blob: 5e8251bc0555bd1be8871ed83514ae2fdd35a5a7 [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",
Frank Barchard146e9992020-10-13 13:38:54 -07001324 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001325 "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",
Frank Barchard146e9992020-10-13 13:38:54 -07001328 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001329 "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",
Marat Dukhanac5f1602020-10-15 21:00:45 -07002005 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2006 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -07002007 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
Marat Dukhanac5f1602020-10-15 21:00:45 -07002008 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2009 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002010]
2011
2012SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002013 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2014 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2015 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07002016 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2017 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2018 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002019 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2020 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2021 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002022 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
2023 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002024 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
2025 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002026 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
2027 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002028 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
2029 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002030 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002031 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
2032 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002033 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002034 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
2035 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
2036 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002037 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
2038 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002039 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
2040 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002041 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
2042 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002043 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002044 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
2045 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002046 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002047 "src/qs8-requantization/precise-ssse3.c",
2048 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002049 "src/qu8-requantization/precise-ssse3.c",
2050 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002051]
2052
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002053SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002054 "src/f32-prelu/gen/sse41-2x4.c",
2055 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002056 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2057 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2058 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2059 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2060 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2061 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002062 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002063 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002064 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002065 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002066 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2067 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002068 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2069 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002070 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2071 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2072 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2073 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2074 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2075 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2076 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2077 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002078 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2079 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2080 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07002081 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2082 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2083 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002084 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2085 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2086 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002087 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
2088 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002089 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
2090 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002091 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
2092 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002093 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
2094 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002095 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002096 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
2097 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002098 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002099 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
2100 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
2101 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002102 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
2103 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002104 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
2105 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002106 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
2107 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002108 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002109 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
2110 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002111 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002112 "src/qs8-requantization/fp32-sse4.c",
2113 "src/qs8-requantization/precise-sse4.c",
2114 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002115 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2116 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2117 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2118 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002119 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2120 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2121 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2122 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002123 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2124 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2125 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2126 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002127 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2128 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2129 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2130 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002131 "src/qu8-requantization/precise-sse4.c",
2132 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002133 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002134 "src/math/roundd-sse41.c",
2135 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002136 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002137]
2138
Marat Dukhan08c4a432019-10-03 09:29:21 -07002139AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002140 "src/f32-clamp/gen/avx-x8.c",
2141 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002142 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2143 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2144 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2145 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
2146 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2147 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2148 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2149 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
2150 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2151 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
2152 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2153 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
2154 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2155 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2156 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2157 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2158 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
2159 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2160 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2161 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2162 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2163 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
2164 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
2165 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
2166 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2167 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2168 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2169 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
2170 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
2171 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002172 "src/f32-hswish/gen/hswish-avx-x8.c",
2173 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002174 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
2175 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
2176 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
2177 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2178 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
2179 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2180 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
2181 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
2182 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002183 "src/f32-prelu/gen/avx-2x8.c",
2184 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002185 "src/f32-relu/gen/avx-x16.c",
2186 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002187 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002188 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2189 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2190 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2191 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2192 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2193 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2194 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2195 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2196 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2197 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2198 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2199 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2200 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2201 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2202 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2203 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2204 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2205 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2206 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2207 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002208 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2209 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2210 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2211 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2212 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2213 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2214 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2215 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002216 "src/f32-vbinary/gen/vmax-avx-x8.c",
2217 "src/f32-vbinary/gen/vmax-avx-x16.c",
2218 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2219 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2220 "src/f32-vbinary/gen/vmin-avx-x8.c",
2221 "src/f32-vbinary/gen/vmin-avx-x16.c",
2222 "src/f32-vbinary/gen/vminc-avx-x8.c",
2223 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002224 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2225 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2226 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2227 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2228 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2229 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2230 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2231 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002232 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2233 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2234 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2235 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002236 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2237 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2238 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2239 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002240 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2241 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002242 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002243 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2244 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2245 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2246 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2247 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2248 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2249 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2250 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002251 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2252 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002253 "src/f32-vunary/gen/vabs-avx-x8.c",
2254 "src/f32-vunary/gen/vabs-avx-x16.c",
2255 "src/f32-vunary/gen/vneg-avx-x8.c",
2256 "src/f32-vunary/gen/vneg-avx-x16.c",
2257 "src/f32-vunary/gen/vsqr-avx-x8.c",
2258 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhand08abcb2020-09-21 00:26:25 -07002259 "src/math/exp-avx-p5.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002260 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002261 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2262 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002263 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002264]
2265
Marat Dukhan1566fee2020-08-02 21:55:41 -07002266XOP_UKERNELS = [
2267 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
2268 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2269 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
2270 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002271 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
2272 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002273 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
2274 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002275 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002276 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
2277 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002278 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002279 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
