blob: 14f2df2d63adab8a77cf1d6038f52730236de71b [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 Barchardf1fd89e2020-10-06 23:27:11 -07002819 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07002820 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002821]
2822
Marat Dukhan1b354632020-03-23 12:50:22 -07002823INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002824 "src/xnnpack/argmaxpool.h",
2825 "src/xnnpack/avgpool.h",
2826 "src/xnnpack/clamp.h",
2827 "src/xnnpack/common.h",
2828 "src/xnnpack/conv.h",
2829 "src/xnnpack/dwconv.h",
2830 "src/xnnpack/gavgpool.h",
2831 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002832 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002833 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002834 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002835 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002836 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002837 "src/xnnpack/lut.h",
2838 "src/xnnpack/math.h",
2839 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002840 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002841 "src/xnnpack/packx.h",
2842 "src/xnnpack/pad.h",
2843 "src/xnnpack/params.h",
2844 "src/xnnpack/pavgpool.h",
2845 "src/xnnpack/ppmm.h",
2846 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002847 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002848 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002849 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002850 "src/xnnpack/rmax.h",
2851 "src/xnnpack/scalar-utils.h",
2852 "src/xnnpack/spmm.h",
2853 "src/xnnpack/unpool.h",
2854 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002855 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002856 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002857 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002858 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002859 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002860 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002861 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002862]
2863
2864INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002865 "include/xnnpack.h",
2866 "src/xnnpack/allocator.h",
2867 "src/xnnpack/compute.h",
2868 "src/xnnpack/im2col.h",
2869 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002870 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002871 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002872 "src/xnnpack/operator.h",
2873 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002874 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002875 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002876 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002877 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002878]
2879
Marat Dukhan1b354632020-03-23 12:50:22 -07002880ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002881 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002882]
2883
Marat Dukhan1b354632020-03-23 12:50:22 -07002884MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002885 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002886 "include/xnnpack.h",
2887]
2888
Marat Dukhan1b354632020-03-23 12:50:22 -07002889MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002890 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002891 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002892 "src/xnnpack/requantization.h",
2893 "include/xnnpack.h",
2894]
2895
2896OPERATOR_TEST_PARAMS_HDRS = [
2897 "src/xnnpack/params.h",
2898 "src/xnnpack/common.h",
2899]
2900
2901WEIGHTS_PACK_HDRS = [
2902 "src/xnnpack/pack.h",
2903 "src/xnnpack/operator.h",
2904 "src/xnnpack/compute.h",
2905]
2906
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002907LOGGING_COPTS = select({
2908 # No logging in optimized mode
2909 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2910 # Full logging in debug mode
2911 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2912 # Error-only logging in default (fastbuild) mode
2913 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2914})
2915
Marat Dukhan3b59de22020-06-03 20:15:19 -07002916LOGGING_SRCS = select({
2917 # No logging in optimized mode
2918 ":optimized_build": [],
2919 "//conditions:default": [
2920 "src/operator-strings.c",
2921 "src/subgraph-strings.c",
2922 ],
2923})
2924
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002925LOGGING_HDRS = [
2926 "src/xnnpack/log.h",
2927]
2928
Marat Dukhan08c4a432019-10-03 09:29:21 -07002929xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002930 name = "tables",
2931 srcs = TABLE_SRCS,
2932 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002933 gcc_copts = xnnpack_gcc_std_copts(),
2934 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002935)
2936
2937xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002938 name = "scalar_ukernels",
2939 srcs = SCALAR_UKERNELS,
2940 hdrs = INTERNAL_HDRS,
2941 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002942 gcc_copts = xnnpack_gcc_std_copts(),
2943 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002944 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002945 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002946 "@FP16",
2947 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002948 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002949 ],
2950)
2951
2952xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002953 name = "scalar_ukernels_test_mode",
2954 srcs = SCALAR_UKERNELS,
2955 hdrs = INTERNAL_HDRS,
2956 aarch32_copts = ["-marm"],
2957 copts = [
2958 "-UNDEBUG",
2959 "-DXNN_TEST_MODE=1",
2960 ],
2961 gcc_copts = xnnpack_gcc_std_copts(),
2962 msvc_copts = xnnpack_msvc_std_copts(),
2963 deps = [
2964 ":tables",
2965 "@FP16",
2966 "@FXdiv",
2967 "@pthreadpool",
2968 ],
2969)
2970
2971xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002972 name = "wasm_ukernels",
2973 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002974 gcc_copts = xnnpack_gcc_std_copts(),
2975 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002976 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002977 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002978 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002979 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002980 "@FP16",
2981 "@FXdiv",
2982 "@pthreadpool",
2983 ],
2984)
2985
2986xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002987 name = "wasm_ukernels_test_mode",
2988 hdrs = INTERNAL_HDRS,
2989 copts = [
2990 "-UNDEBUG",
2991 "-DXNN_TEST_MODE=1",
2992 ],
2993 gcc_copts = xnnpack_gcc_std_copts(),
2994 msvc_copts = xnnpack_msvc_std_copts(),
2995 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002996 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002997 deps = [
2998 ":tables",
2999 "@FP16",
3000 "@FXdiv",
3001 "@pthreadpool",
3002 ],
3003)
3004
3005xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003006 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003007 hdrs = INTERNAL_HDRS,
3008 aarch32_copts = [
3009 "-marm",
3010 "-mfpu=neon",
3011 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003012 gcc_copts = xnnpack_gcc_std_copts(),
3013 gcc_x86_copts = ["-msse2"],
3014 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003015 optimized_copts = [
3016 "-O3",
3017 "-ffast-math",
3018 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003019 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003020 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003021 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003022 "@FP16",
3023 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003024 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003025 ],
3026)
3027
3028xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003029 name = "psimd_fastmath_ukernels_test_mode",
3030 hdrs = INTERNAL_HDRS,
3031 aarch32_copts = [
3032 "-marm",
3033 "-mfpu=neon",
3034 ],
3035 copts = [
3036 "-UNDEBUG",
3037 "-DXNN_TEST_MODE=1",
3038 ],
3039 gcc_copts = xnnpack_gcc_std_copts(),
3040 gcc_x86_copts = ["-msse2"],
3041 msvc_copts = xnnpack_msvc_std_copts(),
3042 optimized_copts = [
3043 "-O3",
3044 "-ffast-math",
3045 ],
3046 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
3047 deps = [
3048 ":tables",
3049 "@FP16",
3050 "@psimd",
3051 "@pthreadpool",
3052 ],
3053)
3054
3055xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003056 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003057 hdrs = INTERNAL_HDRS,
3058 aarch32_copts = [
3059 "-marm",
3060 "-mfpu=neon",
3061 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003062 gcc_copts = xnnpack_gcc_std_copts(),
3063 gcc_x86_copts = ["-msse2"],
3064 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003065 optimized_copts = [
3066 "-O3",
3067 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003068 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003069 deps = [
3070 ":tables",
3071 "@FP16",
3072 "@psimd",
3073 "@pthreadpool",
3074 ],
3075)
3076
3077xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003078 name = "psimd_accmath_ukernels_test_mode",
3079 hdrs = INTERNAL_HDRS,
3080 aarch32_copts = [
3081 "-marm",
3082 "-mfpu=neon",
3083 ],
3084 copts = [
3085 "-UNDEBUG",
3086 "-DXNN_TEST_MODE=1",
3087 ],
3088 gcc_copts = xnnpack_gcc_std_copts(),
3089 gcc_x86_copts = ["-msse2"],
3090 msvc_copts = xnnpack_msvc_std_copts(),
3091 optimized_copts = [
3092 "-O3",
3093 ],
3094 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3095 deps = [
3096 ":tables",
3097 "@FP16",
3098 "@psimd",
3099 "@pthreadpool",
3100 ],
3101)
3102
3103xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003104 name = "neon_ukernels",
3105 hdrs = INTERNAL_HDRS,
3106 aarch32_copts = [
3107 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003108 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003109 "-mfpu=neon",
3110 ],
3111 aarch32_srcs = NEON_UKERNELS,
3112 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003113 gcc_copts = xnnpack_gcc_std_copts(),
3114 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003115 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003116 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003117 "@FP16",
3118 "@pthreadpool",
3119 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003120)
3121
3122xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003123 name = "neon_ukernels_test_mode",
3124 hdrs = INTERNAL_HDRS,
3125 aarch32_copts = [
3126 "-marm",
3127 "-march=armv7-a",
3128 "-mfpu=neon",
3129 ],
3130 aarch32_srcs = NEON_UKERNELS,
3131 aarch64_srcs = NEON_UKERNELS,
3132 copts = [
3133 "-UNDEBUG",
3134 "-DXNN_TEST_MODE=1",
3135 ],
3136 gcc_copts = xnnpack_gcc_std_copts(),
3137 msvc_copts = xnnpack_msvc_std_copts(),
3138 deps = [
3139 ":tables",
3140 "@FP16",
3141 "@pthreadpool",
3142 ],
3143)
3144
3145xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003146 name = "neonfma_ukernels",
3147 hdrs = INTERNAL_HDRS,
3148 aarch32_copts = [
3149 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003150 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003151 "-mfpu=neon-vfpv4",
3152 ],
3153 aarch32_srcs = NEONFMA_UKERNELS,
3154 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003155 apple_aarch32_copts = [
3156 "-mcpu=swift",
3157 "-mtune=generic",
3158 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003159 gcc_copts = xnnpack_gcc_std_copts(),
3160 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003161 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003162 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003163 "@FP16",
3164 "@pthreadpool",
3165 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003166)
3167
3168xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003169 name = "neonfma_ukernels_test_mode",
3170 hdrs = INTERNAL_HDRS,
3171 aarch32_copts = [
3172 "-marm",
3173 "-march=armv7-a",
3174 "-mfpu=neon-vfpv4",
3175 ],
3176 aarch32_srcs = NEONFMA_UKERNELS,
3177 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003178 apple_aarch32_copts = [
3179 "-mcpu=swift",
3180 "-mtune=generic",
3181 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003182 copts = [
3183 "-UNDEBUG",
3184 "-DXNN_TEST_MODE=1",
3185 ],
3186 gcc_copts = xnnpack_gcc_std_copts(),
3187 msvc_copts = xnnpack_msvc_std_copts(),
3188 deps = [
3189 ":tables",
3190 "@FP16",
3191 "@pthreadpool",
3192 ],
3193)
3194
3195xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003196 name = "neonv8_ukernels",
3197 hdrs = INTERNAL_HDRS,
3198 aarch32_copts = [
3199 "-marm",
3200 "-march=armv8-a",
3201 "-mfpu=neon-fp-armv8",
3202 ],
3203 aarch32_srcs = NEONV8_UKERNELS,
3204 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003205 apple_aarch32_copts = [
3206 "-mcpu=cyclone",
3207 "-mtune=generic",
3208 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003209 gcc_copts = xnnpack_gcc_std_copts(),
3210 msvc_copts = xnnpack_msvc_std_copts(),
3211 deps = [
3212 ":tables",
3213 "@FP16",
3214 "@pthreadpool",
3215 ],
3216)
3217
3218xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003219 name = "neonv8_ukernels_test_mode",
3220 hdrs = INTERNAL_HDRS,
3221 aarch32_copts = [
3222 "-marm",
3223 "-march=armv8-a",
3224 "-mfpu=neon-fp-armv8",
3225 ],
3226 aarch32_srcs = NEONV8_UKERNELS,
3227 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003228 apple_aarch32_copts = [
3229 "-mcpu=cyclone",
3230 "-mtune=generic",
3231 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003232 copts = [
3233 "-UNDEBUG",
3234 "-DXNN_TEST_MODE=1",
3235 ],
3236 gcc_copts = xnnpack_gcc_std_copts(),
3237 msvc_copts = xnnpack_msvc_std_copts(),
3238 deps = [
3239 ":tables",
3240 "@FP16",
3241 "@pthreadpool",
3242 ],
3243)
3244
3245xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003246 name = "neonfp16arith_ukernels",
3247 hdrs = INTERNAL_HDRS,
3248 aarch64_copts = ["-march=armv8.2-a+fp16"],
3249 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003250 gcc_copts = xnnpack_gcc_std_copts(),
3251 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003252 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003253 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003254 "@FP16",
3255 "@pthreadpool",
3256 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003257)
3258
3259xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003260 name = "neonfp16arith_ukernels_test_mode",
3261 hdrs = INTERNAL_HDRS,
3262 aarch64_copts = ["-march=armv8.2-a+fp16"],
3263 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3264 copts = [
3265 "-UNDEBUG",
3266 "-DXNN_TEST_MODE=1",
3267 ],
3268 gcc_copts = xnnpack_gcc_std_copts(),
3269 msvc_copts = xnnpack_msvc_std_copts(),
3270 deps = [
3271 ":tables",
3272 "@FP16",
3273 "@pthreadpool",
3274 ],
3275)
3276
3277xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003278 name = "neondot_ukernels",
3279 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003280 aarch32_copts = [
3281 "-marm",
3282 "-march=armv8.2-a+dotprod",
3283 "-mfpu=neon-fp-armv8",
3284 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003285 aarch32_srcs = NEONDOT_UKERNELS,
3286 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3287 aarch64_srcs = NEONDOT_UKERNELS,
3288 gcc_copts = xnnpack_gcc_std_copts(),
3289 msvc_copts = xnnpack_msvc_std_copts(),
3290 deps = [
3291 ":tables",
3292 "@FP16",
3293 "@pthreadpool",
3294 ],
3295)
3296
3297xnnpack_cc_library(
3298 name = "neondot_ukernels_test_mode",
3299 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003300 aarch32_copts = [
3301 "-marm",
3302 "-march=armv8.2-a+dotprod",
