 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005950 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005951 "test/avgpool-microkernel-tester.h",
5952 "src/xnnpack/AlignedAllocator.h",
5953 ] + MICROKERNEL_TEST_HDRS,
5954 deps = MICROKERNEL_TEST_DEPS,
5955)
5956
5957xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005958 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005959 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005960 "test/qu8-dwconv-minmax.cc",
5961 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005962 "src/xnnpack/AlignedAllocator.h",
5963 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005964 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005965)
5966
5967xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005968 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005969 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005970 "test/qu8-igemm-minmax.cc",
5971 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005972 "src/xnnpack/AlignedAllocator.h",
5973 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005974 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005975)
5976
5977xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005978 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005979 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005980 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005981 "test/gavgpool-microkernel-tester.h",
5982 "src/xnnpack/AlignedAllocator.h",
5983 ] + MICROKERNEL_TEST_HDRS,
5984 deps = MICROKERNEL_TEST_DEPS,
5985)
5986
5987xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005988 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005989 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005990 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005991 "test/gemm-microkernel-tester.h",
5992 "src/xnnpack/AlignedAllocator.h",
5993 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005994 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005995)
5996
5997xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005998 name = "qu8_requantization_test",
5999 srcs = [
6000 "src/xnnpack/requantization-stubs.h",
6001 "test/qu8-requantization.cc",
6002 "test/requantization-tester.h",
6003 ] + MICROKERNEL_TEST_HDRS,
6004 deps = MICROKERNEL_TEST_DEPS,
6005)
6006
6007xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006008 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006009 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006010 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006011 "test/vadd-microkernel-tester.h",
6012 ] + MICROKERNEL_TEST_HDRS,
6013 deps = MICROKERNEL_TEST_DEPS,
6014)
6015
6016xnnpack_unit_test(
6017 name = "u8_clamp_test",
6018 srcs = [
6019 "test/u8-clamp.cc",
6020 "test/clamp-microkernel-tester.h",
6021 ] + MICROKERNEL_TEST_HDRS,
6022 deps = MICROKERNEL_TEST_DEPS,
6023)
6024
6025xnnpack_unit_test(
6026 name = "u8_lut32norm_test",
6027 srcs = [
6028 "test/u8-lut32norm.cc",
6029 "test/lut-norm-microkernel-tester.h",
6030 ] + MICROKERNEL_TEST_HDRS,
6031 deps = MICROKERNEL_TEST_DEPS,
6032)
6033
6034xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006035 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006036 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006037 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006038 "test/maxpool-microkernel-tester.h",
6039 ] + MICROKERNEL_TEST_HDRS,
6040 deps = MICROKERNEL_TEST_DEPS,
6041)
6042
6043xnnpack_unit_test(
6044 name = "u8_rmax_test",
6045 srcs = [
6046 "test/u8-rmax.cc",
6047 "test/rmax-microkernel-tester.h",
6048 ] + MICROKERNEL_TEST_HDRS,
6049 deps = MICROKERNEL_TEST_DEPS,
6050)
6051
6052xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006053 name = "x32_fill_test",
6054 srcs = [
6055 "test/x32-fill.cc",
6056 "test/fill-microkernel-tester.h",
6057 ] + MICROKERNEL_TEST_HDRS,
6058 deps = MICROKERNEL_TEST_DEPS,
6059)
6060
6061xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006062 name = "x32_packx_test",
6063 srcs = [
6064 "test/x32-packx.cc",
6065 "test/pack-microkernel-tester.h",
6066 "src/xnnpack/AlignedAllocator.h",
6067 ] + MICROKERNEL_TEST_HDRS,
6068 deps = MICROKERNEL_TEST_DEPS,
6069)
6070
6071xnnpack_unit_test(
6072 name = "x32_pad_test",
6073 srcs = [
6074 "test/x32-pad.cc",
6075 "test/pad-microkernel-tester.h",
6076 ] + MICROKERNEL_TEST_HDRS,
6077 deps = MICROKERNEL_TEST_DEPS,
6078)
6079
6080xnnpack_unit_test(
6081 name = "x32_unpool_test",
6082 srcs = [
6083 "test/x32-unpool.cc",
6084 "test/unpool-microkernel-tester.h",
6085 ] + MICROKERNEL_TEST_HDRS,
6086 deps = MICROKERNEL_TEST_DEPS,
6087)
6088
6089xnnpack_unit_test(
6090 name = "x32_zip_test",
6091 srcs = [
6092 "test/x32-zip.cc",
6093 "test/zip-microkernel-tester.h",
6094 ] + MICROKERNEL_TEST_HDRS,
6095 deps = MICROKERNEL_TEST_DEPS,
6096)
6097
6098xnnpack_unit_test(
6099 name = "x8_lut_test",
6100 srcs = [
6101 "test/x8-lut.cc",
6102 "test/lut-microkernel-tester.h",
6103 ] + MICROKERNEL_TEST_HDRS,
6104 deps = MICROKERNEL_TEST_DEPS,
6105)
6106
6107xnnpack_unit_test(
6108 name = "x8_zip_test",
6109 srcs = [
6110 "test/x8-zip.cc",
6111 "test/zip-microkernel-tester.h",
6112 ] + MICROKERNEL_TEST_HDRS,
6113 deps = MICROKERNEL_TEST_DEPS,
6114)
6115
Marat Dukhan20c3b922020-03-10 03:45:06 -07006116########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006117
6118xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006119 name = "operator_size_test",
6120 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006121 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006122)
6123
Marat Dukhan20c3b922020-03-10 03:45:06 -07006124xnnpack_binary(
6125 name = "subgraph_size_test",
6126 srcs = ["test/subgraph-size.c"],
6127 deps = [":XNNPACK"],
6128)
6129
6130########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006131
6132xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006133 name = "abs_nc_test",
6134 srcs = [
6135 "test/abs-nc.cc",
6136 "test/abs-operator-tester.h",
6137 ],
6138 deps = OPERATOR_TEST_DEPS,
6139)
6140
6141xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006142 name = "add_nd_test",
6143 srcs = [
6144 "test/add-nd.cc",
6145 "test/binary-elementwise-operator-tester.h",
6146 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006147 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006148)
6149