2280 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
2281 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002282 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
2283 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2284 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
2285 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
2286 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
2287 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002288 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002289 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
2290 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002291 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002292 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2293 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2294 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2295 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2296 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2297 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2298 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2299 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002300]
2301
Marat Dukhanfda12b82019-11-21 12:27:59 -08002302FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002303 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2304 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2305 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2306 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
2307 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2308 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2309 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2310 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
2311 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2312 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
2313 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2314 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
2315 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2316 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2317 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2318 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2319 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2320 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
2321 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2322 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2323 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2324 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2325 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2326 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2327 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2328 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2329 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
2330 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
2331 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
2332 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2333 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2334 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
2335 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2336 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2337 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2338 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2339 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2340 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
2341 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
2342 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002343 "src/f32-hswish/gen/hswish-fma3-x8.c",
2344 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002345 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
2346 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2347 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2348 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2349 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2350 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2351 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2352 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2353 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2354 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2355 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2356 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2357 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2358 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002359 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2360 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2361 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2362 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2363 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2364 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2365 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2366 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002367 "src/math/sqrt-fma3-nr1fma.c",
2368 "src/math/sqrt-fma3-nr2fma.c",
2369 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002370]
2371
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002372AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002373 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2374 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2375 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2376 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2377 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2378 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2379 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2380 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2381 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2382 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2383 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2384 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2385 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2386 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2387 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2388 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2389 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2390 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2391 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2392 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2393 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2394 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2395 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2396 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2397 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2398 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2399 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2400 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2401 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2402 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2403 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2404 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2405 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2406 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2407 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2408 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002409 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2410 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2411 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2412 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2413 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2414 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2415 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2416 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2417 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2418 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2419 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2420 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2421 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2422 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2423 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2424 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2425 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2426 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2427 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2428 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2429 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2430 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2431 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2432 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2433 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2434 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2435 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2436 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2437 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2438 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002439 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2440 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2441 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2442 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2443 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2444 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2445 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2446 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2447 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2448 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2449 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2450 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2451 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2452 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2453 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2454 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2455 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2456 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2457 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2458 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2459 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2460 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2461 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2462 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan40135522020-08-07 01:21:00 -07002463 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
2464 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002465 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2466 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2467 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2468 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002469 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2470 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2471 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002472 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2473 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2474 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002475 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2476 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2477 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002478 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2479 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2480 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2481 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2482 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2483 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2484 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2485 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002486 "src/math/exp-avx2-p5.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07002487 "src/math/exp-avx2-lut8-p3-perm.c",
2488 "src/math/exp-avx2-lut8-p4-perm.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002489 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002490 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002491 "src/math/sigmoid-avx2-rr1-p5-div.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002492 "src/math/sigmoid-avx2-rr2-p5-div.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002493 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002494 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002495 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002496 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002497 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2498 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2499 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2500 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2501 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2502 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2503 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2504 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002505]
2506
Marat Dukhan08c4a432019-10-03 09:29:21 -07002507AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002508 "src/f32-clamp/gen/avx512f-x16.c",
2509 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002510 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2511 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2512 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2513 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2514 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2515 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2516 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2517 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2518 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2519 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2520 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2521 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2522 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2523 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2524 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2525 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2526 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2527 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2528 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2529 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2530 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2531 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2532 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2533 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002534 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2535 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002536 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2537 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2538 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2539 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2540 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2541 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002542 "src/f32-prelu/gen/avx512f-2x16.c",