3303 "-mfpu=neon-fp-armv8",
3304 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003305 aarch32_srcs = NEONDOT_UKERNELS,
3306 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3307 aarch64_srcs = NEONDOT_UKERNELS,
3308 copts = [
3309 "-UNDEBUG",
3310 "-DXNN_TEST_MODE=1",
3311 ],
3312 gcc_copts = xnnpack_gcc_std_copts(),
3313 msvc_copts = xnnpack_msvc_std_copts(),
3314 deps = [
3315 ":tables",
3316 "@FP16",
3317 "@pthreadpool",
3318 ],
3319)
3320
3321xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003322 name = "sse2_ukernels",
3323 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003324 gcc_copts = xnnpack_gcc_std_copts(),
3325 gcc_x86_copts = ["-msse2"],
3326 msvc_copts = xnnpack_msvc_std_copts(),
3327 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003328 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003329 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003330 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003331 "@FP16",
3332 "@pthreadpool",
3333 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003334)
3335
3336xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003337 name = "sse2_ukernels_test_mode",
3338 hdrs = INTERNAL_HDRS,
3339 copts = [
3340 "-UNDEBUG",
3341 "-DXNN_TEST_MODE=1",
3342 ],
3343 gcc_copts = xnnpack_gcc_std_copts(),
3344 gcc_x86_copts = ["-msse2"],
3345 msvc_copts = xnnpack_msvc_std_copts(),
3346 msvc_x86_32_copts = ["/arch:SSE2"],
3347 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3348 deps = [
3349 ":tables",
3350 "@FP16",
3351 "@pthreadpool",
3352 ],
3353)
3354
3355xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003356 name = "ssse3_ukernels",
3357 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003358 gcc_copts = xnnpack_gcc_std_copts(),
3359 gcc_x86_copts = ["-mssse3"],
3360 msvc_copts = xnnpack_msvc_std_copts(),
3361 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003362 x86_srcs = SSSE3_UKERNELS,
3363 deps = [
3364 ":tables",
3365 "@FP16",
3366 "@pthreadpool",
3367 ],
3368)
3369
3370xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003371 name = "ssse3_ukernels_test_mode",
3372 hdrs = INTERNAL_HDRS,
3373 copts = [
3374 "-UNDEBUG",
3375 "-DXNN_TEST_MODE=1",
3376 ],
3377 gcc_copts = xnnpack_gcc_std_copts(),
3378 gcc_x86_copts = ["-mssse3"],
3379 msvc_copts = xnnpack_msvc_std_copts(),
3380 msvc_x86_32_copts = ["/arch:SSE2"],
3381 x86_srcs = SSSE3_UKERNELS,
3382 deps = [
3383 ":tables",
3384 "@FP16",
3385 "@pthreadpool",
3386 ],
3387)
3388
3389xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003390 name = "sse41_ukernels",
3391 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003392 gcc_copts = xnnpack_gcc_std_copts(),
3393 gcc_x86_copts = ["-msse4.1"],
3394 msvc_copts = xnnpack_msvc_std_copts(),
3395 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003396 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003397 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003398 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003399 "@FP16",
3400 "@pthreadpool",
3401 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003402)
3403
3404xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003405 name = "sse41_ukernels_test_mode",
3406 hdrs = INTERNAL_HDRS,
3407 copts = [
3408 "-UNDEBUG",
3409 "-DXNN_TEST_MODE=1",
3410 ],
3411 gcc_copts = xnnpack_gcc_std_copts(),
3412 gcc_x86_copts = ["-msse4.1"],
3413 msvc_copts = xnnpack_msvc_std_copts(),
3414 msvc_x86_32_copts = ["/arch:SSE2"],
3415 x86_srcs = SSE41_UKERNELS,
3416 deps = [
3417 ":tables",
3418 "@FP16",
3419 "@pthreadpool",
3420 ],
3421)
3422
3423xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003424 name = "avx_ukernels",
3425 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003426 gcc_copts = xnnpack_gcc_std_copts(),
3427 gcc_x86_copts = ["-mavx"],
3428 msvc_copts = xnnpack_msvc_std_copts(),
3429 msvc_x86_32_copts = ["/arch:AVX"],
3430 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003431 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003432 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003433 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003434 "@FP16",
3435 "@pthreadpool",
3436 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003437)
3438
3439xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003440 name = "avx_ukernels_test_mode",
3441 hdrs = INTERNAL_HDRS,
3442 copts = [
3443 "-UNDEBUG",
3444 "-DXNN_TEST_MODE=1",
3445 ],
3446 gcc_copts = xnnpack_gcc_std_copts(),
3447 gcc_x86_copts = ["-mavx"],
3448 msvc_copts = xnnpack_msvc_std_copts(),
3449 msvc_x86_32_copts = ["/arch:AVX"],
3450 msvc_x86_64_copts = ["/arch:AVX"],
3451 x86_srcs = AVX_UKERNELS,
3452 deps = [
3453 ":tables",
3454 "@FP16",
3455 "@pthreadpool",
3456 ],
3457)
3458
3459xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003460 name = "xop_ukernels",
3461 hdrs = INTERNAL_HDRS,
3462 gcc_copts = xnnpack_gcc_std_copts(),
3463 gcc_x86_copts = ["-mxop"],
3464 msvc_copts = xnnpack_msvc_std_copts(),
3465 msvc_x86_32_copts = ["/arch:AVX"],
3466 msvc_x86_64_copts = ["/arch:AVX"],
3467 x86_srcs = XOP_UKERNELS,
3468 deps = [
3469 ":tables",
3470 "@FP16",
3471 "@pthreadpool",
3472 ],
3473)
3474
3475xnnpack_cc_library(
3476 name = "xop_ukernels_test_mode",
3477 hdrs = INTERNAL_HDRS,
3478 copts = [
3479 "-UNDEBUG",
3480 "-DXNN_TEST_MODE=1",
3481 ],
3482 gcc_copts = xnnpack_gcc_std_copts(),
3483 gcc_x86_copts = ["-mxop"],
3484 msvc_copts = xnnpack_msvc_std_copts(),
3485 msvc_x86_32_copts = ["/arch:AVX"],
3486 msvc_x86_64_copts = ["/arch:AVX"],
3487 x86_srcs = XOP_UKERNELS,
3488 deps = [
3489 ":tables",
3490 "@FP16",
3491 "@pthreadpool",
3492 ],
3493)
3494
3495xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003496 name = "fma3_ukernels",
3497 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003498 gcc_copts = xnnpack_gcc_std_copts(),
3499 gcc_x86_copts = ["-mfma"],
3500 msvc_copts = xnnpack_msvc_std_copts(),
3501 msvc_x86_32_copts = ["/arch:AVX"],
3502 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003503 x86_srcs = FMA3_UKERNELS,
3504 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003505 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003506 "@FP16",
3507 "@pthreadpool",
3508 ],
3509)
3510
3511xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003512 name = "fma3_ukernels_test_mode",
3513 hdrs = INTERNAL_HDRS,
3514 copts = [
3515 "-UNDEBUG",
3516 "-DXNN_TEST_MODE=1",
3517 ],
3518 gcc_copts = xnnpack_gcc_std_copts(),
3519 gcc_x86_copts = ["-mfma"],
3520 msvc_copts = xnnpack_msvc_std_copts(),
3521 msvc_x86_32_copts = ["/arch:AVX"],
3522 msvc_x86_64_copts = ["/arch:AVX"],
3523 x86_srcs = FMA3_UKERNELS,
3524 deps = [
3525 ":tables",
3526 "@FP16",
3527 "@pthreadpool",
3528 ],
3529)
3530
3531xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003532 name = "avx2_ukernels",
3533 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003534 gcc_copts = xnnpack_gcc_std_copts(),
3535 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003536 "-mfma",
3537 "-mavx2",
3538 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003539 msvc_copts = xnnpack_msvc_std_copts(),
3540 msvc_x86_32_copts = ["/arch:AVX2"],
3541 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003542 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003543 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003544 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003545 "@FP16",
3546 "@pthreadpool",
3547 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003548)
3549
3550xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003551 name = "avx2_ukernels_test_mode",
3552 hdrs = INTERNAL_HDRS,
3553 copts = [
3554 "-UNDEBUG",
3555 "-DXNN_TEST_MODE=1",
3556 ],
3557 gcc_copts = xnnpack_gcc_std_copts(),
3558 gcc_x86_copts = [
3559 "-mfma",
3560 "-mavx2",
3561 ],
3562 msvc_copts = xnnpack_msvc_std_copts(),
3563 msvc_x86_32_copts = ["/arch:AVX2"],
3564 msvc_x86_64_copts = ["/arch:AVX2"],
3565 x86_srcs = AVX2_UKERNELS,
3566 deps = [
3567 ":tables",
3568 "@FP16",
3569 "@pthreadpool",
3570 ],
3571)
3572
3573xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003574 name = "avx512f_ukernels",
3575 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003576 gcc_copts = xnnpack_gcc_std_copts(),
3577 gcc_x86_copts = ["-mavx512f"],
3578 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3579 msvc_copts = xnnpack_msvc_std_copts(),
3580 msvc_x86_32_copts = ["/arch:AVX512"],
3581 msvc_x86_64_copts = ["/arch:AVX512"],
3582 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003583 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003584 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003585 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003586 "@FP16",
3587 "@pthreadpool",
3588 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003589)
3590
3591xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003592 name = "avx512f_ukernels_test_mode",
3593 hdrs = INTERNAL_HDRS,
3594 copts = [
3595 "-UNDEBUG",
3596 "-DXNN_TEST_MODE=1",
3597 ],
3598 gcc_copts = xnnpack_gcc_std_copts(),
3599 gcc_x86_copts = ["-mavx512f"],
3600 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3601 msvc_copts = xnnpack_msvc_std_copts(),
3602 msvc_x86_32_copts = ["/arch:AVX512"],
3603 msvc_x86_64_copts = ["/arch:AVX512"],
3604 msys_copts = ["-fno-asynchronous-unwind-tables"],
3605 x86_srcs = AVX512F_UKERNELS,
3606 deps = [
3607 ":tables",
3608 "@FP16",
3609 "@pthreadpool",
3610 ],
3611)
3612
3613xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003614 name = "avx512skx_ukernels",
3615 hdrs = INTERNAL_HDRS,
3616 gcc_copts = xnnpack_gcc_std_copts(),
3617 gcc_x86_copts = [
3618 "-mavx512f",
3619 "-mavx512cd",
3620 "-mavx512bw",
3621 "-mavx512dq",
3622 "-mavx512vl",
3623 ],
3624 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3625 msvc_copts = xnnpack_msvc_std_copts(),
3626 msvc_x86_32_copts = ["/arch:AVX512"],
3627 msvc_x86_64_copts = ["/arch:AVX512"],
3628 msys_copts = ["-fno-asynchronous-unwind-tables"],
3629 x86_srcs = AVX512SKX_UKERNELS,
3630 deps = [
3631 ":tables",
3632 "@FP16",
3633 "@pthreadpool",
3634 ],
3635)
3636
3637xnnpack_cc_library(
3638 name = "avx512skx_ukernels_test_mode",
3639 hdrs = INTERNAL_HDRS,
3640 copts = [
3641 "-UNDEBUG",
3642 "-DXNN_TEST_MODE=1",
3643 ],
3644 gcc_copts = xnnpack_gcc_std_copts(),
3645 gcc_x86_copts = [
3646 "-mavx512f",
3647 "-mavx512cd",
3648 "-mavx512bw",
3649 "-mavx512dq",
3650 "-mavx512vl",
3651 ],
3652 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3653 msvc_copts = xnnpack_msvc_std_copts(),
3654 msvc_x86_32_copts = ["/arch:AVX512"],
3655 msvc_x86_64_copts = ["/arch:AVX512"],
3656 msys_copts = ["-fno-asynchronous-unwind-tables"],
3657 x86_srcs = AVX512SKX_UKERNELS,
3658 deps = [
3659 ":tables",
3660 "@FP16",
3661 "@pthreadpool",
3662 ],
3663)
3664
3665xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003666 name = "asm_ukernels",
3667 hdrs = ["src/xnnpack/assembly.h"],
3668 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07003669 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003670 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003671 wasm_srcs = WASM32_ASM_UKERNELS,
3672 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003673)
3674
Marat Dukhan3b59de22020-06-03 20:15:19 -07003675xnnpack_cc_library(
3676 name = "logging_utils",
3677 srcs = LOGGING_SRCS,
3678 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3679 copts = LOGGING_COPTS + [
3680 "-Isrc",
3681 "-Iinclude",
3682 ] + select({
3683 ":debug_build": [],
3684 "//conditions:default": xnnpack_min_size_copts(),
3685 }),
3686 gcc_copts = xnnpack_gcc_std_copts(),
3687 msvc_copts = xnnpack_msvc_std_copts(),
3688 visibility = xnnpack_visibility(),
3689 deps = [
3690 "@FP16",
3691 "@clog",
3692 "@pthreadpool",
3693 ],
3694)
3695
Marat Dukhan08c4a432019-10-03 09:29:21 -07003696xnnpack_aggregate_library(
3697 name = "ukernels",
3698 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003699 ":neon_ukernels",
3700 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003701 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003702 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003703 ":asm_ukernels",
3704 ],
3705 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003706 ":neon_ukernels",
3707 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003708 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003709 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003710 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003711 ":asm_ukernels",
3712 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003713 generic_deps = [
3714 ":scalar_ukernels",
3715 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003716 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003717 ":psimd_fastmath_ukernels",
3718 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003719 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003720 wasm_deps = [
3721 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003722 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003723 ],
3724 wasmsimd_deps = [
3725 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003726 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003727 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003728 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003729 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003730 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003731 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003732 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003733 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003734 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003735 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003736 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003737 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003738 ],
3739)
3740
Marat Dukhan33fcf782020-05-24 14:27:15 -07003741xnnpack_aggregate_library(
3742 name = "ukernels_test_mode",
3743 aarch32_deps = [
3744 ":neon_ukernels_test_mode",
3745 ":neonfma_ukernels_test_mode",
3746 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003747 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003748 ":asm_ukernels",
3749 ],
3750 aarch64_deps = [
3751 ":neon_ukernels_test_mode",
3752 ":neonfma_ukernels_test_mode",
3753 ":neonv8_ukernels_test_mode",
3754 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003755 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003756 ":asm_ukernels",
3757 ],
3758 generic_deps = [
3759 ":scalar_ukernels_test_mode",
3760 ],
3761 psimd_deps = [
3762 ":psimd_fastmath_ukernels_test_mode",
3763 ":psimd_accmath_ukernels_test_mode",
3764 ],
3765 wasm_deps = [
3766 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003767 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003768 ],
3769 wasmsimd_deps = [
3770 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003771 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003772 ],
3773 x86_deps = [
3774 ":sse2_ukernels_test_mode",
3775 ":ssse3_ukernels_test_mode",
3776 ":sse41_ukernels_test_mode",
3777 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003778 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003779 ":fma3_ukernels_test_mode",
3780 ":avx2_ukernels_test_mode",
3781 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003782 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003783 ],
3784)
3785
Marat Dukhan86df3842020-06-17 14:03:31 -07003786# This target must be compatible with Android NDK r17 and compile with Clang 6.