6150xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006151 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006152 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006153 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006154 "test/argmax-pooling-operator-tester.h",
6155 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006156 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006157)
6158
6159xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006160 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006161 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006162 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006163 "test/average-pooling-operator-tester.h",
6164 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006165 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006166)
6167
6168xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006169 name = "bankers_rounding_nc_test",
6170 srcs = [
6171 "test/bankers-rounding-nc.cc",
6172 "test/bankers-rounding-operator-tester.h",
6173 ],
6174 deps = OPERATOR_TEST_DEPS,
6175)
6176
6177xnnpack_unit_test(
6178 name = "ceiling_nc_test",
6179 srcs = [
6180 "test/ceiling-nc.cc",
6181 "test/ceiling-operator-tester.h",
6182 ],
6183 deps = OPERATOR_TEST_DEPS,
6184)
6185
6186xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006187 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006188 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006189 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006190 "test/channel-shuffle-operator-tester.h",
6191 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006192 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006193)
6194
6195xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006196 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006197 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006198 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006199 "test/clamp-operator-tester.h",
6200 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006201 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006202)
6203
6204xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006205 name = "constant_pad_nd_test",
6206 srcs = [
6207 "test/constant-pad-nd.cc",
6208 "test/constant-pad-operator-tester.h",
6209 ],
6210 deps = OPERATOR_TEST_DEPS,
6211)
6212
6213xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006214 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006215 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006216 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006217 "test/convolution-operator-tester.h",
6218 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006219 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006220)
6221
6222xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006223 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006224 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006225 "test/convolution-nchw.cc",
6226 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006227 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006228 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006229)
6230
6231xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006232 name = "copy_nc_test",
6233 srcs = [
6234 "test/copy-nc.cc",
6235 "test/copy-operator-tester.h",
6236 ],
6237 deps = OPERATOR_TEST_DEPS,
6238)
6239
6240xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006241 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006242 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006243 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006244 "test/deconvolution-operator-tester.h",
6245 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006246 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006247)
6248
6249xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006250 name = "divide_nd_test",
6251 srcs = [
6252 "test/binary-elementwise-operator-tester.h",
6253 "test/divide-nd.cc",
6254 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006255 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006256)
6257
6258xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006259 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006260 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006261 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006262 "test/fully-connected-operator-tester.h",
6263 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006264 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006265)
6266
6267xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006268 name = "floor_nc_test",
6269 srcs = [
6270 "test/floor-nc.cc",
6271 "test/floor-operator-tester.h",
6272 ],
6273 deps = OPERATOR_TEST_DEPS,
6274)
6275
6276xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006277 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006278 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006279 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006280 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006281 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006282 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006283)
6284
6285xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006286 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006287 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006288 "test/global-average-pooling-ncw.cc",
6289 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006290 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006291 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006292)
6293
6294xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006295 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006296 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006297 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006298 "test/hardswish-operator-tester.h",
6299 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006300 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006301)
6302
6303xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006304 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006305 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006306 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006307 "test/leaky-relu-operator-tester.h",
6308 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006309 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006310)