2543 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002544 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2545 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2546 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2547 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2548 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2549 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2550 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2551 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2552 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2553 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2554 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2555 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2556 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2557 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2558 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2559 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2560 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2561 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2562 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2563 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2564 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2565 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2566 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2567 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2568 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2569 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2570 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2571 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2572 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2573 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2574 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2575 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2576 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2577 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2578 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2579 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002580 "src/f32-relu/gen/avx512f-x16.c",
2581 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002582 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002583 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2584 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2585 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2586 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2587 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2588 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2589 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2590 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2591 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2592 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2593 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2594 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2595 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2596 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2597 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2598 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
2599 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2600 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2601 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2602 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2603 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2604 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2605 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2606 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2607 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2608 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2609 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2610 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2611 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2612 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2613 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2614 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
2615 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2616 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2617 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2618 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2619 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2620 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2621 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2622 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2623 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2624 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2625 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2626 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2627 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2628 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2629 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2630 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002631 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2632 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2633 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2634 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2635 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2636 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2637 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2638 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002639 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2640 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2641 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2642 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2643 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2644 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2645 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2646 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002647 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2648 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2649 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2650 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2651 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2652 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2653 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2654 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002655 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2656 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2657 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2658 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002659 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2660 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2661 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2662 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002663 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2664 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002665 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002666 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2667 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2668 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2669 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2670 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2671 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2672 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2673 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2674 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2675 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2676 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2677 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2678 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2679 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2680 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2681 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2682 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2683 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2684 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2685 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2686 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2687 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2688 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2689 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002690 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2691 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2692 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2693 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2694 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2695 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2696 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2697 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002698 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2699 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2700 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2701 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2702 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2703 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2704 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2705 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002706 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2707 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2708 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2709 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2710 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2711 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002712 "src/math/exp-avx512f-p5-scalef.c",
2713 "src/math/exp-avx512f-p5.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07002714 "src/math/exp-avx512f-lut16-p3-perm.c",
Marat Dukhan80a8ac52020-09-21 17:00:27 -07002715 "src/math/exp-avx512f-lut16-p3-perm-scalef.c",
Marat Dukhan8e172942020-09-20 23:56:23 -07002716 "src/math/exp-avx512f-lut32-p2-perm2.c",
Marat Dukhan80a8ac52020-09-21 17:00:27 -07002717 "src/math/exp-avx512f-lut32-p2-perm2-scalef.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002718 "src/math/extexp-avx512f-p5.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002719 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002720 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002721 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002722 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002723 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan47ffc412020-09-20 23:30:28 -07002724 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002725 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
2726 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
2727 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
2728 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2729 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2730 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2731 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
2732 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
2733 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
2734 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2735 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
2736 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002737 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
2738 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
2739 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
2740 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
2741 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
2742 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002743 "src/math/sqrt-avx512f-nr1fma.c",
2744 "src/math/sqrt-avx512f-nr2fma.c",
2745 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002746]
2747
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002748AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002749 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2750 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002751 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2752 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2753 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2754 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2755 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2756 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2757 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2758 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2759]
2760
Frank Barchardbcedc082020-08-17 18:00:51 -07002761WASM32_ASM_UKERNELS = [
2762 "src/f32-relu/wasm_shr_x1.S",
2763 "src/f32-relu/wasm_shr_x2.S",
2764 "src/f32-relu/wasm_shr_x4.S",
2765]
2766
Marat Dukhan08c4a432019-10-03 09:29:21 -07002767AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002768 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002769 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002770 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2771 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002772 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002773 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002774 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002775 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002776 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2777 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002778 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2779 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2780 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2781 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002782]
2783
2784AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002785 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002786 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2787 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002788 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002789 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2790 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002791 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2792 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2793 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002794 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002795 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2796 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2797 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002798 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002799 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2800 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002801 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002802 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2803 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2804 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2805 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2806 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2807 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2808 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2809 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2810 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2811 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2812 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2813 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2814 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2815 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2816 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2817 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2818 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2819 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002820 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2821 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002822 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002823 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2824 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2825 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2826 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2827 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2828 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2829 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2830 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2831 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2832 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2833 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2834 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2835 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2836 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2837 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2838 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2839 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2840 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002841 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2842 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2843 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2844 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2845 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2846 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2847 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2848 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2849 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2850 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2851 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2852 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2853 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2854 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2855 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2856 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2857 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2858 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07002859 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07002860 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07002861 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002862]