3787# It is similar to :ukernels target, but excludes microkernels which require
3788# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003789xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003790 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003791 aarch32_deps = [
3792 ":neon_ukernels",
3793 ":neonfma_ukernels",
3794 ":neonv8_ukernels",
3795 ":asm_ukernels",
3796 ],
3797 aarch64_deps = [
3798 ":neon_ukernels",
3799 ":neonfma_ukernels",
3800 ":neonv8_ukernels",
3801 ":asm_ukernels",
3802 ],
3803 generic_deps = [
3804 ":scalar_ukernels",
3805 ],
3806 psimd_deps = [
3807 ":psimd_fastmath_ukernels",
3808 ":psimd_accmath_ukernels",
3809 ],
3810 wasm_deps = [
3811 ":wasm_ukernels",
3812 ],
3813 wasmsimd_deps = [
3814 ":wasm_ukernels",
3815 ],
3816 x86_deps = [
3817 ":sse2_ukernels",
3818 ":ssse3_ukernels",
3819 ":sse41_ukernels",
3820 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003821 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003822 ":fma3_ukernels",
3823 ":avx2_ukernels",
3824 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003825 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003826 ],
3827)
3828
Marat Dukhan08c4a432019-10-03 09:29:21 -07003829xnnpack_cc_library(
3830 name = "im2col",
3831 srcs = ["src/im2col.c"],
3832 hdrs = [
3833 "src/xnnpack/common.h",
3834 "src/xnnpack/im2col.h",
3835 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003836 gcc_copts = xnnpack_gcc_std_copts(),
3837 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003838)
3839
3840xnnpack_cc_library(
3841 name = "indirection",
3842 srcs = ["src/indirection.c"],
3843 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003844 gcc_copts = xnnpack_gcc_std_copts(),
3845 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003846 deps = [
3847 "@FP16",
3848 "@FXdiv",
3849 "@pthreadpool",
3850 ],
3851)
3852
3853xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003854 name = "indirection_test_mode",
3855 srcs = ["src/indirection.c"],
3856 hdrs = INTERNAL_HDRS,
3857 copts = [
3858 "-UNDEBUG",
3859 "-DXNN_TEST_MODE=1",
3860 ],
3861 gcc_copts = xnnpack_gcc_std_copts(),
3862 msvc_copts = xnnpack_msvc_std_copts(),
3863 deps = [
3864 "@FP16",
3865 "@FXdiv",
3866 "@pthreadpool",
3867 ],
3868)
3869
3870xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003871 name = "packing",
3872 srcs = ["src/packing.c"],
3873 hdrs = INTERNAL_HDRS,
3874 gcc_copts = xnnpack_gcc_std_copts(),
3875 msvc_copts = xnnpack_msvc_std_copts(),
3876 deps = [
3877 "@FP16",
3878 "@FXdiv",
3879 "@pthreadpool",
3880 ],
3881)
3882
3883xnnpack_cc_library(
3884 name = "packing_test_mode",
3885 srcs = ["src/packing.c"],
3886 hdrs = INTERNAL_HDRS,
3887 copts = [
3888 "-UNDEBUG",
3889 "-DXNN_TEST_MODE=1",
3890 ],
3891 gcc_copts = xnnpack_gcc_std_copts(),
3892 msvc_copts = xnnpack_msvc_std_copts(),
3893 deps = [
3894 "@FP16",
3895 "@FXdiv",
3896 "@pthreadpool",
3897 ],
3898)
3899
3900xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003901 name = "operator_run",
3902 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003903 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003904 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003905 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3906 "//conditions:default": [],
3907 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003908 gcc_copts = xnnpack_gcc_std_copts(),
3909 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003910 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003911 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003912 "@FP16",
3913 "@FXdiv",
3914 "@clog",
3915 "@pthreadpool",
3916 ],
3917)
3918
Chao Mei6ddfc602020-05-13 22:29:36 -07003919xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003920 name = "operator_run_test_mode",
3921 srcs = ["src/operator-run.c"],
3922 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3923 copts = LOGGING_COPTS + [
3924 "-UNDEBUG",
3925 "-DXNN_TEST_MODE=1",
3926 ] + select({
3927 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3928 "//conditions:default": [],
3929 }),
3930 gcc_copts = xnnpack_gcc_std_copts(),
3931 msvc_copts = xnnpack_msvc_std_copts(),
3932 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003933 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003934 "@FP16",
3935 "@FXdiv",
3936 "@clog",
3937 "@pthreadpool",
3938 ],
3939)
3940
3941xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003942 name = "memory_planner",
3943 srcs = ["src/memory-planner.c"],
3944 hdrs = INTERNAL_HDRS,
3945 defines = select({
3946 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3947 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3948 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3949 }),
3950 gcc_copts = xnnpack_gcc_std_copts(),
3951 msvc_copts = xnnpack_msvc_std_copts(),
3952 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003953 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003954 "@pthreadpool",
3955 ],
3956)
3957
Marat Dukhan33fcf782020-05-24 14:27:15 -07003958xnnpack_cc_library(
3959 name = "memory_planner_test_mode",
3960 srcs = ["src/memory-planner.c"],
3961 hdrs = INTERNAL_HDRS,
3962 copts = [
3963 "-UNDEBUG",
3964 "-DXNN_TEST_MODE=1",
3965 ],
3966 defines = select({
3967 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3968 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3969 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3970 }),
3971 gcc_copts = xnnpack_gcc_std_copts(),
3972 msvc_copts = xnnpack_msvc_std_copts(),
3973 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003974 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003975 "@pthreadpool",
3976 ],
3977)
3978
Marat Dukhan08c4a432019-10-03 09:29:21 -07003979cc_library(
3980 name = "enable_assembly",
3981 defines = select({
3982 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3983 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003984 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003985 }),
3986)
3987
Marat Dukhan9de90e02020-06-18 16:04:12 -07003988cc_library(
3989 name = "enable_sparse",
3990 defines = select({
3991 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3992 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003993 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003994 }),
3995)
3996
Marat Dukhancf056b22019-10-07 10:26:29 -07003997xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003998 name = "operators",
3999 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004000 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004001 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004002 ],
4003 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004004 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004005 "-Isrc",
4006 "-Iinclude",
4007 ] + select({
4008 ":debug_build": [],
4009 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004010 }) + select({
4011 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4012 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004013 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004014 gcc_copts = xnnpack_gcc_std_copts(),
4015 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004016 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004017 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004018 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004019 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004020 "@FP16",
4021 "@FXdiv",
4022 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004023 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004024 ],
4025)
4026
Marat Dukhan10a38082020-04-17 03:58:35 -07004027xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004028 name = "operators_test_mode",
4029 srcs = OPERATOR_SRCS + [
4030 "src/memory.c",
4031 "src/operator-delete.c",
4032 ],
4033 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4034 copts = LOGGING_COPTS + [
4035 "-Isrc",
4036 "-Iinclude",
4037 "-UNDEBUG",
4038 "-DXNN_TEST_MODE=1",
4039 ] + select({
4040 ":debug_build": [],
4041 "//conditions:default": xnnpack_min_size_copts(),
4042 }) + select({
4043 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4044 "//conditions:default": [],
4045 }),
4046 gcc_copts = xnnpack_gcc_std_copts(),
4047 msvc_copts = xnnpack_msvc_std_copts(),
4048 deps = [
4049 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004050 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004051 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004052 "@FP16",
4053 "@FXdiv",
4054 "@clog",
4055 "@pthreadpool",
4056 ],
4057)
4058
4059xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004060 name = "XNNPACK",
4061 srcs = [
4062 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004063 "src/runtime.c",
4064 "src/subgraph.c",
4065 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004066 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004067 hdrs = ["include/xnnpack.h"],
4068 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004069 "-Isrc",
4070 "-Iinclude",
4071 ] + select({
4072 ":debug_build": [],
4073 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004074 }) + select({
4075 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4076 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004077 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004078 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004079 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004080 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004081 visibility = xnnpack_visibility(),
4082 deps = [
4083 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004084 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004085 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004086 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004087 ":operator_run",
4088 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004089 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004090 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004091 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004092 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004093 ] + select({
4094 ":emscripten": [],
4095 "//conditions:default": ["@cpuinfo"],
4096 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004097)
4098
Marat Dukhan10a38082020-04-17 03:58:35 -07004099xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004100 name = "XNNPACK_test_mode",
4101 srcs = [
4102 "src/init.c",
4103 "src/runtime.c",
4104 "src/subgraph.c",
4105 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004106 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004107 hdrs = ["include/xnnpack.h"],
4108 copts = LOGGING_COPTS + [
4109 "-Isrc",
4110 "-Iinclude",
4111 "-UNDEBUG",
4112 "-DXNN_TEST_MODE=1",
4113 ] + select({
4114 ":debug_build": [],
4115 "//conditions:default": xnnpack_min_size_copts(),
4116 }) + select({
4117 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4118 "//conditions:default": [],
4119 }),
4120 gcc_copts = xnnpack_gcc_std_copts(),
4121 includes = ["include"],
4122 msvc_copts = xnnpack_msvc_std_copts(),
4123 visibility = xnnpack_visibility(),
4124 deps = [
4125 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004126 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004127 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004128 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004129 ":operator_run_test_mode",
4130 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004131 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004132 "@clog",
4133 "@FP16",
4134 "@pthreadpool",
4135 ] + select({
4136 ":emscripten": [],
4137 "//conditions:default": ["@cpuinfo"],
4138 }),
4139)
4140
4141xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004142 name = "xnnpack_f32",
4143 srcs = [
4144 "src/init.c",
4145 "src/runtime.c",
4146 "src/subgraph.c",
4147 "src/tensor.c",
4148 ] + SUBGRAPH_SRCS,
4149 hdrs = ["include/xnnpack.h"],
4150 copts = LOGGING_COPTS + [
4151 "-Isrc",
4152 "-Iinclude",
4153 ] + select({
4154 ":debug_build": [],
4155 "//conditions:default": xnnpack_min_size_copts(),
4156 }) + select({
4157 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4158 "//conditions:default": [],
4159 }),
4160 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004161 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004162 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004163 "XNN_NO_U8_OPERATORS",
4164 "XNN_NO_X8_OPERATORS",
4165 "XNN_NO_F16_OPERATORS",
4166 "XNN_NO_X16_OPERATORS",
4167 ],
4168 gcc_copts = xnnpack_gcc_std_copts(),
4169 includes = ["include"],
4170 msvc_copts = xnnpack_msvc_std_copts(),
4171 visibility = xnnpack_visibility(),
4172 deps = [
4173 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004174 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004175 ":logging_utils",
4176 ":memory_planner",
4177 ":operator_run",
4178 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004179 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004180 "@clog",
4181 "@FP16",
4182 "@pthreadpool",
4183 ] + select({
4184 ":emscripten": [],
4185 "//conditions:default": ["@cpuinfo"],
4186 }),
4187)
4188
4189xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004190 name = "xnnpack_operators_nhwc_f32",
4191 srcs = [
4192 "src/init.c",
4193 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004194 hdrs = ["include/xnnpack.h"],
4195 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004196 "-Isrc",
4197 "-Iinclude",
4198 ] + select({
4199 ":debug_build": [],
4200 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004201 }) + select({
4202 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4203 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004204 }),
4205 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004206 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004207 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004208 "XNN_NO_U8_OPERATORS",
4209 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004210 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004211 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004212 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004213 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004214 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004215 visibility = xnnpack_visibility(),
4216 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004217 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004218 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004219 ":operator_run",
4220 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004221 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004222 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004223 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004224 ] + select({
4225 ":emscripten": [],
4226 "//conditions:default": ["@cpuinfo"],
4227 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004228)
4229
Marat Dukhancf056b22019-10-07 10:26:29 -07004230xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004231 name = "bench_utils",
4232 srcs = ["bench/utils.cc"],
4233 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004234 deps = [
4235 "@com_google_benchmark//:benchmark",
4236 "@cpuinfo",
4237 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004238)
4239
Frank Barchard7e955972019-10-11 10:34:25 -07004240######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004241
4242xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004243 name = "qs8_gemm_bench",
4244 srcs = [
4245 "bench/gemm.h",
4246 "bench/qs8-gemm.cc",
4247 "src/xnnpack/AlignedAllocator.h",
4248 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
4249 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
4250)
4251
4252xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004253 name = "qs8_requantization_bench",
4254 srcs = [
4255 "bench/qs8-requantization.cc",
4256 "src/xnnpack/requantization-stubs.h",
4257 "src/xnnpack/AlignedAllocator.h",
4258 ] + MICROKERNEL_BENCHMARK_HDRS,
4259 deps = MICROKERNEL_BENCHMARK_DEPS,
4260)
4261