6311
6312xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006313 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006314 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006315 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006316 "test/max-pooling-operator-tester.h",
6317 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006318 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006319)
6320
6321xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006322 name = "maximum_nd_test",
6323 srcs = [
6324 "test/binary-elementwise-operator-tester.h",
6325 "test/maximum-nd.cc",
6326 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006327 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006328)
6329
6330xnnpack_unit_test(
6331 name = "minimum_nd_test",
6332 srcs = [
6333 "test/binary-elementwise-operator-tester.h",
6334 "test/minimum-nd.cc",
6335 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006336 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006337)
6338
6339xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006340 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006341 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006342 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006343 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006344 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006345 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006346)
6347
6348xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006349 name = "negate_nc_test",
6350 srcs = [
6351 "test/negate-nc.cc",
6352 "test/negate-operator-tester.h",
6353 ],
6354 deps = OPERATOR_TEST_DEPS,
6355)
6356
6357xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006358 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006359 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006360 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006361 "test/prelu-operator-tester.h",
6362 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006363 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006364)
6365
6366xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006367 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006368 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006369 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006370 "test/resize-bilinear-operator-tester.h",
6371 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006372 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006373)
6374
6375xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006376 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006377 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006378 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006379 "test/sigmoid-operator-tester.h",
6380 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006381 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006382)
6383
6384xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006385 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006386 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006387 "test/softmax-nc.cc",
6388 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006389 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006390 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006391)
6392
6393xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006394 name = "square_nc_test",
6395 srcs = [
6396 "test/square-nc.cc",
6397 "test/square-operator-tester.h",
6398 ],
6399 deps = OPERATOR_TEST_DEPS,
6400)
6401
6402xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006403 name = "square_root_nc_test",
6404 srcs = [
6405 "test/square-root-nc.cc",
6406 "test/square-root-operator-tester.h",
6407 ],
6408 deps = OPERATOR_TEST_DEPS,
6409)
6410
6411xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006412 name = "squared_difference_nd_test",
6413 srcs = [
6414 "test/binary-elementwise-operator-tester.h",
6415 "test/squared-difference-nd.cc",
6416 ],
6417 deps = OPERATOR_TEST_DEPS,
6418)
6419
6420xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006421 name = "subtract_nd_test",
6422 srcs = [
6423 "test/binary-elementwise-operator-tester.h",
6424 "test/subtract-nd.cc",
6425 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006426 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006427)
6428
6429xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006430 name = "truncation_nc_test",
6431 srcs = [
6432 "test/truncation-nc.cc",
6433 "test/truncation-operator-tester.h",
6434 ],
6435 deps = OPERATOR_TEST_DEPS,
6436)
6437
6438xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006439 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006440 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006441 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006442 "test/unpooling-operator-tester.h",
6443 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006444 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006445)
6446
Chao Mei6ddfc602020-05-13 22:29:36 -07006447############################### Misc unit tests ###############################
6448
6449xnnpack_unit_test(
6450 name = "memory_planner_test",
6451 srcs = [
6452 "test/memory-planner-test.cc",
6453 ],
6454 deps = [
6455 ":XNNPACK",
6456 ":memory_planner",
6457 ],
6458)
6459
Marat Dukhan08c4a432019-10-03 09:29:21 -07006460############################# Build configurations #############################
6461
Marat Dukhanb8642352019-10-30 15:43:02 -07006462# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006463config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006464 name = "xnn_enable_assembly_explicit_true",
6465 define_values = {"xnn_enable_assembly": "true"},
6466)
6467
6468# Disables usage of assembly kernels.
6469config_setting(
6470 name = "xnn_enable_assembly_explicit_false",
6471 define_values = {"xnn_enable_assembly": "false"},
6472)
6473
Marat Dukhan9de90e02020-06-18 16:04:12 -07006474# Enables usage of sparse inference.
6475config_setting(
6476 name = "xnn_enable_sparse_explicit_true",
6477 define_values = {"xnn_enable_sparse": "true"},
6478)
6479
6480# Disables usage of sparse inference.
6481config_setting(
6482 name = "xnn_enable_sparse_explicit_false",
6483 define_values = {"xnn_enable_sparse": "false"},
6484)
6485
Marat Dukhan05702cf2020-03-26 15:41:33 -07006486# Disables usage of HMP-aware optimizations.