2863
Marat Dukhan1b354632020-03-23 12:50:22 -07002864INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002865 "src/xnnpack/argmaxpool.h",
2866 "src/xnnpack/avgpool.h",
2867 "src/xnnpack/clamp.h",
2868 "src/xnnpack/common.h",
2869 "src/xnnpack/conv.h",
2870 "src/xnnpack/dwconv.h",
2871 "src/xnnpack/gavgpool.h",
2872 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002873 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002874 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002875 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002876 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002877 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002878 "src/xnnpack/lut.h",
2879 "src/xnnpack/math.h",
2880 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002881 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002882 "src/xnnpack/packx.h",
2883 "src/xnnpack/pad.h",
2884 "src/xnnpack/params.h",
2885 "src/xnnpack/pavgpool.h",
2886 "src/xnnpack/ppmm.h",
2887 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002888 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002889 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002890 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002891 "src/xnnpack/rmax.h",
2892 "src/xnnpack/scalar-utils.h",
2893 "src/xnnpack/spmm.h",
2894 "src/xnnpack/unpool.h",
2895 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002896 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002897 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002898 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002899 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002900 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002901 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002902 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002903]
2904
2905INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002906 "include/xnnpack.h",
2907 "src/xnnpack/allocator.h",
2908 "src/xnnpack/compute.h",
2909 "src/xnnpack/im2col.h",
2910 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002911 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002912 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002913 "src/xnnpack/operator.h",
2914 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002915 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002916 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002917 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002918 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002919]
2920
Marat Dukhan1b354632020-03-23 12:50:22 -07002921ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002922 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002923]
2924
Marat Dukhan1b354632020-03-23 12:50:22 -07002925MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002926 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002927 "include/xnnpack.h",
2928]
2929
Marat Dukhan1b354632020-03-23 12:50:22 -07002930MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002931 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002932 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002933 "src/xnnpack/requantization.h",
2934 "include/xnnpack.h",
2935]
2936
2937OPERATOR_TEST_PARAMS_HDRS = [
2938 "src/xnnpack/params.h",
2939 "src/xnnpack/common.h",
2940]
2941
2942WEIGHTS_PACK_HDRS = [
2943 "src/xnnpack/pack.h",
2944 "src/xnnpack/operator.h",
2945 "src/xnnpack/compute.h",
2946]
2947
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002948LOGGING_COPTS = select({
2949 # No logging in optimized mode
2950 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2951 # Full logging in debug mode
2952 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2953 # Error-only logging in default (fastbuild) mode
2954 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2955})
2956
Marat Dukhan3b59de22020-06-03 20:15:19 -07002957LOGGING_SRCS = select({
2958 # No logging in optimized mode
2959 ":optimized_build": [],
2960 "//conditions:default": [
2961 "src/operator-strings.c",
2962 "src/subgraph-strings.c",
2963 ],
2964})
2965
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002966LOGGING_HDRS = [
2967 "src/xnnpack/log.h",
2968]
2969
Marat Dukhan08c4a432019-10-03 09:29:21 -07002970xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002971 name = "tables",
2972 srcs = TABLE_SRCS,
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 Dukhan3a77ea72019-12-23 12:10:24 -08002976)
2977
2978xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002979 name = "scalar_ukernels",
2980 srcs = SCALAR_UKERNELS,
2981 hdrs = INTERNAL_HDRS,
2982 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002983 gcc_copts = xnnpack_gcc_std_copts(),
2984 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002985 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002986 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002987 "@FP16",
2988 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002989 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002990 ],
2991)
2992
2993xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002994 name = "scalar_ukernels_test_mode",
2995 srcs = SCALAR_UKERNELS,
2996 hdrs = INTERNAL_HDRS,
2997 aarch32_copts = ["-marm"],
2998 copts = [
2999 "-UNDEBUG",
3000 "-DXNN_TEST_MODE=1",
3001 ],
3002 gcc_copts = xnnpack_gcc_std_copts(),
3003 msvc_copts = xnnpack_msvc_std_copts(),
3004 deps = [
3005 ":tables",
3006 "@FP16",
3007 "@FXdiv",
3008 "@pthreadpool",
3009 ],
3010)
3011
3012xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003013 name = "wasm_ukernels",
3014 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003015 gcc_copts = xnnpack_gcc_std_copts(),
3016 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003017 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003018 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003019 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003020 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003021 "@FP16",
3022 "@FXdiv",
3023 "@pthreadpool",
3024 ],
3025)
3026
3027xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003028 name = "wasm_ukernels_test_mode",
3029 hdrs = INTERNAL_HDRS,
3030 copts = [
3031 "-UNDEBUG",
3032 "-DXNN_TEST_MODE=1",
3033 ],
3034 gcc_copts = xnnpack_gcc_std_copts(),
3035 msvc_copts = xnnpack_msvc_std_copts(),
3036 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003037 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003038 deps = [
3039 ":tables",
3040 "@FP16",
3041 "@FXdiv",
3042 "@pthreadpool",
3043 ],
3044)
3045
3046xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003047 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003048 hdrs = INTERNAL_HDRS,
3049 aarch32_copts = [
3050 "-marm",
3051 "-mfpu=neon",
3052 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003053 gcc_copts = xnnpack_gcc_std_copts(),
3054 gcc_x86_copts = ["-msse2"],
3055 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003056 optimized_copts = [
3057 "-O3",
3058 "-ffast-math",
3059 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003060 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003061 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003062 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003063 "@FP16",
3064 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003065 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003066 ],
3067)
3068
3069xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003070 name = "psimd_fastmath_ukernels_test_mode",
3071 hdrs = INTERNAL_HDRS,
3072 aarch32_copts = [
3073 "-marm",
3074 "-mfpu=neon",
3075 ],
3076 copts = [
3077 "-UNDEBUG",
3078 "-DXNN_TEST_MODE=1",
3079 ],
3080 gcc_copts = xnnpack_gcc_std_copts(),
3081 gcc_x86_copts = ["-msse2"],
3082 msvc_copts = xnnpack_msvc_std_copts(),
3083 optimized_copts = [
3084 "-O3",
3085 "-ffast-math",
3086 ],
3087 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
3088 deps = [
3089 ":tables",
3090 "@FP16",
3091 "@psimd",
3092 "@pthreadpool",
3093 ],
3094)
3095
3096xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003097 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003098 hdrs = INTERNAL_HDRS,
3099 aarch32_copts = [
3100 "-marm",
3101 "-mfpu=neon",
3102 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003103 gcc_copts = xnnpack_gcc_std_copts(),
3104 gcc_x86_copts = ["-msse2"],
3105 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003106 optimized_copts = [
3107 "-O3",
3108 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003109 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003110 deps = [
3111 ":tables",
3112 "@FP16",
3113 "@psimd",
3114 "@pthreadpool",
3115 ],
3116)
3117
3118xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003119 name = "psimd_accmath_ukernels_test_mode",
3120 hdrs = INTERNAL_HDRS,
3121 aarch32_copts = [
3122 "-marm",
3123 "-mfpu=neon",
3124 ],
3125 copts = [
3126 "-UNDEBUG",
3127 "-DXNN_TEST_MODE=1",
3128 ],
3129 gcc_copts = xnnpack_gcc_std_copts(),
3130 gcc_x86_copts = ["-msse2"],
3131 msvc_copts = xnnpack_msvc_std_copts(),
3132 optimized_copts = [
3133 "-O3",
3134 ],
3135 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3136 deps = [
3137 ":tables",
3138 "@FP16",
3139 "@psimd",
3140 "@pthreadpool",
3141 ],
3142)
3143
3144xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003145 name = "neon_ukernels",
3146 hdrs = INTERNAL_HDRS,
3147 aarch32_copts = [
3148 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003149 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003150 "-mfpu=neon",
3151 ],
3152 aarch32_srcs = NEON_UKERNELS,
3153 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003154 gcc_copts = xnnpack_gcc_std_copts(),
3155 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003156 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003157 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003158 "@FP16",
3159 "@pthreadpool",
3160 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003161)
3162
3163xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003164 name = "neon_ukernels_test_mode",
3165 hdrs = INTERNAL_HDRS,
3166 aarch32_copts = [
3167 "-marm",
3168 "-march=armv7-a",
3169 "-mfpu=neon",
3170 ],
3171 aarch32_srcs = NEON_UKERNELS,
3172 aarch64_srcs = NEON_UKERNELS,
3173 copts = [
3174 "-UNDEBUG",
3175 "-DXNN_TEST_MODE=1",
3176 ],
3177 gcc_copts = xnnpack_gcc_std_copts(),
3178 msvc_copts = xnnpack_msvc_std_copts(),
3179 deps = [
3180 ":tables",
3181 "@FP16",
3182 "@pthreadpool",
3183 ],
3184)
3185
3186xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003187 name = "neonfma_ukernels",
3188 hdrs = INTERNAL_HDRS,
3189 aarch32_copts = [
3190 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003191 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003192 "-mfpu=neon-vfpv4",
3193 ],
3194 aarch32_srcs = NEONFMA_UKERNELS,
3195 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003196 apple_aarch32_copts = [
3197 "-mcpu=swift",
3198 "-mtune=generic",
3199 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003200 gcc_copts = xnnpack_gcc_std_copts(),
3201 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003202 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003203 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003204 "@FP16",
3205 "@pthreadpool",
3206 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003207)
3208
3209xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003210 name = "neonfma_ukernels_test_mode",
3211 hdrs = INTERNAL_HDRS,
3212 aarch32_copts = [
3213 "-marm",
3214 "-march=armv7-a",
3215 "-mfpu=neon-vfpv4",
3216 ],
3217 aarch32_srcs = NEONFMA_UKERNELS,
3218 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003219 apple_aarch32_copts = [
3220 "-mcpu=swift",
3221 "-mtune=generic",
3222 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003223 copts = [
3224 "-UNDEBUG",
3225 "-DXNN_TEST_MODE=1",
3226 ],
3227 gcc_copts = xnnpack_gcc_std_copts(),
3228 msvc_copts = xnnpack_msvc_std_copts(),
3229 deps = [
3230 ":tables",
3231 "@FP16",
3232 "@pthreadpool",
3233 ],
3234)
3235
3236xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003237 name = "neonv8_ukernels",
3238 hdrs = INTERNAL_HDRS,
3239 aarch32_copts = [
3240 "-marm",
3241 "-march=armv8-a",
3242 "-mfpu=neon-fp-armv8",
3243 ],
3244 aarch32_srcs = NEONV8_UKERNELS,
3245 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003246 apple_aarch32_copts = [
3247 "-mcpu=cyclone",
3248 "-mtune=generic",
3249 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003250 gcc_copts = xnnpack_gcc_std_copts(),
3251 msvc_copts = xnnpack_msvc_std_copts(),
3252 deps = [
3253 ":tables",
3254 "@FP16",
3255 "@pthreadpool",
3256 ],
3257)
3258
3259xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003260 name = "neonv8_ukernels_test_mode",
3261 hdrs = INTERNAL_HDRS,
3262 aarch32_copts = [
3263 "-marm",
3264 "-march=armv8-a",
3265 "-mfpu=neon-fp-armv8",
3266 ],
3267 aarch32_srcs = NEONV8_UKERNELS,
3268 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003269 apple_aarch32_copts = [
3270 "-mcpu=cyclone",
3271 "-mtune=generic",
3272 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003273 copts = [
3274 "-UNDEBUG",
3275 "-DXNN_TEST_MODE=1",
3276 ],
3277 gcc_copts = xnnpack_gcc_std_copts(),
3278 msvc_copts = xnnpack_msvc_std_copts(),
3279 deps = [
3280 ":tables",
3281 "@FP16",
3282 "@pthreadpool",
3283 ],
3284)
3285
3286xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003287 name = "neonfp16arith_ukernels",
3288 hdrs = INTERNAL_HDRS,
3289 aarch64_copts = ["-march=armv8.2-a+fp16"],
3290 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003291 gcc_copts = xnnpack_gcc_std_copts(),
3292 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003293 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003294 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003295 "@FP16",
3296 "@pthreadpool",
3297 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003298)
3299
3300xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003301 name = "neonfp16arith_ukernels_test_mode",
3302 hdrs = INTERNAL_HDRS,
3303 aarch64_copts = ["-march=armv8.2-a+fp16"],
3304 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3305 copts = [
3306 "-UNDEBUG",
3307 "-DXNN_TEST_MODE=1",
3308 ],
3309 gcc_copts = xnnpack_gcc_std_copts(),
3310 msvc_copts = xnnpack_msvc_std_copts(),
3311 deps = [
3312 ":tables",
3313 "@FP16",
3314 "@pthreadpool",
3315 ],