4262xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004263 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004264 srcs = [
4265 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004266 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004267 "src/xnnpack/AlignedAllocator.h",
4268 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004269 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004270 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004271)
4272
4273xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004274 name = "qu8_requantization_bench",
4275 srcs = [
4276 "bench/qu8-requantization.cc",
4277 "src/xnnpack/requantization-stubs.h",
4278 "src/xnnpack/AlignedAllocator.h",
4279 ] + MICROKERNEL_BENCHMARK_HDRS,
4280 deps = MICROKERNEL_BENCHMARK_DEPS,
4281)
4282
4283xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004284 name = "f16_igemm_bench",
4285 srcs = [
4286 "bench/f16-igemm.cc",
4287 "bench/conv.h",
4288 "bench/google/conv.h",
4289 "src/xnnpack/AlignedAllocator.h",
4290 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004291 deps = MICROKERNEL_BENCHMARK_DEPS + [
4292 ":indirection",
4293 ":packing",
4294 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004295)
4296
4297xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004298 name = "f16_gemm_bench",
4299 srcs = [
4300 "bench/f16-gemm.cc",
4301 "bench/gemm.h",
4302 "src/xnnpack/AlignedAllocator.h",
4303 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004304 deps = MICROKERNEL_BENCHMARK_DEPS + [
4305 ":packing",
4306 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004307)
4308
4309xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004310 name = "f16_spmm_bench",
4311 srcs = [
4312 "bench/f16-spmm.cc",
4313 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004314 "src/xnnpack/AlignedAllocator.h",
4315 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004316 deps = MICROKERNEL_BENCHMARK_DEPS,
4317)
4318
4319xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004320 name = "f32_igemm_bench",
4321 srcs = [
4322 "bench/f32-igemm.cc",
4323 "bench/conv.h",
4324 "src/xnnpack/AlignedAllocator.h",
4325 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004326 deps = MICROKERNEL_BENCHMARK_DEPS + [
4327 ":indirection",
4328 ":packing",
4329 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004330)
4331
4332xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004333 name = "f16_relu_bench",
4334 srcs = [
4335 "bench/f16-relu.cc",
4336 "src/xnnpack/AlignedAllocator.h",
4337 ] + MICROKERNEL_BENCHMARK_HDRS,
4338 deps = MICROKERNEL_BENCHMARK_DEPS,
4339)
4340
4341xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004342 name = "f32_conv_hwc_bench",
4343 srcs = [
4344 "bench/f32-conv-hwc.cc",
4345 "bench/dconv.h",
4346 "src/xnnpack/AlignedAllocator.h",
4347 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004348 deps = MICROKERNEL_BENCHMARK_DEPS + [
4349 ":packing",
4350 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004351)
4352
4353xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004354 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004355 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004356 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004357 "bench/dconv.h",
4358 "src/xnnpack/AlignedAllocator.h",
4359 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004360 deps = MICROKERNEL_BENCHMARK_DEPS + [
4361 ":packing",
4362 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004363)
4364
4365xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004366 name = "f16_dwconv_bench",
4367 srcs = [
4368 "bench/f16-dwconv.cc",
4369 "bench/dwconv.h",
4370 "bench/google/dwconv.h",
4371 "src/xnnpack/AlignedAllocator.h",
4372 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004373 deps = MICROKERNEL_BENCHMARK_DEPS + [
4374 ":indirection",
4375 ":packing",
4376 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004377)
4378
4379xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004380 name = "f32_dwconv_bench",
4381 srcs = [
4382 "bench/f32-dwconv.cc",
4383 "bench/dwconv.h",
4384 "src/xnnpack/AlignedAllocator.h",
4385 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004386 deps = MICROKERNEL_BENCHMARK_DEPS + [
4387 ":indirection",
4388 ":packing",
4389 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004390)
4391
4392xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004393 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004394 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004395 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004396 "bench/dwconv.h",
4397 "src/xnnpack/AlignedAllocator.h",
4398 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004399 deps = MICROKERNEL_BENCHMARK_DEPS + [
4400 ":indirection",
4401 ":packing",
4402 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004403)
4404
4405xnnpack_benchmark(
4406 name = "f32_gemm_bench",
4407 srcs = [
4408 "bench/f32-gemm.cc",
4409 "bench/gemm.h",
4410 "src/xnnpack/AlignedAllocator.h",
4411 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004412 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004413 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004414)
4415
4416xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004417 name = "f32_hswish_bench",
4418 srcs = [
4419 "bench/f32-hswish.cc",
4420 "src/xnnpack/AlignedAllocator.h",
4421 ] + MICROKERNEL_BENCHMARK_HDRS,
4422 deps = MICROKERNEL_BENCHMARK_DEPS,
4423)
4424
4425xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004426 name = "f32_raddexpminusmax_bench",
4427 srcs = [
4428 "bench/f32-raddexpminusmax.cc",
4429 "src/xnnpack/AlignedAllocator.h",
4430 ] + MICROKERNEL_BENCHMARK_HDRS,
4431 deps = MICROKERNEL_BENCHMARK_DEPS,
4432)
4433
4434xnnpack_benchmark(
4435 name = "f32_raddextexp_bench",
4436 srcs = [
4437 "bench/f32-raddextexp.cc",
4438 "src/xnnpack/AlignedAllocator.h",
4439 ] + MICROKERNEL_BENCHMARK_HDRS,
4440 deps = MICROKERNEL_BENCHMARK_DEPS,
4441)
4442
4443xnnpack_benchmark(
4444 name = "f32_raddstoreexpminusmax_bench",
4445 srcs = [
4446 "bench/f32-raddstoreexpminusmax.cc",
4447 "src/xnnpack/AlignedAllocator.h",
4448 ] + MICROKERNEL_BENCHMARK_HDRS,
4449 deps = MICROKERNEL_BENCHMARK_DEPS,
4450)
4451
4452xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004453 name = "f32_relu_bench",
4454 srcs = [
4455 "bench/f32-relu.cc",
4456 "src/xnnpack/AlignedAllocator.h",
4457 ] + MICROKERNEL_BENCHMARK_HDRS,
4458 deps = MICROKERNEL_BENCHMARK_DEPS,
4459)
4460
4461xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004462 name = "f32_rmax_bench",
4463 srcs = [
4464 "bench/f32-rmax.cc",
4465 "src/xnnpack/AlignedAllocator.h",
4466 ] + MICROKERNEL_BENCHMARK_HDRS,
4467 deps = MICROKERNEL_BENCHMARK_DEPS,
4468)
4469
4470xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004471 name = "f32_sigmoid_bench",
4472 srcs = [
4473 "bench/f32-sigmoid.cc",
4474 "src/xnnpack/AlignedAllocator.h",
4475 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004476 deps = MICROKERNEL_BENCHMARK_DEPS,
4477)
4478
4479xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004480 name = "f32_spmm_bench",
4481 srcs = [
4482 "bench/f32-spmm.cc",
4483 "bench/gemm.h",
4484 "src/xnnpack/AlignedAllocator.h",
4485 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004486 deps = MICROKERNEL_BENCHMARK_DEPS,
4487)
4488
4489xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004490 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004491 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004492 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004493 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004494 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004495 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004496)
4497
4498xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004499 name = "f32_vscaleexpminusmax_bench",
4500 srcs = [
4501 "bench/f32-vscaleexpminusmax.cc",
4502 "src/xnnpack/AlignedAllocator.h",
4503 ] + MICROKERNEL_BENCHMARK_HDRS,
4504 deps = MICROKERNEL_BENCHMARK_DEPS,
4505)
4506
4507xnnpack_benchmark(
4508 name = "f32_vscaleextexp_bench",
4509 srcs = [
4510 "bench/f32-vscaleextexp.cc",
4511 "src/xnnpack/AlignedAllocator.h",
4512 ] + MICROKERNEL_BENCHMARK_HDRS,
4513 deps = MICROKERNEL_BENCHMARK_DEPS,
4514)
4515
4516xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004517 name = "f32_vsqrt_bench",
4518 srcs = [
4519 "bench/f32-vsqrt.cc",
4520 "src/xnnpack/AlignedAllocator.h",
4521 ] + MICROKERNEL_BENCHMARK_HDRS,
4522 deps = MICROKERNEL_BENCHMARK_DEPS,
4523)
4524
4525xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004526 name = "f32_im2col_gemm_bench",
4527 srcs = [
4528 "bench/f32-im2col-gemm.cc",
4529 "bench/conv.h",
4530 "src/xnnpack/AlignedAllocator.h",
4531 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004532 deps = MICROKERNEL_BENCHMARK_DEPS + [
4533 ":im2col",
4534 ":packing",
4535 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004536)
4537
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004538xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004539 name = "rounding_bench",
4540 srcs = [
4541 "bench/rounding.cc",
4542 "src/xnnpack/math-stubs.h",
4543 "src/xnnpack/AlignedAllocator.h",
4544 ] + MICROKERNEL_BENCHMARK_HDRS,
4545 deps = MICROKERNEL_BENCHMARK_DEPS,
4546)
4547
Marat Dukhan08c4a432019-10-03 09:29:21 -07004548########################### Benchmarks for operators ###########################
4549
4550xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004551 name = "average_pooling_bench",
4552 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004553 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004554 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004555 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004556)
4557
4558xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004559 name = "bankers_rounding_bench",
4560 srcs = ["bench/bankers-rounding.cc"],
4561 copts = xnnpack_optional_tflite_copts(),
4562 tags = ["nowin32"],
4563 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4564)
4565
4566xnnpack_benchmark(
4567 name = "ceiling_bench",
4568 srcs = ["bench/ceiling.cc"],
4569 copts = xnnpack_optional_tflite_copts(),
4570 tags = ["nowin32"],
4571 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4572)
4573
4574xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004575 name = "channel_shuffle_bench",
4576 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004577 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004578)
4579
4580xnnpack_benchmark(
4581 name = "convolution_bench",
4582 srcs = ["bench/convolution.cc"],
4583 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004584 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004585 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004586)
4587
4588xnnpack_benchmark(
4589 name = "deconvolution_bench",
4590 srcs = ["bench/deconvolution.cc"],
4591 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004592 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004593 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004594)
4595
4596xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004597 name = "floor_bench",
4598 srcs = ["bench/floor.cc"],
4599 copts = xnnpack_optional_tflite_copts(),
4600 tags = ["nowin32"],
4601 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4602)
4603
4604xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004605 name = "global_average_pooling_bench",
4606 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004607 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004608)
4609
4610xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004611 name = "hardswish_bench",
4612 srcs = ["bench/hardswish.cc"],
4613 copts = xnnpack_optional_tflite_copts(),
4614 tags = ["nowin32"],
4615 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4616)
4617
4618xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004619 name = "max_pooling_bench",
4620 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004621 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004622)
4623
4624xnnpack_benchmark(
4625 name = "sigmoid_bench",
4626 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004627 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004628 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004629 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004630)
4631
4632xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004633 name = "prelu_bench",
4634 srcs = ["bench/prelu.cc"],
4635 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004636 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004637 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004638)
4639
4640xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004641 name = "softmax_bench",
4642 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004643 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004644 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004645 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004646)
4647
Marat Dukhan87727142020-06-24 15:24:10 -07004648xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004649 name = "square_root_bench",
4650 srcs = ["bench/square-root.cc"],
4651 copts = xnnpack_optional_tflite_copts(),
4652 tags = ["nowin32"],
4653 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4654)
4655
4656xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004657 name = "truncation_bench",
4658 srcs = ["bench/truncation.cc"],
4659 deps = OPERATOR_BENCHMARK_DEPS,
4660)
4661
Marat Dukhanc068bb62019-10-04 13:24:39 -07004662############################# End-to-end benchmarks ############################
4663
4664cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004665 name = "fp32_mobilenet_v1",
4666 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004667 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004668 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004669 linkstatic = True,
4670 deps = [
4671 ":XNNPACK",
4672 "@pthreadpool",
4673 ],
4674)
4675
4676cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004677 name = "fp16_mobilenet_v1",
4678 srcs = ["models/fp16-mobilenet-v1.cc"],
4679 hdrs = ["models/models.h"],
4680 copts = xnnpack_std_cxxopts(),
4681 linkstatic = True,
4682 deps = [
4683 ":XNNPACK",
4684 "@FP16",
4685 "@pthreadpool",
4686 ],
4687)
4688
4689cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004690 name = "qs8_mobilenet_v1",
4691 srcs = ["models/qs8-mobilenet-v1.cc"],
4692 hdrs = ["models/models.h"],
4693 copts = xnnpack_std_cxxopts(),
4694 linkstatic = True,
4695 deps = [
4696 ":XNNPACK",
4697 "@pthreadpool",
4698 ],
4699)
4700
4701cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004702 name = "qs8_mobilenet_v2",
4703 srcs = ["models/qs8-mobilenet-v2.cc"],
4704 hdrs = ["models/models.h"],
4705 copts = xnnpack_std_cxxopts(),
4706 linkstatic = True,
4707 deps = [
4708 ":XNNPACK",
4709 "@pthreadpool",
4710 ],
4711)
4712
4713cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004714 name = "fp32_mobilenet_v2",
4715 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004716 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004717 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004718 linkstatic = True,
4719 deps = [
4720 ":XNNPACK",
4721 "@pthreadpool",
4722 ],
4723)
4724
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004725cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004726 name = "fp16_mobilenet_v2",
4727 srcs = ["models/fp16-mobilenet-v2.cc"],
4728 hdrs = ["models/models.h"],
4729 copts = xnnpack_std_cxxopts(),
4730 linkstatic = True,
4731 deps = [
4732 ":XNNPACK",
4733 "@FP16",
4734 "@pthreadpool",
4735 ],
4736)
4737
4738cc_library(
4739 name = "fp32_mobilenet_v3_large",
4740 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004741 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004742 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004743 linkstatic = True,
4744 deps = [
4745 ":XNNPACK",
4746 "@pthreadpool",
4747 ],
4748)
4749