6487config_setting(
6488 name = "xnn_enable_hmp_explicit_false",
6489 define_values = {"xnn_enable_hmp": "false"},
6490)
6491
Chao Mei6ddfc602020-05-13 22:29:36 -07006492# Enable usage of optimized memory allocation
6493config_setting(
6494 name = "xnn_enable_memopt_explicit_true",
6495 define_values = {"xnnpack_enable_memopt": "true"},
6496)
6497
6498# Disable usage of optimized memory allocation
6499config_setting(
6500 name = "xnn_enable_memopt_explicit_false",
6501 define_values = {"xnnpack_enable_memopt": "false"},
6502)
6503
Marat Dukhanb8642352019-10-30 15:43:02 -07006504# Builds with -c dbg
6505config_setting(
6506 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006507 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006508 "compilation_mode": "dbg",
6509 },
6510)
6511
6512# Builds with -c opt
6513config_setting(
6514 name = "optimized_build",
6515 values = {
6516 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006517 },
6518)
6519
6520config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006521 name = "linux_k8",
6522 values = {"cpu": "k8"},
6523)
6524
6525config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006526 name = "linux_arm",
6527 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006528)
6529
6530config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006531 name = "linux_armeabi",
6532 values = {"cpu": "armeabi"},
6533)
6534
6535config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006536 name = "linux_armhf",
6537 values = {"cpu": "armhf"},
6538)
6539
6540config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006541 name = "linux_armv7a",
6542 values = {"cpu": "armv7a"},
6543)
6544
6545config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006546 name = "linux_aarch64",
6547 values = {"cpu": "aarch64"},
6548)
6549
6550config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006551 name = "android",
6552 values = {"crosstool_top": "//external:android/crosstool"},
6553)
6554
6555config_setting(
6556 name = "android_armv7",
6557 values = {
6558 "crosstool_top": "//external:android/crosstool",
6559 "cpu": "armeabi-v7a",
6560 },
6561)
6562
6563config_setting(
6564 name = "android_arm64",
6565 values = {
6566 "crosstool_top": "//external:android/crosstool",
6567 "cpu": "arm64-v8a",
6568 },
6569)
6570
6571config_setting(
6572 name = "android_x86",
6573 values = {
6574 "crosstool_top": "//external:android/crosstool",
6575 "cpu": "x86",
6576 },
6577)
6578
6579config_setting(
6580 name = "android_x86_64",
6581 values = {
6582 "crosstool_top": "//external:android/crosstool",
6583 "cpu": "x86_64",
6584 },
6585)
6586
6587config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006588 name = "windows_x86_64",
6589 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006590)
6591
6592config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006593 name = "windows_x86_64_clang",
6594 values = {
6595 "compiler": "clang-cl",
6596 "cpu": "x64_windows",
6597 },
6598)
6599
6600config_setting(
6601 name = "windows_x86_64_mingw",
6602 values = {
6603 "compiler": "mingw-gcc",
6604 "cpu": "x64_windows",
6605 },
6606)
6607
6608config_setting(
6609 name = "windows_x86_64_msys",
6610 values = {
6611 "compiler": "msys-gcc",
6612 "cpu": "x64_windows",
6613 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006614)
6615
6616config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006617 name = "macos_x86_64",
6618 values = {
6619 "apple_platform_type": "macos",
6620 "cpu": "darwin",
6621 },
6622)
6623
6624config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006625 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006626 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006627)
6628
6629config_setting(
6630 name = "emscripten_wasm",
6631 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006632 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006633 "cpu": "wasm",
6634 },
6635)
6636
6637config_setting(
6638 name = "emscripten_wasmsimd",
6639 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006640 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006641 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006642 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006643 },
6644)
6645
6646config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006647 name = "ios_armv7",
6648 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006649 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006650 "cpu": "ios_armv7",
6651 },
6652)
6653
6654config_setting(
6655 name = "ios_arm64",
6656 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006657 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006658 "cpu": "ios_arm64",
6659 },
6660)
6661
6662config_setting(
6663 name = "ios_arm64e",
6664 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006665 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006666 "cpu": "ios_arm64e",
6667 },
6668)
6669
6670config_setting(
6671 name = "ios_x86",
6672 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006673 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006674 "cpu": "ios_i386",
6675 },
6676)
6677
6678config_setting(
6679 name = "ios_x86_64",
6680 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006681 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006682 "cpu": "ios_x86_64",
6683 },
6684)
6685
6686config_setting(
6687 name = "watchos_armv7k",
6688 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006689 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006690 "cpu": "watchos_armv7k",
6691 },
6692)
6693
6694config_setting(
6695 name = "watchos_arm64_32",
6696 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006697 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006698 "cpu": "watchos_arm64_32",
6699 },
6700)
6701
6702config_setting(
6703 name = "watchos_x86",
6704 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006705 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006706 "cpu": "watchos_i386",
6707 },
6708)
6709
6710config_setting(
6711 name = "watchos_x86_64",
6712 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006713 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006714 "cpu": "watchos_x86_64",
6715 },
6716)
6717
6718config_setting(
6719 name = "tvos_arm64",
6720 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006721 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006722 "cpu": "tvos_arm64",
6723 },
6724)
6725
6726config_setting(
6727 name = "tvos_x86_64",
6728 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006729 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006730 "cpu": "tvos_x86_64",
6731 },
6732)