3316)
3317
3318xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003319 name = "neondot_ukernels",
3320 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003321 aarch32_copts = [
3322 "-marm",
3323 "-march=armv8.2-a+dotprod",
3324 "-mfpu=neon-fp-armv8",
3325 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003326 aarch32_srcs = NEONDOT_UKERNELS,
3327 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3328 aarch64_srcs = NEONDOT_UKERNELS,
3329 gcc_copts = xnnpack_gcc_std_copts(),
3330 msvc_copts = xnnpack_msvc_std_copts(),
3331 deps = [
3332 ":tables",
3333 "@FP16",
3334 "@pthreadpool",
3335 ],
3336)
3337
3338xnnpack_cc_library(
3339 name = "neondot_ukernels_test_mode",
3340 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003341 aarch32_copts = [
3342 "-marm",
3343 "-march=armv8.2-a+dotprod",
3344 "-mfpu=neon-fp-armv8",
3345 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003346 aarch32_srcs = NEONDOT_UKERNELS,
3347 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3348 aarch64_srcs = NEONDOT_UKERNELS,
3349 copts = [
3350 "-UNDEBUG",
3351 "-DXNN_TEST_MODE=1",
3352 ],
3353 gcc_copts = xnnpack_gcc_std_copts(),
3354 msvc_copts = xnnpack_msvc_std_copts(),
3355 deps = [
3356 ":tables",
3357 "@FP16",
3358 "@pthreadpool",
3359 ],
3360)
3361
3362xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003363 name = "sse2_ukernels",
3364 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003365 gcc_copts = xnnpack_gcc_std_copts(),
3366 gcc_x86_copts = ["-msse2"],
3367 msvc_copts = xnnpack_msvc_std_copts(),
3368 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003369 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003370 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003371 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003372 "@FP16",
3373 "@pthreadpool",
3374 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003375)
3376
3377xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003378 name = "sse2_ukernels_test_mode",
3379 hdrs = INTERNAL_HDRS,
3380 copts = [
3381 "-UNDEBUG",
3382 "-DXNN_TEST_MODE=1",
3383 ],
3384 gcc_copts = xnnpack_gcc_std_copts(),
3385 gcc_x86_copts = ["-msse2"],
3386 msvc_copts = xnnpack_msvc_std_copts(),
3387 msvc_x86_32_copts = ["/arch:SSE2"],
3388 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3389 deps = [
3390 ":tables",
3391 "@FP16",
3392 "@pthreadpool",
3393 ],
3394)
3395
3396xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003397 name = "ssse3_ukernels",
3398 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003399 gcc_copts = xnnpack_gcc_std_copts(),
3400 gcc_x86_copts = ["-mssse3"],
3401 msvc_copts = xnnpack_msvc_std_copts(),
3402 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003403 x86_srcs = SSSE3_UKERNELS,
3404 deps = [
3405 ":tables",
3406 "@FP16",
3407 "@pthreadpool",
3408 ],
3409)
3410
3411xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003412 name = "ssse3_ukernels_test_mode",
3413 hdrs = INTERNAL_HDRS,
3414 copts = [
3415 "-UNDEBUG",
3416 "-DXNN_TEST_MODE=1",
3417 ],
3418 gcc_copts = xnnpack_gcc_std_copts(),
3419 gcc_x86_copts = ["-mssse3"],
3420 msvc_copts = xnnpack_msvc_std_copts(),
3421 msvc_x86_32_copts = ["/arch:SSE2"],
3422 x86_srcs = SSSE3_UKERNELS,
3423 deps = [
3424 ":tables",
3425 "@FP16",
3426 "@pthreadpool",
3427 ],
3428)
3429
3430xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003431 name = "sse41_ukernels",
3432 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003433 gcc_copts = xnnpack_gcc_std_copts(),
3434 gcc_x86_copts = ["-msse4.1"],
3435 msvc_copts = xnnpack_msvc_std_copts(),
3436 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003437 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003438 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003439 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003440 "@FP16",
3441 "@pthreadpool",
3442 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003443)
3444
3445xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003446 name = "sse41_ukernels_test_mode",
3447 hdrs = INTERNAL_HDRS,
3448 copts = [
3449 "-UNDEBUG",
3450 "-DXNN_TEST_MODE=1",
3451 ],
3452 gcc_copts = xnnpack_gcc_std_copts(),
3453 gcc_x86_copts = ["-msse4.1"],
3454 msvc_copts = xnnpack_msvc_std_copts(),
3455 msvc_x86_32_copts = ["/arch:SSE2"],
3456 x86_srcs = SSE41_UKERNELS,
3457 deps = [
3458 ":tables",
3459 "@FP16",
3460 "@pthreadpool",
3461 ],
3462)
3463
3464xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003465 name = "avx_ukernels",
3466 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003467 gcc_copts = xnnpack_gcc_std_copts(),
3468 gcc_x86_copts = ["-mavx"],
3469 msvc_copts = xnnpack_msvc_std_copts(),
3470 msvc_x86_32_copts = ["/arch:AVX"],
3471 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003472 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003473 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003474 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003475 "@FP16",
3476 "@pthreadpool",
3477 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003478)
3479
3480xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003481 name = "avx_ukernels_test_mode",
3482 hdrs = INTERNAL_HDRS,
3483 copts = [
3484 "-UNDEBUG",
3485 "-DXNN_TEST_MODE=1",
3486 ],
3487 gcc_copts = xnnpack_gcc_std_copts(),
3488 gcc_x86_copts = ["-mavx"],
3489 msvc_copts = xnnpack_msvc_std_copts(),
3490 msvc_x86_32_copts = ["/arch:AVX"],
3491 msvc_x86_64_copts = ["/arch:AVX"],
3492 x86_srcs = AVX_UKERNELS,
3493 deps = [
3494 ":tables",
3495 "@FP16",
3496 "@pthreadpool",
3497 ],
3498)
3499
3500xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003501 name = "xop_ukernels",
3502 hdrs = INTERNAL_HDRS,
3503 gcc_copts = xnnpack_gcc_std_copts(),
3504 gcc_x86_copts = ["-mxop"],
3505 msvc_copts = xnnpack_msvc_std_copts(),
3506 msvc_x86_32_copts = ["/arch:AVX"],
3507 msvc_x86_64_copts = ["/arch:AVX"],
3508 x86_srcs = XOP_UKERNELS,
3509 deps = [
3510 ":tables",
3511 "@FP16",
3512 "@pthreadpool",
3513 ],
3514)
3515
3516xnnpack_cc_library(
3517 name = "xop_ukernels_test_mode",
3518 hdrs = INTERNAL_HDRS,
3519 copts = [
3520 "-UNDEBUG",
3521 "-DXNN_TEST_MODE=1",
3522 ],
3523 gcc_copts = xnnpack_gcc_std_copts(),
3524 gcc_x86_copts = ["-mxop"],
3525 msvc_copts = xnnpack_msvc_std_copts(),
3526 msvc_x86_32_copts = ["/arch:AVX"],
3527 msvc_x86_64_copts = ["/arch:AVX"],
3528 x86_srcs = XOP_UKERNELS,
3529 deps = [
3530 ":tables",
3531 "@FP16",
3532 "@pthreadpool",
3533 ],
3534)
3535
3536xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003537 name = "fma3_ukernels",
3538 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003539 gcc_copts = xnnpack_gcc_std_copts(),
3540 gcc_x86_copts = ["-mfma"],
3541 msvc_copts = xnnpack_msvc_std_copts(),
3542 msvc_x86_32_copts = ["/arch:AVX"],
3543 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003544 x86_srcs = FMA3_UKERNELS,
3545 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003546 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003547 "@FP16",
3548 "@pthreadpool",
3549 ],
3550)
3551
3552xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003553 name = "fma3_ukernels_test_mode",
3554 hdrs = INTERNAL_HDRS,
3555 copts = [
3556 "-UNDEBUG",
3557 "-DXNN_TEST_MODE=1",
3558 ],
3559 gcc_copts = xnnpack_gcc_std_copts(),
3560 gcc_x86_copts = ["-mfma"],
3561 msvc_copts = xnnpack_msvc_std_copts(),
3562 msvc_x86_32_copts = ["/arch:AVX"],
3563 msvc_x86_64_copts = ["/arch:AVX"],
3564 x86_srcs = FMA3_UKERNELS,
3565 deps = [
3566 ":tables",
3567 "@FP16",
3568 "@pthreadpool",
3569 ],
3570)
3571
3572xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003573 name = "avx2_ukernels",
3574 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003575 gcc_copts = xnnpack_gcc_std_copts(),
3576 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003577 "-mfma",
3578 "-mavx2",
3579 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003580 msvc_copts = xnnpack_msvc_std_copts(),
3581 msvc_x86_32_copts = ["/arch:AVX2"],
3582 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003583 x86_srcs = AVX2_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 Dukhan6adff4e2019-10-14 18:32:07 -07003589)
3590
3591xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003592 name = "avx2_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 = [
3600 "-mfma",
3601 "-mavx2",
3602 ],
3603 msvc_copts = xnnpack_msvc_std_copts(),
3604 msvc_x86_32_copts = ["/arch:AVX2"],
3605 msvc_x86_64_copts = ["/arch:AVX2"],
3606 x86_srcs = AVX2_UKERNELS,
3607 deps = [
3608 ":tables",
3609 "@FP16",
3610 "@pthreadpool",
3611 ],
3612)
3613
3614xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003615 name = "avx512f_ukernels",
3616 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003617 gcc_copts = xnnpack_gcc_std_copts(),
3618 gcc_x86_copts = ["-mavx512f"],
3619 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3620 msvc_copts = xnnpack_msvc_std_copts(),
3621 msvc_x86_32_copts = ["/arch:AVX512"],
3622 msvc_x86_64_copts = ["/arch:AVX512"],
3623 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003624 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003625 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003626 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003627 "@FP16",
3628 "@pthreadpool",
3629 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003630)
3631
3632xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003633 name = "avx512f_ukernels_test_mode",
3634 hdrs = INTERNAL_HDRS,
3635 copts = [
3636 "-UNDEBUG",
3637 "-DXNN_TEST_MODE=1",
3638 ],
3639 gcc_copts = xnnpack_gcc_std_copts(),
3640 gcc_x86_copts = ["-mavx512f"],
3641 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3642 msvc_copts = xnnpack_msvc_std_copts(),
3643 msvc_x86_32_copts = ["/arch:AVX512"],
3644 msvc_x86_64_copts = ["/arch:AVX512"],
3645 msys_copts = ["-fno-asynchronous-unwind-tables"],
3646 x86_srcs = AVX512F_UKERNELS,
3647 deps = [
3648 ":tables",
3649 "@FP16",
3650 "@pthreadpool",
3651 ],
3652)
3653
3654xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003655 name = "avx512skx_ukernels",
3656 hdrs = INTERNAL_HDRS,
3657 gcc_copts = xnnpack_gcc_std_copts(),
3658 gcc_x86_copts = [
3659 "-mavx512f",
3660 "-mavx512cd",
3661 "-mavx512bw",
3662 "-mavx512dq",
3663 "-mavx512vl",
3664 ],
3665 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3666 msvc_copts = xnnpack_msvc_std_copts(),
3667 msvc_x86_32_copts = ["/arch:AVX512"],
3668 msvc_x86_64_copts = ["/arch:AVX512"],
3669 msys_copts = ["-fno-asynchronous-unwind-tables"],
3670 x86_srcs = AVX512SKX_UKERNELS,
3671 deps = [
3672 ":tables",
3673 "@FP16",
3674 "@pthreadpool",
3675 ],
3676)
3677
3678xnnpack_cc_library(
3679 name = "avx512skx_ukernels_test_mode",
3680 hdrs = INTERNAL_HDRS,
3681 copts = [
3682 "-UNDEBUG",
3683 "-DXNN_TEST_MODE=1",
3684 ],
3685 gcc_copts = xnnpack_gcc_std_copts(),
3686 gcc_x86_copts = [
3687 "-mavx512f",
3688 "-mavx512cd",
3689 "-mavx512bw",
3690 "-mavx512dq",
3691 "-mavx512vl",
3692 ],
3693 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3694 msvc_copts = xnnpack_msvc_std_copts(),
3695 msvc_x86_32_copts = ["/arch:AVX512"],
3696 msvc_x86_64_copts = ["/arch:AVX512"],
3697 msys_copts = ["-fno-asynchronous-unwind-tables"],
3698 x86_srcs = AVX512SKX_UKERNELS,
3699 deps = [
3700 ":tables",
3701 "@FP16",
3702 "@pthreadpool",
3703 ],
3704)
3705
3706xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003707 name = "asm_ukernels",
3708 hdrs = ["src/xnnpack/assembly.h"],
3709 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07003710 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003711 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003712 wasm_srcs = WASM32_ASM_UKERNELS,
3713 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003714)
3715
Marat Dukhan3b59de22020-06-03 20:15:19 -07003716xnnpack_cc_library(
3717 name = "logging_utils",
3718 srcs = LOGGING_SRCS,
3719 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3720 copts = LOGGING_COPTS + [
3721 "-Isrc",
3722 "-Iinclude",
3723 ] + select({
3724 ":debug_build": [],
3725 "//conditions:default": xnnpack_min_size_copts(),
3726 }),
3727 gcc_copts = xnnpack_gcc_std_copts(),
3728 msvc_copts = xnnpack_msvc_std_copts(),
3729 visibility = xnnpack_visibility(),
3730 deps = [
3731 "@FP16",
3732 "@clog",
3733 "@pthreadpool",
3734 ],
3735)
3736
Marat Dukhan08c4a432019-10-03 09:29:21 -07003737xnnpack_aggregate_library(
3738 name = "ukernels",
3739 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003740 ":neon_ukernels",
3741 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003742 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003743 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003744 ":asm_ukernels",
3745 ],
3746 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003747 ":neon_ukernels",
3748 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003749 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003750 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003751 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003752 ":asm_ukernels",
3753 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003754 generic_deps = [
3755 ":scalar_ukernels",
3756 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003757 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003758 ":psimd_fastmath_ukernels",
3759 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003760 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003761 wasm_deps = [
3762 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003763 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003764 ],
3765 wasmsimd_deps = [
3766 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003767 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003768 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003769 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003770 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003771 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003772 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003773 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003774 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003775 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003776 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003777 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003778 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003779 ],
3780)
3781
Marat Dukhan33fcf782020-05-24 14:27:15 -07003782xnnpack_aggregate_library(
3783 name = "ukernels_test_mode",
3784 aarch32_deps = [
3785 ":neon_ukernels_test_mode",
3786 ":neonfma_ukernels_test_mode",
3787 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003788 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003789 ":asm_ukernels",
3790 ],
3791 aarch64_deps = [
3792 ":neon_ukernels_test_mode",
3793 ":neonfma_ukernels_test_mode",
3794 ":neonv8_ukernels_test_mode",
3795 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003796 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003797 ":asm_ukernels",
3798 ],
3799 generic_deps = [
3800 ":scalar_ukernels_test_mode",
3801 ],
3802 psimd_deps = [
3803 ":psimd_fastmath_ukernels_test_mode",
3804 ":psimd_accmath_ukernels_test_mode",
3805 ],
3806 wasm_deps = [
3807 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003808 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003809 ],
3810 wasmsimd_deps = [
3811 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003812 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003813 ],
3814 x86_deps = [
3815 ":sse2_ukernels_test_mode",
3816 ":ssse3_ukernels_test_mode",
3817 ":sse41_ukernels_test_mode",
3818 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003819 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003820 ":fma3_ukernels_test_mode",
3821 ":avx2_ukernels_test_mode",
3822 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003823 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003824 ],
3825)
3826
Marat Dukhan86df3842020-06-17 14:03:31 -07003827# This target must be compatible with Android NDK r17 and compile with Clang 6.