4750cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004751 name = "fp16_mobilenet_v3_large",
4752 srcs = ["models/fp16-mobilenet-v3-large.cc"],
4753 hdrs = ["models/models.h"],
4754 copts = xnnpack_std_cxxopts(),
4755 linkstatic = True,
4756 deps = [
4757 ":XNNPACK",
4758 "@FP16",
4759 "@pthreadpool",
4760 ],
4761)
4762
4763cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004764 name = "fp32_mobilenet_v3_small",
4765 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004766 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004767 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004768 linkstatic = True,
4769 deps = [
4770 ":XNNPACK",
4771 "@pthreadpool",
4772 ],
4773)
4774
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004775cc_library(
4776 name = "fp16_mobilenet_v3_small",
4777 srcs = ["models/fp16-mobilenet-v3-small.cc"],
4778 hdrs = ["models/models.h"],
4779 copts = xnnpack_std_cxxopts(),
4780 linkstatic = True,
4781 deps = [
4782 ":XNNPACK",
4783 "@FP16",
4784 "@pthreadpool",
4785 ],
4786)
4787
Marat Dukhanc068bb62019-10-04 13:24:39 -07004788xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004789 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004790 srcs = [
4791 "bench/f32-dwconv-e2e.cc",
4792 "bench/end2end.h",
4793 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004794 deps = MICROKERNEL_BENCHMARK_DEPS + [
4795 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004796 ":fp32_mobilenet_v1",
4797 ":fp32_mobilenet_v2",
4798 ":fp32_mobilenet_v3_large",
4799 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004800 ],
4801)
4802
4803xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004804 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004805 srcs = [
4806 "bench/f32-gemm-e2e.cc",
4807 "bench/end2end.h",
4808 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004809 deps = MICROKERNEL_BENCHMARK_DEPS + [
4810 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004811 ":fp32_mobilenet_v1",
4812 ":fp32_mobilenet_v2",
4813 ":fp32_mobilenet_v3_large",
4814 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004815 ],
4816)
4817
4818xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004819 name = "end2end_bench",
4820 srcs = ["bench/end2end.cc"],
4821 deps = [
4822 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004823 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004824 ":fp16_mobilenet_v1",
4825 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004826 ":fp16_mobilenet_v3_large",
4827 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004828 ":fp32_mobilenet_v1",
4829 ":fp32_mobilenet_v2",
4830 ":fp32_mobilenet_v3_large",
4831 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004832 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07004833 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004834 "@pthreadpool",
4835 ],
4836)
4837
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004838#################### Accuracy evaluation for math functions ####################
4839
4840xnnpack_benchmark(
4841 name = "f32_exp_eval",
4842 srcs = [
4843 "eval/f32-exp.cc",
4844 "src/xnnpack/AlignedAllocator.h",
4845 ] + ACCURACY_EVAL_HDRS,
4846 deps = ACCURACY_EVAL_DEPS,
4847)
4848
Marat Dukhan515c9772019-10-17 18:07:57 -07004849xnnpack_benchmark(
4850 name = "f32_expminus_eval",
4851 srcs = [
4852 "eval/f32-expminus.cc",
4853 "src/xnnpack/AlignedAllocator.h",
4854 ] + ACCURACY_EVAL_HDRS,
4855 deps = ACCURACY_EVAL_DEPS,
4856)
4857
Marat Dukhan98ba4412019-10-23 02:14:28 -07004858xnnpack_benchmark(
4859 name = "f32_extexp_eval",
4860 srcs = [
4861 "eval/f32-extexp.cc",
4862 "src/xnnpack/AlignedAllocator.h",
4863 ] + ACCURACY_EVAL_HDRS,
4864 deps = ACCURACY_EVAL_DEPS,
4865)
4866
Marat Dukhan8853b822020-05-07 12:19:01 -07004867xnnpack_unit_test(
4868 name = "f32_roundne_eval",
4869 srcs = [
4870 "eval/f32-roundne.cc",
4871 "src/xnnpack/AlignedAllocator.h",
4872 "src/xnnpack/math-stubs.h",
4873 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004874 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004875 deps = MICROKERNEL_TEST_DEPS,
4876)
4877
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004878xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004879 name = "f32_roundd_eval",
4880 srcs = [
4881 "eval/f32-roundd.cc",
4882 "src/xnnpack/AlignedAllocator.h",
4883 "src/xnnpack/math-stubs.h",
4884 ] + MICROKERNEL_TEST_HDRS,
4885 automatic = False,
4886 deps = MICROKERNEL_TEST_DEPS,
4887)
4888
4889xnnpack_unit_test(
4890 name = "f32_roundu_eval",
4891 srcs = [
4892 "eval/f32-roundu.cc",
4893 "src/xnnpack/AlignedAllocator.h",
4894 "src/xnnpack/math-stubs.h",
4895 ] + MICROKERNEL_TEST_HDRS,
4896 automatic = False,
4897 deps = MICROKERNEL_TEST_DEPS,
4898)
4899
4900xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004901 name = "f32_roundz_eval",
4902 srcs = [
4903 "eval/f32-roundz.cc",
4904 "src/xnnpack/AlignedAllocator.h",
4905 "src/xnnpack/math-stubs.h",
4906 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004907 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004908 deps = MICROKERNEL_TEST_DEPS,
4909)
4910
Marat Dukhan346a9e52019-11-15 09:06:30 -08004911xnnpack_benchmark(
4912 name = "f32_sigmoid_eval",
4913 srcs = [
4914 "eval/f32-sigmoid.cc",
4915 "src/xnnpack/AlignedAllocator.h",
4916 ] + ACCURACY_EVAL_HDRS,
4917 deps = ACCURACY_EVAL_DEPS,
4918)
4919
Marat Dukhan84000762020-06-29 18:38:43 -07004920xnnpack_benchmark(
4921 name = "f32_sqrt_eval",
4922 srcs = [
4923 "eval/f32-sqrt.cc",
4924 "src/xnnpack/AlignedAllocator.h",
4925 ] + ACCURACY_EVAL_HDRS,
4926 deps = ACCURACY_EVAL_DEPS + [
4927 ":bench_utils",
4928 ],
4929)
4930
Marat Dukhan08c4a432019-10-03 09:29:21 -07004931######################### Unit tests for micro-kernels #########################
4932
4933xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004934 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004935 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004936 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004937 "test/gemm-microkernel-tester.h",
4938 "src/xnnpack/AlignedAllocator.h",
4939 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004940 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004941)
4942
4943xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004944 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004945 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004946 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004947 "test/spmm-microkernel-tester.h",
4948 "src/xnnpack/AlignedAllocator.h",
4949 ] + MICROKERNEL_TEST_HDRS,
4950 deps = MICROKERNEL_TEST_DEPS,
4951)
4952
4953xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004954 name = "f32_argmaxpool_test",
4955 srcs = [
4956 "test/f32-argmaxpool.cc",
4957 "test/argmaxpool-microkernel-tester.h",
4958 "src/xnnpack/AlignedAllocator.h",
4959 ] + MICROKERNEL_TEST_HDRS,
4960 deps = MICROKERNEL_TEST_DEPS,
4961)
4962
4963xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004964 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004965 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004966 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004967 "test/avgpool-microkernel-tester.h",
4968 "src/xnnpack/AlignedAllocator.h",
4969 ] + MICROKERNEL_TEST_HDRS,
4970 deps = MICROKERNEL_TEST_DEPS,
4971)
4972
4973xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004974 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004975 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004976 "test/f32-ibilinear.cc",
4977 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004978 "src/xnnpack/AlignedAllocator.h",
4979 ] + MICROKERNEL_TEST_HDRS,
4980 deps = MICROKERNEL_TEST_DEPS,
4981)
4982
4983xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004984 name = "f16_clamp_test",
4985 srcs = [
4986 "test/f16-clamp.cc",
4987 "test/clamp-microkernel-tester.h",
4988 ] + MICROKERNEL_TEST_HDRS,
4989 deps = MICROKERNEL_TEST_DEPS,
4990)
4991
4992xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004993 name = "f32_clamp_test",
4994 srcs = [
4995 "test/f32-clamp.cc",
4996 "test/clamp-microkernel-tester.h",
4997 ] + MICROKERNEL_TEST_HDRS,
4998 deps = MICROKERNEL_TEST_DEPS,
4999)
5000
5001xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005002 name = "f32_igemm_test",
5003 srcs = [
5004 "test/f32-igemm.cc",
5005 "test/gemm-microkernel-tester.h",
5006 "src/xnnpack/AlignedAllocator.h",
5007 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005008 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005009)
5010
5011xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005012 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005013 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005014 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005015 "test/gemm-microkernel-tester.h",
5016 "src/xnnpack/AlignedAllocator.h",
5017 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005018 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005019)
5020
5021xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005022 name = "f32_igemm_minmax_test",
5023 srcs = [
5024 "test/f32-igemm-minmax.cc",
5025 "test/gemm-microkernel-tester.h",
5026 "src/xnnpack/AlignedAllocator.h",
5027 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005028 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005029)
5030
5031xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005032 name = "f16_igemm_minmax_test",
5033 srcs = [
5034 "test/f16-igemm-minmax.cc",
5035 "test/gemm-microkernel-tester.h",
5036 "src/xnnpack/AlignedAllocator.h",
5037 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005038 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005039)
5040
5041xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005042 name = "f32_conv_hwc_test",
5043 srcs = [
5044 "test/f32-conv-hwc.cc",
5045 "test/conv-hwc-microkernel-tester.h",
5046 "src/xnnpack/AlignedAllocator.h",
5047 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005048 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005049)
5050
5051xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005052 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005053 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005054 "test/f32-conv-hwc2chw.cc",
5055 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005056 "src/xnnpack/AlignedAllocator.h",
5057 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005058 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005059)
5060
5061xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005062 name = "f32_dwconv_test",
5063 srcs = [
5064 "test/f32-dwconv.cc",
5065 "test/dwconv-microkernel-tester.h",
5066 "src/xnnpack/AlignedAllocator.h",
5067 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005068 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005069)
5070
5071xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005072 name = "f16_dwconv_minmax_test",
5073 srcs = [
5074 "test/f16-dwconv-minmax.cc",
5075 "test/dwconv-microkernel-tester.h",
5076 "src/xnnpack/AlignedAllocator.h",
5077 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005078 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005079)
5080
5081xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005082 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005083 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005084 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005085 "test/dwconv-microkernel-tester.h",
5086 "src/xnnpack/AlignedAllocator.h",
5087 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005088 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005089)
5090
5091xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005092 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005093 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005094 "test/f32-dwconv-chw.cc",
5095 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005096 "src/xnnpack/AlignedAllocator.h",
5097 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005098 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005099)
5100
5101xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005102 name = "f16_gavgpool_minmax_test",
5103 srcs = [
5104 "test/f16-gavgpool-minmax.cc",
5105 "test/gavgpool-microkernel-tester.h",
5106 "src/xnnpack/AlignedAllocator.h",
5107 ] + MICROKERNEL_TEST_HDRS,
5108 deps = MICROKERNEL_TEST_DEPS,
5109)
5110
5111xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005112 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005113 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005114 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005115 "test/gavgpool-microkernel-tester.h",
5116 "src/xnnpack/AlignedAllocator.h",
5117 ] + MICROKERNEL_TEST_HDRS,
5118 deps = MICROKERNEL_TEST_DEPS,
5119)
5120
5121xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005122 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005123 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005124 "test/f32-gavgpool-cw.cc",
5125 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005126 "src/xnnpack/AlignedAllocator.h",
5127 ] + MICROKERNEL_TEST_HDRS,
5128 deps = MICROKERNEL_TEST_DEPS,
5129)
5130
5131xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005132 name = "f32_gemm_test",
5133 srcs = [
5134 "test/f32-gemm.cc",
5135 "test/gemm-microkernel-tester.h",
5136 "src/xnnpack/AlignedAllocator.h",
5137 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005138 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005139)
5140
5141xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005142 name = "f32_gemm_relu_test",
5143 srcs = [
5144 "test/f32-gemm-relu.cc",
5145 "test/gemm-microkernel-tester.h",
5146 "src/xnnpack/AlignedAllocator.h",
5147 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005148 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005149)
5150
5151xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005152 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005153 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005154 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005155 "test/gemm-microkernel-tester.h",
5156 "src/xnnpack/AlignedAllocator.h",
5157 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005158 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005159)
5160
5161xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005162 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005163 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005164 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005165 "test/gemm-microkernel-tester.h",
5166 "src/xnnpack/AlignedAllocator.h",
5167 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005168 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005169)
5170
5171xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005172 name = "f16_hswish_test",
5173 srcs = [
5174 "test/f16-hswish.cc",
5175 "test/hswish-microkernel-tester.h",
5176 ] + MICROKERNEL_TEST_HDRS,
5177 deps = MICROKERNEL_TEST_DEPS,
5178)
5179
5180xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005181 name = "f32_hswish_test",
5182 srcs = [
5183 "test/f32-hswish.cc",
5184 "test/hswish-microkernel-tester.h",
5185 ] + MICROKERNEL_TEST_HDRS,
5186 deps = MICROKERNEL_TEST_DEPS,
5187)
5188
5189xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005190 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005191 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005192 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005193 "test/maxpool-microkernel-tester.h",
5194 ] + MICROKERNEL_TEST_HDRS,
5195 deps = MICROKERNEL_TEST_DEPS,
5196)
5197
5198xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005199 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005200 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005201 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005202 "test/avgpool-microkernel-tester.h",
5203 "src/xnnpack/AlignedAllocator.h",
5204 ] + MICROKERNEL_TEST_HDRS,
5205 deps = MICROKERNEL_TEST_DEPS,
5206)
5207