3828# It is similar to :ukernels target, but excludes microkernels which require
3829# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003830xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003831 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003832 aarch32_deps = [
3833 ":neon_ukernels",
3834 ":neonfma_ukernels",
3835 ":neonv8_ukernels",
3836 ":asm_ukernels",
3837 ],
3838 aarch64_deps = [
3839 ":neon_ukernels",
3840 ":neonfma_ukernels",
3841 ":neonv8_ukernels",
3842 ":asm_ukernels",
3843 ],
3844 generic_deps = [
3845 ":scalar_ukernels",
3846 ],
3847 psimd_deps = [
3848 ":psimd_fastmath_ukernels",
3849 ":psimd_accmath_ukernels",
3850 ],
3851 wasm_deps = [
3852 ":wasm_ukernels",
3853 ],
3854 wasmsimd_deps = [
3855 ":wasm_ukernels",
3856 ],
3857 x86_deps = [
3858 ":sse2_ukernels",
3859 ":ssse3_ukernels",
3860 ":sse41_ukernels",
3861 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003862 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003863 ":fma3_ukernels",
3864 ":avx2_ukernels",
3865 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003866 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003867 ],
3868)
3869
Marat Dukhan08c4a432019-10-03 09:29:21 -07003870xnnpack_cc_library(
3871 name = "im2col",
3872 srcs = ["src/im2col.c"],
3873 hdrs = [
3874 "src/xnnpack/common.h",
3875 "src/xnnpack/im2col.h",
3876 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003877 gcc_copts = xnnpack_gcc_std_copts(),
3878 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003879)
3880
3881xnnpack_cc_library(
3882 name = "indirection",
3883 srcs = ["src/indirection.c"],
3884 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003885 gcc_copts = xnnpack_gcc_std_copts(),
3886 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003887 deps = [
3888 "@FP16",
3889 "@FXdiv",
3890 "@pthreadpool",
3891 ],
3892)
3893
3894xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003895 name = "indirection_test_mode",
3896 srcs = ["src/indirection.c"],
3897 hdrs = INTERNAL_HDRS,
3898 copts = [
3899 "-UNDEBUG",
3900 "-DXNN_TEST_MODE=1",
3901 ],
3902 gcc_copts = xnnpack_gcc_std_copts(),
3903 msvc_copts = xnnpack_msvc_std_copts(),
3904 deps = [
3905 "@FP16",
3906 "@FXdiv",
3907 "@pthreadpool",
3908 ],
3909)
3910
3911xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003912 name = "packing",
3913 srcs = ["src/packing.c"],
3914 hdrs = INTERNAL_HDRS,
3915 gcc_copts = xnnpack_gcc_std_copts(),
3916 msvc_copts = xnnpack_msvc_std_copts(),
3917 deps = [
3918 "@FP16",
3919 "@FXdiv",
3920 "@pthreadpool",
3921 ],
3922)
3923
3924xnnpack_cc_library(
3925 name = "packing_test_mode",
3926 srcs = ["src/packing.c"],
3927 hdrs = INTERNAL_HDRS,
3928 copts = [
3929 "-UNDEBUG",
3930 "-DXNN_TEST_MODE=1",
3931 ],
3932 gcc_copts = xnnpack_gcc_std_copts(),
3933 msvc_copts = xnnpack_msvc_std_copts(),
3934 deps = [
3935 "@FP16",
3936 "@FXdiv",
3937 "@pthreadpool",
3938 ],
3939)
3940
3941xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003942 name = "operator_run",
3943 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003944 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003945 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003946 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3947 "//conditions:default": [],
3948 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003949 gcc_copts = xnnpack_gcc_std_copts(),
3950 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003951 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003952 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003953 "@FP16",
3954 "@FXdiv",
3955 "@clog",
3956 "@pthreadpool",
3957 ],
3958)
3959
Chao Mei6ddfc602020-05-13 22:29:36 -07003960xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003961 name = "operator_run_test_mode",
3962 srcs = ["src/operator-run.c"],
3963 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3964 copts = LOGGING_COPTS + [
3965 "-UNDEBUG",
3966 "-DXNN_TEST_MODE=1",
3967 ] + select({
3968 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3969 "//conditions:default": [],
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 "@FP16",
3976 "@FXdiv",
3977 "@clog",
3978 "@pthreadpool",
3979 ],
3980)
3981
3982xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003983 name = "memory_planner",
3984 srcs = ["src/memory-planner.c"],
3985 hdrs = INTERNAL_HDRS,
3986 defines = select({
3987 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3988 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3989 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3990 }),
3991 gcc_copts = xnnpack_gcc_std_copts(),
3992 msvc_copts = xnnpack_msvc_std_copts(),
3993 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003994 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003995 "@pthreadpool",
3996 ],
3997)
3998
Marat Dukhan33fcf782020-05-24 14:27:15 -07003999xnnpack_cc_library(
4000 name = "memory_planner_test_mode",
4001 srcs = ["src/memory-planner.c"],
4002 hdrs = INTERNAL_HDRS,
4003 copts = [
4004 "-UNDEBUG",
4005 "-DXNN_TEST_MODE=1",
4006 ],
4007 defines = select({
4008 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4009 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4010 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4011 }),
4012 gcc_copts = xnnpack_gcc_std_copts(),
4013 msvc_copts = xnnpack_msvc_std_copts(),
4014 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004015 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004016 "@pthreadpool",
4017 ],
4018)
4019
Marat Dukhan08c4a432019-10-03 09:29:21 -07004020cc_library(
4021 name = "enable_assembly",
4022 defines = select({
4023 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4024 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004025 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004026 }),
4027)
4028
Marat Dukhan9de90e02020-06-18 16:04:12 -07004029cc_library(
4030 name = "enable_sparse",
4031 defines = select({
4032 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4033 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07004034 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004035 }),
4036)
4037
Marat Dukhancf056b22019-10-07 10:26:29 -07004038xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004039 name = "operators",
4040 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004041 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004042 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004043 ],
4044 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004045 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004046 "-Isrc",
4047 "-Iinclude",
4048 ] + select({
4049 ":debug_build": [],
4050 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004051 }) + select({
4052 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4053 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004054 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004055 gcc_copts = xnnpack_gcc_std_copts(),
4056 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004057 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004058 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004059 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004060 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004061 "@FP16",
4062 "@FXdiv",
4063 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004064 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004065 ],
4066)
4067
Marat Dukhan10a38082020-04-17 03:58:35 -07004068xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004069 name = "operators_test_mode",
4070 srcs = OPERATOR_SRCS + [
4071 "src/memory.c",
4072 "src/operator-delete.c",
4073 ],
4074 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4075 copts = LOGGING_COPTS + [
4076 "-Isrc",
4077 "-Iinclude",
4078 "-UNDEBUG",
4079 "-DXNN_TEST_MODE=1",
4080 ] + select({
4081 ":debug_build": [],
4082 "//conditions:default": xnnpack_min_size_copts(),
4083 }) + select({
4084 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4085 "//conditions:default": [],
4086 }),
4087 gcc_copts = xnnpack_gcc_std_copts(),
4088 msvc_copts = xnnpack_msvc_std_copts(),
4089 deps = [
4090 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004091 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004092 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004093 "@FP16",
4094 "@FXdiv",
4095 "@clog",
4096 "@pthreadpool",