5208xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005209 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005210 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005211 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005212 "test/gemm-microkernel-tester.h",
5213 "src/xnnpack/AlignedAllocator.h",
5214 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005215 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005216)
5217
5218xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005219 name = "f16_prelu_test",
5220 srcs = [
5221 "test/f16-prelu.cc",
5222 "test/prelu-microkernel-tester.h",
5223 "src/xnnpack/AlignedAllocator.h",
5224 ] + MICROKERNEL_TEST_HDRS,
5225 deps = MICROKERNEL_TEST_DEPS,
5226)
5227
5228xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005229 name = "f32_prelu_test",
5230 srcs = [
5231 "test/f32-prelu.cc",
5232 "test/prelu-microkernel-tester.h",
5233 "src/xnnpack/AlignedAllocator.h",
5234 ] + MICROKERNEL_TEST_HDRS,
5235 deps = MICROKERNEL_TEST_DEPS,
5236)
5237
5238xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005239 name = "f32_raddexpminusmax_test",
5240 srcs = [
5241 "test/f32-raddexpminusmax.cc",
5242 "test/raddexpminusmax-microkernel-tester.h",
5243 ] + MICROKERNEL_TEST_HDRS,
5244 deps = MICROKERNEL_TEST_DEPS,
5245)
5246
5247xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005248 name = "f32_raddextexp_test",
5249 srcs = [
5250 "test/f32-raddextexp.cc",
5251 "test/raddextexp-microkernel-tester.h",
5252 ] + MICROKERNEL_TEST_HDRS,
5253 deps = MICROKERNEL_TEST_DEPS,
5254)
5255
5256xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005257 name = "f32_raddstoreexpminusmax_test",
5258 srcs = [
5259 "test/f32-raddstoreexpminusmax.cc",
5260 "test/raddstoreexpminusmax-microkernel-tester.h",
5261 ] + MICROKERNEL_TEST_HDRS,
5262 deps = MICROKERNEL_TEST_DEPS,
5263)
5264
5265xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005266 name = "f32_relu_test",
5267 srcs = [
5268 "test/f32-relu.cc",
5269 "test/vunary-microkernel-tester.h",
5270 ] + MICROKERNEL_TEST_HDRS,
5271 deps = MICROKERNEL_TEST_DEPS,
5272)
5273
5274xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005275 name = "f32_rmax_test",
5276 srcs = [
5277 "test/f32-rmax.cc",
5278 "test/rmax-microkernel-tester.h",
5279 ] + MICROKERNEL_TEST_HDRS,
5280 deps = MICROKERNEL_TEST_DEPS,
5281)
5282
5283xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005284 name = "f32_sigmoid_test",
5285 srcs = [
5286 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005287 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005288 ] + MICROKERNEL_TEST_HDRS,
5289 deps = MICROKERNEL_TEST_DEPS,
5290)
5291
5292xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005293 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005294 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005295 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005296 "test/spmm-microkernel-tester.h",
5297 "src/xnnpack/AlignedAllocator.h",
5298 ] + MICROKERNEL_TEST_HDRS,
5299 deps = MICROKERNEL_TEST_DEPS,
5300)
5301
5302xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005303 name = "f16_vadd_minmax_test",
5304 srcs = [
5305 "test/f16-vadd-minmax.cc",
5306 "test/vbinary-microkernel-tester.h",
5307 ] + MICROKERNEL_TEST_HDRS,
5308 deps = MICROKERNEL_TEST_DEPS,
5309)
5310
5311xnnpack_unit_test(
5312 name = "f16_vaddc_minmax_test",
5313 srcs = [
5314 "test/f16-vaddc-minmax.cc",
5315 "test/vbinaryc-microkernel-tester.h",
5316 ] + MICROKERNEL_TEST_HDRS,
5317 deps = MICROKERNEL_TEST_DEPS,
5318)
5319
5320xnnpack_unit_test(
5321 name = "f16_vdiv_minmax_test",
5322 srcs = [
5323 "test/f16-vdiv-minmax.cc",
5324 "test/vbinary-microkernel-tester.h",
5325 ] + MICROKERNEL_TEST_HDRS,
5326 deps = MICROKERNEL_TEST_DEPS,
5327)
5328
5329xnnpack_unit_test(
5330 name = "f16_vdivc_minmax_test",
5331 srcs = [
5332 "test/f16-vdivc-minmax.cc",
5333 "test/vbinaryc-microkernel-tester.h",
5334 ] + MICROKERNEL_TEST_HDRS,
5335 deps = MICROKERNEL_TEST_DEPS,
5336)
5337
5338xnnpack_unit_test(
5339 name = "f16_vrdivc_minmax_test",
5340 srcs = [
5341 "test/f16-vrdivc-minmax.cc",
5342 "test/vbinaryc-microkernel-tester.h",
5343 ] + MICROKERNEL_TEST_HDRS,
5344 deps = MICROKERNEL_TEST_DEPS,
5345)
5346
5347xnnpack_unit_test(
5348 name = "f16_vmax_test",
5349 srcs = [
5350 "test/f16-vmax.cc",
5351 "test/vbinary-microkernel-tester.h",
5352 ] + MICROKERNEL_TEST_HDRS,
5353 deps = MICROKERNEL_TEST_DEPS,
5354)
5355
5356xnnpack_unit_test(
5357 name = "f16_vmaxc_test",
5358 srcs = [
5359 "test/f16-vmaxc.cc",
5360 "test/vbinaryc-microkernel-tester.h",
5361 ] + MICROKERNEL_TEST_HDRS,
5362 deps = MICROKERNEL_TEST_DEPS,
5363)
5364
5365xnnpack_unit_test(
5366 name = "f16_vmin_test",
5367 srcs = [
5368 "test/f16-vmin.cc",
5369 "test/vbinary-microkernel-tester.h",
5370 ] + MICROKERNEL_TEST_HDRS,
5371 deps = MICROKERNEL_TEST_DEPS,
5372)
5373
5374xnnpack_unit_test(
5375 name = "f16_vminc_test",
5376 srcs = [
5377 "test/f16-vminc.cc",
5378 "test/vbinaryc-microkernel-tester.h",
5379 ] + MICROKERNEL_TEST_HDRS,
5380 deps = MICROKERNEL_TEST_DEPS,
5381)
5382
5383xnnpack_unit_test(
5384 name = "f16_vmul_minmax_test",
5385 srcs = [
5386 "test/f16-vmul-minmax.cc",
5387 "test/vbinary-microkernel-tester.h",
5388 ] + MICROKERNEL_TEST_HDRS,
5389 deps = MICROKERNEL_TEST_DEPS,
5390)
5391
5392xnnpack_unit_test(
5393 name = "f16_vmulc_minmax_test",
5394 srcs = [
5395 "test/f16-vmulc-minmax.cc",
5396 "test/vbinaryc-microkernel-tester.h",
5397 ] + MICROKERNEL_TEST_HDRS,
5398 deps = MICROKERNEL_TEST_DEPS,
5399)
5400
5401xnnpack_unit_test(
5402 name = "f16_vsub_minmax_test",
5403 srcs = [
5404 "test/f16-vsub-minmax.cc",
5405 "test/vbinary-microkernel-tester.h",
5406 ] + MICROKERNEL_TEST_HDRS,
5407 deps = MICROKERNEL_TEST_DEPS,
5408)
5409
5410xnnpack_unit_test(
5411 name = "f16_vsubc_minmax_test",
5412 srcs = [
5413 "test/f16-vsubc-minmax.cc",
5414 "test/vbinaryc-microkernel-tester.h",
5415 ] + MICROKERNEL_TEST_HDRS,
5416 deps = MICROKERNEL_TEST_DEPS,
5417)
5418
5419xnnpack_unit_test(
5420 name = "f16_vrsubc_minmax_test",
5421 srcs = [
5422 "test/f16-vrsubc-minmax.cc",
5423 "test/vbinaryc-microkernel-tester.h",
5424 ] + MICROKERNEL_TEST_HDRS,
5425 deps = MICROKERNEL_TEST_DEPS,
5426)
5427
5428xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005429 name = "f32_vabs_test",
5430 srcs = [
5431 "test/f32-vabs.cc",
5432 "test/vunary-microkernel-tester.h",
5433 ] + MICROKERNEL_TEST_HDRS,
5434 deps = MICROKERNEL_TEST_DEPS,
5435)
5436
5437xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005438 name = "f32_vadd_test",
5439 srcs = [
5440 "test/f32-vadd.cc",
5441 "test/vbinary-microkernel-tester.h",
5442 ] + MICROKERNEL_TEST_HDRS,
5443 deps = MICROKERNEL_TEST_DEPS,
5444)
5445
5446xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005447 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005448 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005449 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005450 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005451 ] + MICROKERNEL_TEST_HDRS,
5452 deps = MICROKERNEL_TEST_DEPS,
5453)
5454
5455xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005456 name = "f32_vadd_relu_test",
5457 srcs = [
5458 "test/f32-vadd-relu.cc",
5459 "test/vbinary-microkernel-tester.h",
5460 ] + MICROKERNEL_TEST_HDRS,
5461 deps = MICROKERNEL_TEST_DEPS,
5462)
5463
5464xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005465 name = "f32_vaddc_test",
5466 srcs = [
5467 "test/f32-vaddc.cc",
5468 "test/vbinaryc-microkernel-tester.h",
5469 ] + MICROKERNEL_TEST_HDRS,
5470 deps = MICROKERNEL_TEST_DEPS,
5471)
5472
5473xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005474 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005475 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005476 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005477 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005478 ] + MICROKERNEL_TEST_HDRS,
5479 deps = MICROKERNEL_TEST_DEPS,
5480)
5481
5482xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005483 name = "f32_vaddc_relu_test",
5484 srcs = [
5485 "test/f32-vaddc-relu.cc",
5486 "test/vbinaryc-microkernel-tester.h",
5487 ] + MICROKERNEL_TEST_HDRS,
5488 deps = MICROKERNEL_TEST_DEPS,
5489)
5490
5491xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005492 name = "f32_vdiv_test",
5493 srcs = [
5494 "test/f32-vdiv.cc",
5495 "test/vbinary-microkernel-tester.h",
5496 ] + MICROKERNEL_TEST_HDRS,
5497 deps = MICROKERNEL_TEST_DEPS,
5498)
5499
5500xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005501 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005502 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005503 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005504 "test/vbinary-microkernel-tester.h",
5505 ] + MICROKERNEL_TEST_HDRS,
5506 deps = MICROKERNEL_TEST_DEPS,
5507)
5508
5509xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005510 name = "f32_vdiv_relu_test",
5511 srcs = [
5512 "test/f32-vdiv-relu.cc",
5513 "test/vbinary-microkernel-tester.h",
5514 ] + MICROKERNEL_TEST_HDRS,
5515 deps = MICROKERNEL_TEST_DEPS,
5516)
5517
5518xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005519 name = "f32_vdivc_test",
5520 srcs = [
5521 "test/f32-vdivc.cc",
5522 "test/vbinaryc-microkernel-tester.h",
5523 ] + MICROKERNEL_TEST_HDRS,
5524 deps = MICROKERNEL_TEST_DEPS,
5525)
5526
5527xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005528 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005529 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005530 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005531 "test/vbinaryc-microkernel-tester.h",
5532 ] + MICROKERNEL_TEST_HDRS,
5533 deps = MICROKERNEL_TEST_DEPS,
5534)
5535
5536xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005537 name = "f32_vdivc_relu_test",
5538 srcs = [
5539 "test/f32-vdivc-relu.cc",
5540 "test/vbinaryc-microkernel-tester.h",
5541 ] + MICROKERNEL_TEST_HDRS,
5542 deps = MICROKERNEL_TEST_DEPS,
5543)
5544
5545xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005546 name = "f32_vrdivc_test",
5547 srcs = [
5548 "test/f32-vrdivc.cc",
5549 "test/vbinaryc-microkernel-tester.h",
5550 ] + MICROKERNEL_TEST_HDRS,
5551 deps = MICROKERNEL_TEST_DEPS,
5552)
5553
5554xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005555 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005556 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005557 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005558 "test/vbinaryc-microkernel-tester.h",
5559 ] + MICROKERNEL_TEST_HDRS,
5560 deps = MICROKERNEL_TEST_DEPS,
5561)
5562
5563xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005564 name = "f32_vrdivc_relu_test",
5565 srcs = [
5566 "test/f32-vrdivc-relu.cc",
5567 "test/vbinaryc-microkernel-tester.h",
5568 ] + MICROKERNEL_TEST_HDRS,
5569 deps = MICROKERNEL_TEST_DEPS,
5570)
5571
5572xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005573 name = "f32_vmax_test",
5574 srcs = [
5575 "test/f32-vmax.cc",
5576 "test/vbinary-microkernel-tester.h",
5577 ] + MICROKERNEL_TEST_HDRS,
5578 deps = MICROKERNEL_TEST_DEPS,
5579)
5580
5581xnnpack_unit_test(
5582 name = "f32_vmaxc_test",
5583 srcs = [
5584 "test/f32-vmaxc.cc",
5585 "test/vbinaryc-microkernel-tester.h",
5586 ] + MICROKERNEL_TEST_HDRS,
5587 deps = MICROKERNEL_TEST_DEPS,
5588)
5589
5590xnnpack_unit_test(
5591 name = "f32_vmin_test",
5592 srcs = [
5593 "test/f32-vmin.cc",
5594 "test/vbinary-microkernel-tester.h",
5595 ] + MICROKERNEL_TEST_HDRS,
5596 deps = MICROKERNEL_TEST_DEPS,
5597)
5598
5599xnnpack_unit_test(
5600 name = "f32_vminc_test",
5601 srcs = [
5602 "test/f32-vminc.cc",
5603 "test/vbinaryc-microkernel-tester.h",
5604 ] + MICROKERNEL_TEST_HDRS,
5605 deps = MICROKERNEL_TEST_DEPS,
5606)
5607
5608xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005609 name = "f32_vmul_test",
5610 srcs = [
5611 "test/f32-vmul.cc",
5612 "test/vbinary-microkernel-tester.h",
5613 ] + MICROKERNEL_TEST_HDRS,
5614 deps = MICROKERNEL_TEST_DEPS,
5615)
5616
5617xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005618 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005619 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005620 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005621 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005622 ] + MICROKERNEL_TEST_HDRS,
5623 deps = MICROKERNEL_TEST_DEPS,
5624)
5625
5626xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005627 name = "f32_vmul_relu_test",
5628 srcs = [
5629 "test/f32-vmul-relu.cc",
5630 "test/vbinary-microkernel-tester.h",
5631 ] + MICROKERNEL_TEST_HDRS,
5632 deps = MICROKERNEL_TEST_DEPS,
5633)
5634
5635xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005636 name = "f32_vmulc_test",
5637 srcs = [
5638 "test/f32-vmulc.cc",
5639 "test/vbinaryc-microkernel-tester.h",
5640 ] + MICROKERNEL_TEST_HDRS,
5641 deps = MICROKERNEL_TEST_DEPS,
5642)
5643
5644xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005645 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005646 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005647 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005648 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005649 ] + MICROKERNEL_TEST_HDRS,
5650 deps = MICROKERNEL_TEST_DEPS,
5651)
5652
5653xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005654 name = "f32_vmulc_relu_test",
5655 srcs = [
5656 "test/f32-vmulc-relu.cc",
5657 "test/vbinaryc-microkernel-tester.h",
5658 ] + MICROKERNEL_TEST_HDRS,
5659 deps = MICROKERNEL_TEST_DEPS,
5660)
5661
5662xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005663 name = "f16_vmulcaddc_minmax_test",
5664 srcs = [
5665 "test/f16-vmulcaddc-minmax.cc",
5666 "test/vmulcaddc-microkernel-tester.h",
5667 "src/xnnpack/AlignedAllocator.h",
5668 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005669 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005670)
5671
5672xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005673 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005674 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005675 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005676 "test/vmulcaddc-microkernel-tester.h",
5677 "src/xnnpack/AlignedAllocator.h",
5678 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005679 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005680)
5681
5682xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005683 name = "f32_vlrelu_test",
5684 srcs = [
5685 "test/f32-vlrelu.cc",
5686 "test/vunary-microkernel-tester.h",
5687 ] + MICROKERNEL_TEST_HDRS,
5688 deps = MICROKERNEL_TEST_DEPS,
5689)
5690
5691xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005692 name = "f32_vneg_test",
5693 srcs = [
5694 "test/f32-vneg.cc",
5695 "test/vunary-microkernel-tester.h",
5696 ] + MICROKERNEL_TEST_HDRS,
5697 deps = MICROKERNEL_TEST_DEPS,
5698)
5699
5700xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005701 name = "f32_vrndne_test",
5702 srcs = [
5703 "test/f32-vrndne.cc",
5704 "test/vunary-microkernel-tester.h",
5705 ] + MICROKERNEL_TEST_HDRS,
5706 deps = MICROKERNEL_TEST_DEPS,
5707)
5708
5709xnnpack_unit_test(
5710 name = "f32_vrndz_test",
5711 srcs = [
5712 "test/f32-vrndz.cc",
5713 "test/vunary-microkernel-tester.h",
5714 ] + MICROKERNEL_TEST_HDRS,
5715 deps = MICROKERNEL_TEST_DEPS,
5716)
5717
5718xnnpack_unit_test(
5719 name = "f32_vrndu_test",
5720 srcs = [
5721 "test/f32-vrndu.cc",
5722 "test/vunary-microkernel-tester.h",
5723 ] + MICROKERNEL_TEST_HDRS,
5724 deps = MICROKERNEL_TEST_DEPS,
5725)
5726
5727xnnpack_unit_test(
5728 name = "f32_vrndd_test",
5729 srcs = [
5730 "test/f32-vrndd.cc",
5731 "test/vunary-microkernel-tester.h",
5732 ] + MICROKERNEL_TEST_HDRS,
5733 deps = MICROKERNEL_TEST_DEPS,
5734)
5735
5736xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005737 name = "f32_vscale_test",
5738 srcs = [
5739 "test/f32-vscale.cc",
5740 "test/vscale-microkernel-tester.h",
5741 ] + MICROKERNEL_TEST_HDRS,
5742 deps = MICROKERNEL_TEST_DEPS,
5743)
5744
5745xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005746 name = "f32_vscaleexpminusmax_test",
5747 srcs = [
5748 "test/f32-vscaleexpminusmax.cc",
5749 "test/vscaleexpminusmax-microkernel-tester.h",
5750 ] + MICROKERNEL_TEST_HDRS,
5751 deps = MICROKERNEL_TEST_DEPS,
5752)
5753
5754xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005755 name = "f32_vscaleextexp_test",
5756 srcs = [
5757 "test/f32-vscaleextexp.cc",
5758 "test/vscaleextexp-microkernel-tester.h",
5759 ] + MICROKERNEL_TEST_HDRS,
5760 deps = MICROKERNEL_TEST_DEPS,
5761)
5762
5763xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005764 name = "f32_vsqr_test",
5765 srcs = [
5766 "test/f32-vsqr.cc",
5767 "test/vunary-microkernel-tester.h",
5768 ] + MICROKERNEL_TEST_HDRS,
5769 deps = MICROKERNEL_TEST_DEPS,
5770)
5771
5772xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005773 name = "f32_vsqrdiff_test",
5774 srcs = [
5775 "test/f32-vsqrdiff.cc",
5776 "test/vbinary-microkernel-tester.h",
5777 ] + MICROKERNEL_TEST_HDRS,
5778 deps = MICROKERNEL_TEST_DEPS,
5779)
5780
5781xnnpack_unit_test(
5782 name = "f32_vsqrdiffc_test",
5783 srcs = [
5784 "test/f32-vsqrdiffc.cc",
5785 "test/vbinaryc-microkernel-tester.h",
5786 ] + MICROKERNEL_TEST_HDRS,
5787 deps = MICROKERNEL_TEST_DEPS,
5788)
5789
5790xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005791 name = "f32_vsqrt_test",
5792 srcs = [
5793 "test/f32-vsqrt.cc",
5794 "test/vunary-microkernel-tester.h",
5795 ] + MICROKERNEL_TEST_HDRS,
5796 deps = MICROKERNEL_TEST_DEPS,
5797)
5798
5799xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005800 name = "f32_vsub_test",
5801 srcs = [
5802 "test/f32-vsub.cc",
5803 "test/vbinary-microkernel-tester.h",
5804 ] + MICROKERNEL_TEST_HDRS,
5805 deps = MICROKERNEL_TEST_DEPS,
5806)
5807
5808xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005809 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005810 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005811 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005812 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005813 ] + MICROKERNEL_TEST_HDRS,
5814 deps = MICROKERNEL_TEST_DEPS,
5815)
5816
5817xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005818 name = "f32_vsub_relu_test",
5819 srcs = [
5820 "test/f32-vsub-relu.cc",
5821 "test/vbinary-microkernel-tester.h",
5822 ] + MICROKERNEL_TEST_HDRS,
5823 deps = MICROKERNEL_TEST_DEPS,
5824)
5825
5826xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005827 name = "f32_vsubc_test",
5828 srcs = [
5829 "test/f32-vsubc.cc",
5830 "test/vbinaryc-microkernel-tester.h",
5831 ] + MICROKERNEL_TEST_HDRS,
5832 deps = MICROKERNEL_TEST_DEPS,
5833)
5834
5835xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005836 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005837 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005838 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005839 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005840 ] + MICROKERNEL_TEST_HDRS,
5841 deps = MICROKERNEL_TEST_DEPS,
5842)
5843
5844xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005845 name = "f32_vsubc_relu_test",
5846 srcs = [
5847 "test/f32-vsubc-relu.cc",
5848 "test/vbinaryc-microkernel-tester.h",
5849 ] + MICROKERNEL_TEST_HDRS,
5850 deps = MICROKERNEL_TEST_DEPS,
5851)
5852
5853xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005854 name = "f32_vrsubc_test",
5855 srcs = [
5856 "test/f32-vrsubc.cc",
5857 "test/vbinaryc-microkernel-tester.h",
5858 ] + MICROKERNEL_TEST_HDRS,
5859 deps = MICROKERNEL_TEST_DEPS,
5860)
5861
5862xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005863 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005864 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005865 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005866 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005867 ] + MICROKERNEL_TEST_HDRS,
5868 deps = MICROKERNEL_TEST_DEPS,
5869)
5870
5871xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005872 name = "f32_vrsubc_relu_test",
5873 srcs = [
5874 "test/f32-vrsubc-relu.cc",
5875 "test/vbinaryc-microkernel-tester.h",
5876 ] + MICROKERNEL_TEST_HDRS,
5877 deps = MICROKERNEL_TEST_DEPS,
5878)
5879
5880xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005881 name = "qs8_dwconv_minmax_test",
5882 srcs = [
5883 "test/qs8-dwconv-minmax.cc",
5884 "test/dwconv-microkernel-tester.h",
5885 "src/xnnpack/AlignedAllocator.h",
5886 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5887 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5888)
5889
5890xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005891 name = "qs8_gavgpool_minmax_test",
5892 srcs = [
5893 "test/qs8-gavgpool-minmax.cc",
5894 "test/gavgpool-microkernel-tester.h",
5895 "src/xnnpack/AlignedAllocator.h",
5896 ] + MICROKERNEL_TEST_HDRS,
5897 deps = MICROKERNEL_TEST_DEPS,
5898)
5899
5900xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005901 name = "qs8_gemm_minmax_test",
5902 srcs = [
5903 "test/qs8-gemm-minmax.cc",
5904 "test/gemm-microkernel-tester.h",
5905 "src/xnnpack/AlignedAllocator.h",
5906 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5907 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5908)
5909
5910xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005911 name = "qs8_igemm_minmax_test",
5912 srcs = [
5913 "test/qs8-igemm-minmax.cc",
5914 "test/gemm-microkernel-tester.h",
5915 "src/xnnpack/AlignedAllocator.h",
5916 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5917 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5918)
5919
5920xnnpack_unit_test(
5921 name = "qs8_requantization_test",
5922 srcs = [
5923 "src/xnnpack/requantization-stubs.h",
5924 "test/qs8-requantization.cc",
5925 "test/requantization-tester.h",
5926 ] + MICROKERNEL_TEST_HDRS,
5927 deps = MICROKERNEL_TEST_DEPS,
5928)
5929
5930xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07005931 name = "qs8_vadd_minmax_test",
5932 srcs = [
5933 "test/qs8-vadd-minmax.cc",
5934 "test/vadd-microkernel-tester.h",
5935 ] + MICROKERNEL_TEST_HDRS,
5936 deps = MICROKERNEL_TEST_DEPS,
5937)
5938
5939xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07005940 name = "qs8_vaddc_minmax_test",
5941 srcs = [
5942 "test/qs8-vaddc-minmax.cc",
5943 "test/vaddc-microkernel-tester.h",
5944 ] + MICROKERNEL_TEST_HDRS,
5945 deps = MICROKERNEL_TEST_DEPS,
5946)
5947
5948xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005949 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005950 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005951 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005952 "test/avgpool-microkernel-tester.h",
5953 "src/xnnpack/AlignedAllocator.h",
5954 ] + MICROKERNEL_TEST_HDRS,
5955 deps = MICROKERNEL_TEST_DEPS,
5956)
5957
5958xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005959 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005960 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005961 "test/qu8-dwconv-minmax.cc",
5962 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005963 "src/xnnpack/AlignedAllocator.h",
5964 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005965 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005966)
5967
5968xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005969 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005970 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005971 "test/qu8-igemm-minmax.cc",
5972 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005973 "src/xnnpack/AlignedAllocator.h",
5974 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005975 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005976)
5977
5978xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005979 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005980 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005981 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005982 "test/gavgpool-microkernel-tester.h",
5983 "src/xnnpack/AlignedAllocator.h",
5984 ] + MICROKERNEL_TEST_HDRS,
5985 deps = MICROKERNEL_TEST_DEPS,
5986)
5987
5988xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005989 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005990 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005991 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005992 "test/gemm-microkernel-tester.h",
5993 "src/xnnpack/AlignedAllocator.h",
5994 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005995 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005996)
5997
5998xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005999 name = "qu8_requantization_test",
6000 srcs = [
6001 "src/xnnpack/requantization-stubs.h",
6002 "test/qu8-requantization.cc",
6003 "test/requantization-tester.h",
6004 ] + MICROKERNEL_TEST_HDRS,
6005 deps = MICROKERNEL_TEST_DEPS,
6006)
6007
6008xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006009 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006010 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006011 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006012 "test/vadd-microkernel-tester.h",
6013 ] + MICROKERNEL_TEST_HDRS,
6014 deps = MICROKERNEL_TEST_DEPS,
6015)
6016
6017xnnpack_unit_test(
6018 name = "u8_clamp_test",
6019 srcs = [
6020 "test/u8-clamp.cc",
6021 "test/clamp-microkernel-tester.h",
6022 ] + MICROKERNEL_TEST_HDRS,
6023 deps = MICROKERNEL_TEST_DEPS,
6024)
6025
6026xnnpack_unit_test(
6027 name = "u8_lut32norm_test",
6028 srcs = [
6029 "test/u8-lut32norm.cc",
6030 "test/lut-norm-microkernel-tester.h",
6031 ] + MICROKERNEL_TEST_HDRS,
6032 deps = MICROKERNEL_TEST_DEPS,
6033)
6034
6035xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006036 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006037 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006038 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006039 "test/maxpool-microkernel-tester.h",
6040 ] + MICROKERNEL_TEST_HDRS,
6041 deps = MICROKERNEL_TEST_DEPS,
6042)
6043
6044xnnpack_unit_test(
6045 name = "u8_rmax_test",
6046 srcs = [
6047 "test/u8-rmax.cc",
6048 "test/rmax-microkernel-tester.h",
6049 ] + MICROKERNEL_TEST_HDRS,
6050 deps = MICROKERNEL_TEST_DEPS,
6051)
6052
6053xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006054 name = "x32_fill_test",
6055 srcs = [
6056 "test/x32-fill.cc",
6057 "test/fill-microkernel-tester.h",
6058 ] + MICROKERNEL_TEST_HDRS,
6059 deps = MICROKERNEL_TEST_DEPS,
6060)
6061
6062xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006063 name = "x32_packx_test",
6064 srcs = [
6065 "test/x32-packx.cc",
6066 "test/pack-microkernel-tester.h",
6067 "src/xnnpack/AlignedAllocator.h",
6068 ] + MICROKERNEL_TEST_HDRS,
6069 deps = MICROKERNEL_TEST_DEPS,
6070)
6071
6072xnnpack_unit_test(
6073 name = "x32_pad_test",
6074 srcs = [
6075 "test/x32-pad.cc",
6076 "test/pad-microkernel-tester.h",
6077 ] + MICROKERNEL_TEST_HDRS,
6078 deps = MICROKERNEL_TEST_DEPS,
6079)
6080
6081xnnpack_unit_test(
6082 name = "x32_unpool_test",
6083 srcs = [
6084 "test/x32-unpool.cc",
6085 "test/unpool-microkernel-tester.h",
6086 ] + MICROKERNEL_TEST_HDRS,
6087 deps = MICROKERNEL_TEST_DEPS,
6088)
6089
6090xnnpack_unit_test(
6091 name = "x32_zip_test",
6092 srcs = [
6093 "test/x32-zip.cc",
6094 "test/zip-microkernel-tester.h",
6095 ] + MICROKERNEL_TEST_HDRS,
6096 deps = MICROKERNEL_TEST_DEPS,
6097)
6098
6099xnnpack_unit_test(
6100 name = "x8_lut_test",
6101 srcs = [
6102 "test/x8-lut.cc",
6103 "test/lut-microkernel-tester.h",
6104 ] + MICROKERNEL_TEST_HDRS,
6105 deps = MICROKERNEL_TEST_DEPS,
6106)
6107
6108xnnpack_unit_test(
6109 name = "x8_zip_test",
6110 srcs = [
6111 "test/x8-zip.cc",
6112 "test/zip-microkernel-tester.h",
6113 ] + MICROKERNEL_TEST_HDRS,
6114 deps = MICROKERNEL_TEST_DEPS,
6115)
6116
Marat Dukhan20c3b922020-03-10 03:45:06 -07006117########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006118
6119xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006120 name = "operator_size_test",
6121 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006122 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006123)
6124
Marat Dukhan20c3b922020-03-10 03:45:06 -07006125xnnpack_binary(
6126 name = "subgraph_size_test",
6127 srcs = ["test/subgraph-size.c"],
6128 deps = [":XNNPACK"],
6129)
6130
6131########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006132
6133xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006134 name = "abs_nc_test",
6135 srcs = [
6136 "test/abs-nc.cc",
6137 "test/abs-operator-tester.h",
6138 ],
6139 deps = OPERATOR_TEST_DEPS,
6140)
6141
6142xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006143 name = "add_nd_test",
6144 srcs = [
6145 "test/add-nd.cc",
6146 "test/binary-elementwise-operator-tester.h",
6147 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006148 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006149)
6150
6151xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006152 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006153 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006154 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006155 "test/argmax-pooling-operator-tester.h",
6156 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006157 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006158)
6159
6160xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006161 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006162 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006163 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006164 "test/average-pooling-operator-tester.h",
6165 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006166 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006167)
6168
6169xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006170 name = "bankers_rounding_nc_test",
6171 srcs = [
6172 "test/bankers-rounding-nc.cc",
6173 "test/bankers-rounding-operator-tester.h",
6174 ],
6175 deps = OPERATOR_TEST_DEPS,
6176)
6177
6178xnnpack_unit_test(
6179 name = "ceiling_nc_test",
6180 srcs = [
6181 "test/ceiling-nc.cc",
6182 "test/ceiling-operator-tester.h",
6183 ],
6184 deps = OPERATOR_TEST_DEPS,
6185)
6186
6187xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006188 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006189 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006190 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006191 "test/channel-shuffle-operator-tester.h",
6192 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006193 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006194)
6195
6196xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006197 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006198 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006199 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006200 "test/clamp-operator-tester.h",
6201 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006202 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006203)
6204
6205xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006206 name = "constant_pad_nd_test",
6207 srcs = [
6208 "test/constant-pad-nd.cc",
6209 "test/constant-pad-operator-tester.h",
6210 ],
6211 deps = OPERATOR_TEST_DEPS,
6212)
6213
6214xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006215 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006216 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006217 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006218 "test/convolution-operator-tester.h",
6219 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006220 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006221)
6222
6223xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006224 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006225 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006226 "test/convolution-nchw.cc",
6227 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006228 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006229 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006230)
6231
6232xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006233 name = "copy_nc_test",
6234 srcs = [
6235 "test/copy-nc.cc",
6236 "test/copy-operator-tester.h",
6237 ],