4097 ],
4098)
4099
4100xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004101 name = "XNNPACK",
4102 srcs = [
4103 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004104 "src/runtime.c",
4105 "src/subgraph.c",
4106 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004107 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004108 hdrs = ["include/xnnpack.h"],
4109 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004110 "-Isrc",
4111 "-Iinclude",
4112 ] + select({
4113 ":debug_build": [],
4114 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004115 }) + select({
4116 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4117 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004118 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004119 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004120 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004121 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004122 visibility = xnnpack_visibility(),
4123 deps = [
4124 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004125 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004126 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004127 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004128 ":operator_run",
4129 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004130 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004131 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004132 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004133 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004134 ] + select({
4135 ":emscripten": [],
4136 "//conditions:default": ["@cpuinfo"],
4137 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004138)
4139
Marat Dukhan10a38082020-04-17 03:58:35 -07004140xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004141 name = "XNNPACK_test_mode",
4142 srcs = [
4143 "src/init.c",
4144 "src/runtime.c",
4145 "src/subgraph.c",
4146 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004147 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004148 hdrs = ["include/xnnpack.h"],
4149 copts = LOGGING_COPTS + [
4150 "-Isrc",
4151 "-Iinclude",
4152 "-UNDEBUG",
4153 "-DXNN_TEST_MODE=1",
4154 ] + select({
4155 ":debug_build": [],
4156 "//conditions:default": xnnpack_min_size_copts(),
4157 }) + select({
4158 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4159 "//conditions:default": [],
4160 }),
4161 gcc_copts = xnnpack_gcc_std_copts(),
4162 includes = ["include"],
4163 msvc_copts = xnnpack_msvc_std_copts(),
4164 visibility = xnnpack_visibility(),
4165 deps = [
4166 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004167 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004168 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004169 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004170 ":operator_run_test_mode",
4171 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004172 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004173 "@clog",
4174 "@FP16",
4175 "@pthreadpool",
4176 ] + select({
4177 ":emscripten": [],
4178 "//conditions:default": ["@cpuinfo"],
4179 }),
4180)
4181
4182xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004183 name = "xnnpack_f32",
4184 srcs = [
4185 "src/init.c",
4186 "src/runtime.c",
4187 "src/subgraph.c",
4188 "src/tensor.c",
4189 ] + SUBGRAPH_SRCS,
4190 hdrs = ["include/xnnpack.h"],
4191 copts = LOGGING_COPTS + [
4192 "-Isrc",
4193 "-Iinclude",
4194 ] + select({
4195 ":debug_build": [],
4196 "//conditions:default": xnnpack_min_size_copts(),
4197 }) + select({
4198 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4199 "//conditions:default": [],
4200 }),
4201 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004202 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004203 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004204 "XNN_NO_U8_OPERATORS",
4205 "XNN_NO_X8_OPERATORS",
4206 "XNN_NO_F16_OPERATORS",
4207 "XNN_NO_X16_OPERATORS",
4208 ],
4209 gcc_copts = xnnpack_gcc_std_copts(),
4210 includes = ["include"],
4211 msvc_copts = xnnpack_msvc_std_copts(),
4212 visibility = xnnpack_visibility(),
4213 deps = [
4214 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004215 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004216 ":logging_utils",
4217 ":memory_planner",
4218 ":operator_run",
4219 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004220 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004221 "@clog",
4222 "@FP16",
4223 "@pthreadpool",
4224 ] + select({
4225 ":emscripten": [],
4226 "//conditions:default": ["@cpuinfo"],
4227 }),
4228)
4229
4230xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004231 name = "xnnpack_operators_nhwc_f32",
4232 srcs = [
4233 "src/init.c",
4234 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004235 hdrs = ["include/xnnpack.h"],
4236 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004237 "-Isrc",
4238 "-Iinclude",
4239 ] + select({
4240 ":debug_build": [],
4241 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004242 }) + select({
4243 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4244 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004245 }),
4246 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004247 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004248 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004249 "XNN_NO_U8_OPERATORS",
4250 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004251 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004252 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004253 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004254 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004255 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004256 visibility = xnnpack_visibility(),
4257 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004258 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004259 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004260 ":operator_run",
4261 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004262 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004263 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004264 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004265 ] + select({
4266 ":emscripten": [],
4267 "//conditions:default": ["@cpuinfo"],
4268 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004269)
4270
Marat Dukhancf056b22019-10-07 10:26:29 -07004271xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004272 name = "bench_utils",
4273 srcs = ["bench/utils.cc"],
4274 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004275 deps = [
4276 "@com_google_benchmark//:benchmark",
4277 "@cpuinfo",
4278 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004279)
4280
Frank Barchard7e955972019-10-11 10:34:25 -07004281######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004282
4283xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004284 name = "qs8_gemm_bench",
4285 srcs = [
4286 "bench/gemm.h",
4287 "bench/qs8-gemm.cc",
4288 "src/xnnpack/AlignedAllocator.h",
4289 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004290 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4291 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004292)
4293
4294xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004295 name = "qs8_requantization_bench",
4296 srcs = [
4297 "bench/qs8-requantization.cc",
4298 "src/xnnpack/requantization-stubs.h",
4299 "src/xnnpack/AlignedAllocator.h",
4300 ] + MICROKERNEL_BENCHMARK_HDRS,
4301 deps = MICROKERNEL_BENCHMARK_DEPS,
4302)
4303
4304xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004305 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004306 srcs = [
4307 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004308 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004309 "src/xnnpack/AlignedAllocator.h",
4310 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004311 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004312 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004313)
4314
4315xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004316 name = "qu8_requantization_bench",
4317 srcs = [
4318 "bench/qu8-requantization.cc",
4319 "src/xnnpack/requantization-stubs.h",
4320 "src/xnnpack/AlignedAllocator.h",
4321 ] + MICROKERNEL_BENCHMARK_HDRS,
4322 deps = MICROKERNEL_BENCHMARK_DEPS,