6238 deps = OPERATOR_TEST_DEPS,
6239)
6240
6241xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006242 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006243 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006244 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006245 "test/deconvolution-operator-tester.h",
6246 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006247 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006248)
6249
6250xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006251 name = "divide_nd_test",
6252 srcs = [
6253 "test/binary-elementwise-operator-tester.h",
6254 "test/divide-nd.cc",
6255 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006256 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006257)
6258
6259xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006260 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006261 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006262 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006263 "test/fully-connected-operator-tester.h",
6264 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006265 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006266)
6267
6268xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006269 name = "floor_nc_test",
6270 srcs = [
6271 "test/floor-nc.cc",
6272 "test/floor-operator-tester.h",
6273 ],
6274 deps = OPERATOR_TEST_DEPS,
6275)
6276
6277xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006278 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006279 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006280 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006281 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006282 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006283 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006284)
6285
6286xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006287 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006288 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006289 "test/global-average-pooling-ncw.cc",
6290 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006291 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006292 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006293)
6294
6295xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006296 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006297 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006298 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006299 "test/hardswish-operator-tester.h",
6300 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006301 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006302)
6303
6304xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006305 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006306 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006307 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006308 "test/leaky-relu-operator-tester.h",
6309 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006310 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006311)
6312
6313xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006314 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006315 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006316 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006317 "test/max-pooling-operator-tester.h",
6318 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006319 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006320)
6321
6322xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006323 name = "maximum_nd_test",
6324 srcs = [
6325 "test/binary-elementwise-operator-tester.h",
6326 "test/maximum-nd.cc",
6327 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006328 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006329)
6330
6331xnnpack_unit_test(
6332 name = "minimum_nd_test",
6333 srcs = [
6334 "test/binary-elementwise-operator-tester.h",
6335 "test/minimum-nd.cc",
6336 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006337 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006338)
6339
6340xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006341 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006342 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006343 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006344 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006345 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006346 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006347)
6348
6349xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006350 name = "negate_nc_test",
6351 srcs = [
6352 "test/negate-nc.cc",
6353 "test/negate-operator-tester.h",
6354 ],
6355 deps = OPERATOR_TEST_DEPS,
6356)
6357
6358xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006359 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006360 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006361 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006362 "test/prelu-operator-tester.h",
6363 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006364 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006365)
6366
6367xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006368 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006369 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006370 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006371 "test/resize-bilinear-operator-tester.h",
6372 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006373 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006374)
6375
6376xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006377 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006378 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006379 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006380 "test/sigmoid-operator-tester.h",
6381 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006382 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006383)
6384
6385xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006386 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006387 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006388 "test/softmax-nc.cc",
6389 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006390 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006391 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006392)
6393
6394xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006395 name = "square_nc_test",
6396 srcs = [
6397 "test/square-nc.cc",
6398 "test/square-operator-tester.h",
6399 ],
6400 deps = OPERATOR_TEST_DEPS,
6401)
6402
6403xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006404 name = "square_root_nc_test",
6405 srcs = [
6406 "test/square-root-nc.cc",
6407 "test/square-root-operator-tester.h",
6408 ],
6409 deps = OPERATOR_TEST_DEPS,
6410)
6411
6412xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006413 name = "squared_difference_nd_test",
6414 srcs = [
6415 "test/binary-elementwise-operator-tester.h",
6416 "test/squared-difference-nd.cc",
6417 ],
6418 deps = OPERATOR_TEST_DEPS,
6419)
6420
6421xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006422 name = "subtract_nd_test",
6423 srcs = [
6424 "test/binary-elementwise-operator-tester.h",
6425 "test/subtract-nd.cc",
6426 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006427 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006428)
6429
6430xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006431 name = "truncation_nc_test",
6432 srcs = [
6433 "test/truncation-nc.cc",
6434 "test/truncation-operator-tester.h",
6435 ],
6436 deps = OPERATOR_TEST_DEPS,
6437)
6438
6439xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006440 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006441 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006442 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006443 "test/unpooling-operator-tester.h",
6444 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006445 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006446)
6447
Chao Mei6ddfc602020-05-13 22:29:36 -07006448############################### Misc unit tests ###############################
6449
6450xnnpack_unit_test(
6451 name = "memory_planner_test",
6452 srcs = [
6453 "test/memory-planner-test.cc",
6454 ],
6455 deps = [
6456 ":XNNPACK",
6457 ":memory_planner",
6458 ],
6459)
6460
Marat Dukhan08c4a432019-10-03 09:29:21 -07006461############################# Build configurations #############################
6462
Marat Dukhanb8642352019-10-30 15:43:02 -07006463# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006464config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006465 name = "xnn_enable_assembly_explicit_true",
6466 define_values = {"xnn_enable_assembly": "true"},
6467)
6468
6469# Disables usage of assembly kernels.
6470config_setting(
6471 name = "xnn_enable_assembly_explicit_false",
6472 define_values = {"xnn_enable_assembly": "false"},
6473)
6474
Marat Dukhan9de90e02020-06-18 16:04:12 -07006475# Enables usage of sparse inference.
6476config_setting(
6477 name = "xnn_enable_sparse_explicit_true",
6478 define_values = {"xnn_enable_sparse": "true"},
6479)
6480
6481# Disables usage of sparse inference.
6482config_setting(
6483 name = "xnn_enable_sparse_explicit_false",
6484 define_values = {"xnn_enable_sparse": "false"},
6485)
6486
Marat Dukhan05702cf2020-03-26 15:41:33 -07006487# Disables usage of HMP-aware optimizations.
6488config_setting(
6489 name = "xnn_enable_hmp_explicit_false",
6490 define_values = {"xnn_enable_hmp": "false"},
6491)
6492
Chao Mei6ddfc602020-05-13 22:29:36 -07006493# Enable usage of optimized memory allocation
6494config_setting(
6495 name = "xnn_enable_memopt_explicit_true",
6496 define_values = {"xnnpack_enable_memopt": "true"},
6497)
6498
6499# Disable usage of optimized memory allocation
6500config_setting(
6501 name = "xnn_enable_memopt_explicit_false",
6502 define_values = {"xnnpack_enable_memopt": "false"},
6503)
6504
Marat Dukhanb8642352019-10-30 15:43:02 -07006505# Builds with -c dbg
6506config_setting(
6507 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006508 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006509 "compilation_mode": "dbg",
6510 },
6511)
6512
6513# Builds with -c opt
6514config_setting(
6515 name = "optimized_build",
6516 values = {
6517 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006518 },
6519)
6520
6521config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006522 name = "linux_k8",
6523 values = {"cpu": "k8"},
6524)
6525
6526config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006527 name = "linux_arm",
6528 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006529)
6530
6531config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006532 name = "linux_armeabi",
6533 values = {"cpu": "armeabi"},
6534)
6535
6536config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006537 name = "linux_armhf",
6538 values = {"cpu": "armhf"},
6539)
6540
6541config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006542 name = "linux_armv7a",
6543 values = {"cpu": "armv7a"},
6544)
6545
6546config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006547 name = "linux_aarch64",
6548 values = {"cpu": "aarch64"},
6549)
6550
6551config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006552 name = "android",
6553 values = {"crosstool_top": "//external:android/crosstool"},
6554)
6555
6556config_setting(
6557 name = "android_armv7",
6558 values = {
6559 "crosstool_top": "//external:android/crosstool",
6560 "cpu": "armeabi-v7a",
6561 },
6562)
6563
6564config_setting(
6565 name = "android_arm64",
6566 values = {
6567 "crosstool_top": "//external:android/crosstool",
6568 "cpu": "arm64-v8a",
6569 },
6570)
6571
6572config_setting(
6573 name = "android_x86",
6574 values = {
6575 "crosstool_top": "//external:android/crosstool",
6576 "cpu": "x86",
6577 },
6578)
6579
6580config_setting(
6581 name = "android_x86_64",
6582 values = {
6583 "crosstool_top": "//external:android/crosstool",
6584 "cpu": "x86_64",
6585 },
6586)
6587
6588config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006589 name = "windows_x86_64",
6590 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006591)
6592
6593config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006594 name = "windows_x86_64_clang",
6595 values = {
6596 "compiler": "clang-cl",
6597 "cpu": "x64_windows",
6598 },
6599)
6600
6601config_setting(
6602 name = "windows_x86_64_mingw",
6603 values = {
6604 "compiler": "mingw-gcc",
6605 "cpu": "x64_windows",
6606 },
6607)
6608
6609config_setting(
6610 name = "windows_x86_64_msys",
6611 values = {
6612 "compiler": "msys-gcc",
6613 "cpu": "x64_windows",
6614 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006615)
6616
6617config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006618 name = "macos_x86_64",
6619 values = {
6620 "apple_platform_type": "macos",
6621 "cpu": "darwin",
6622 },
6623)
6624
6625config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006626 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006627 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006628)
6629
6630config_setting(
6631 name = "emscripten_wasm",
6632 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006633 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006634 "cpu": "wasm",
6635 },
6636)
6637
6638config_setting(
6639 name = "emscripten_wasmsimd",
6640 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006641 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006642 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006643 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006644 },
6645)
6646
6647config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006648 name = "ios_armv7",
6649 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006650 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006651 "cpu": "ios_armv7",
6652 },
6653)
6654
6655config_setting(
6656 name = "ios_arm64",
6657 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006658 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006659 "cpu": "ios_arm64",
6660 },
6661)
6662
6663config_setting(
6664 name = "ios_arm64e",
6665 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006666 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006667 "cpu": "ios_arm64e",
6668 },
6669)
6670
6671config_setting(
6672 name = "ios_x86",
6673 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006674 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006675 "cpu": "ios_i386",
6676 },
6677)
6678
6679config_setting(
6680 name = "ios_x86_64",
6681 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006682 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006683 "cpu": "ios_x86_64",
6684 },
6685)
6686
6687config_setting(
6688 name = "watchos_armv7k",
6689 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006690 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006691 "cpu": "watchos_armv7k",
6692 },
6693)
6694
6695config_setting(
6696 name = "watchos_arm64_32",
6697 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006698 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006699 "cpu": "watchos_arm64_32",
6700 },
6701)
6702
6703config_setting(
6704 name = "watchos_x86",
6705 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006706 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006707 "cpu": "watchos_i386",
6708 },
6709)
6710
6711config_setting(
6712 name = "watchos_x86_64",
6713 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006714 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006715 "cpu": "watchos_x86_64",
6716 },
6717)
6718
6719config_setting(
6720 name = "tvos_arm64",
6721 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006722 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006723 "cpu": "tvos_arm64",
6724 },
6725)
6726
6727config_setting(
6728 name = "tvos_x86_64",
6729 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006730 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006731 "cpu": "tvos_x86_64",
6732 },
6733)