4323)
4324
4325xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004326 name = "f16_igemm_bench",
4327 srcs = [
4328 "bench/f16-igemm.cc",
4329 "bench/conv.h",
4330 "bench/google/conv.h",
4331 "src/xnnpack/AlignedAllocator.h",
4332 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004333 deps = MICROKERNEL_BENCHMARK_DEPS + [
4334 ":indirection",
4335 ":packing",
4336 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004337)
4338
4339xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004340 name = "f16_gemm_bench",
4341 srcs = [
4342 "bench/f16-gemm.cc",
4343 "bench/gemm.h",
4344 "src/xnnpack/AlignedAllocator.h",
4345 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004346 deps = MICROKERNEL_BENCHMARK_DEPS + [
4347 ":packing",
4348 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004349)
4350
4351xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004352 name = "f16_spmm_bench",
4353 srcs = [
4354 "bench/f16-spmm.cc",
4355 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004356 "src/xnnpack/AlignedAllocator.h",
4357 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004358 deps = MICROKERNEL_BENCHMARK_DEPS,
4359)
4360
4361xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004362 name = "f32_igemm_bench",
4363 srcs = [
4364 "bench/f32-igemm.cc",
4365 "bench/conv.h",
4366 "src/xnnpack/AlignedAllocator.h",
4367 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004368 deps = MICROKERNEL_BENCHMARK_DEPS + [
4369 ":indirection",
4370 ":packing",
4371 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004372)
4373
4374xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004375 name = "f16_relu_bench",
4376 srcs = [
4377 "bench/f16-relu.cc",
4378 "src/xnnpack/AlignedAllocator.h",
4379 ] + MICROKERNEL_BENCHMARK_HDRS,
4380 deps = MICROKERNEL_BENCHMARK_DEPS,
4381)
4382
4383xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004384 name = "f32_conv_hwc_bench",
4385 srcs = [
4386 "bench/f32-conv-hwc.cc",
4387 "bench/dconv.h",
4388 "src/xnnpack/AlignedAllocator.h",
4389 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004390 deps = MICROKERNEL_BENCHMARK_DEPS + [
4391 ":packing",
4392 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004393)
4394
4395xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004396 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004397 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004398 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004399 "bench/dconv.h",
4400 "src/xnnpack/AlignedAllocator.h",
4401 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004402 deps = MICROKERNEL_BENCHMARK_DEPS + [
4403 ":packing",
4404 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004405)
4406
4407xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004408 name = "f16_dwconv_bench",
4409 srcs = [
4410 "bench/f16-dwconv.cc",
4411 "bench/dwconv.h",
4412 "bench/google/dwconv.h",
4413 "src/xnnpack/AlignedAllocator.h",
4414 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004415 deps = MICROKERNEL_BENCHMARK_DEPS + [
4416 ":indirection",
4417 ":packing",
4418 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004419)
4420
4421xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004422 name = "f32_dwconv_bench",
4423 srcs = [
4424 "bench/f32-dwconv.cc",
4425 "bench/dwconv.h",
4426 "src/xnnpack/AlignedAllocator.h",
4427 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004428 deps = MICROKERNEL_BENCHMARK_DEPS + [
4429 ":indirection",
4430 ":packing",
4431 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004432)
4433
4434xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004435 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004436 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004437 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004438 "bench/dwconv.h",
4439 "src/xnnpack/AlignedAllocator.h",
4440 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004441 deps = MICROKERNEL_BENCHMARK_DEPS + [
4442 ":indirection",
4443 ":packing",
4444 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004445)
4446
4447xnnpack_benchmark(
4448 name = "f32_gemm_bench",
4449 srcs = [
4450 "bench/f32-gemm.cc",
4451 "bench/gemm.h",
4452 "src/xnnpack/AlignedAllocator.h",
4453 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004454 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004455 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004456)
4457
4458xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004459 name = "f32_hswish_bench",
4460 srcs = [
4461 "bench/f32-hswish.cc",
4462 "src/xnnpack/AlignedAllocator.h",
4463 ] + MICROKERNEL_BENCHMARK_HDRS,
4464 deps = MICROKERNEL_BENCHMARK_DEPS,
4465)
4466
4467xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004468 name = "f32_raddexpminusmax_bench",
4469 srcs = [
4470 "bench/f32-raddexpminusmax.cc",
4471 "src/xnnpack/AlignedAllocator.h",
4472 ] + MICROKERNEL_BENCHMARK_HDRS,
4473 deps = MICROKERNEL_BENCHMARK_DEPS,
4474)
4475
4476xnnpack_benchmark(
4477 name = "f32_raddextexp_bench",
4478 srcs = [
4479 "bench/f32-raddextexp.cc",
4480 "src/xnnpack/AlignedAllocator.h",
4481 ] + MICROKERNEL_BENCHMARK_HDRS,
4482 deps = MICROKERNEL_BENCHMARK_DEPS,
4483)
4484
4485xnnpack_benchmark(
4486 name = "f32_raddstoreexpminusmax_bench",
4487 srcs = [
4488 "bench/f32-raddstoreexpminusmax.cc",
4489 "src/xnnpack/AlignedAllocator.h",
4490 ] + MICROKERNEL_BENCHMARK_HDRS,
4491 deps = MICROKERNEL_BENCHMARK_DEPS,
4492)
4493
4494xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004495 name = "f32_relu_bench",
4496 srcs = [
4497 "bench/f32-relu.cc",
4498 "src/xnnpack/AlignedAllocator.h",
4499 ] + MICROKERNEL_BENCHMARK_HDRS,
4500 deps = MICROKERNEL_BENCHMARK_DEPS,
4501)
4502
4503xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004504 name = "f32_rmax_bench",
4505 srcs = [
4506 "bench/f32-rmax.cc",
4507 "src/xnnpack/AlignedAllocator.h",
4508 ] + MICROKERNEL_BENCHMARK_HDRS,
4509 deps = MICROKERNEL_BENCHMARK_DEPS,
4510)
4511
4512xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004513 name = "f32_sigmoid_bench",
4514 srcs = [
4515 "bench/f32-sigmoid.cc",
4516 "src/xnnpack/AlignedAllocator.h",
4517 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004518 deps = MICROKERNEL_BENCHMARK_DEPS,
4519)
4520
4521xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004522 name = "f32_spmm_bench",
4523 srcs = [
4524 "bench/f32-spmm.cc",
4525 "bench/gemm.h",
4526 "src/xnnpack/AlignedAllocator.h",
4527 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004528 deps = MICROKERNEL_BENCHMARK_DEPS,
4529)
4530
4531xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004532 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004533 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004534 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004535 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004536 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004537 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004538)
4539
4540xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004541 name = "f32_vscaleexpminusmax_bench",
4542 srcs = [
4543 "bench/f32-vscaleexpminusmax.cc",
4544 "src/xnnpack/AlignedAllocator.h",
4545 ] + MICROKERNEL_BENCHMARK_HDRS,
4546 deps = MICROKERNEL_BENCHMARK_DEPS,
4547)
4548
4549xnnpack_benchmark(
4550 name = "f32_vscaleextexp_bench",
4551 srcs = [
4552 "bench/f32-vscaleextexp.cc",
4553 "src/xnnpack/AlignedAllocator.h",
4554 ] + MICROKERNEL_BENCHMARK_HDRS,
4555 deps = MICROKERNEL_BENCHMARK_DEPS,
4556)
4557
4558xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004559 name = "f32_vsqrt_bench",
4560 srcs = [
4561 "bench/f32-vsqrt.cc",
4562 "src/xnnpack/AlignedAllocator.h",
4563 ] + MICROKERNEL_BENCHMARK_HDRS,
4564 deps = MICROKERNEL_BENCHMARK_DEPS,
4565)
4566
4567xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004568 name = "f32_im2col_gemm_bench",
4569 srcs = [
4570 "bench/f32-im2col-gemm.cc",
4571 "bench/conv.h",
4572 "src/xnnpack/AlignedAllocator.h",
4573 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004574 deps = MICROKERNEL_BENCHMARK_DEPS + [
4575 ":im2col",
4576 ":packing",
4577 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004578)
4579
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004580xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004581 name = "rounding_bench",
4582 srcs = [
4583 "bench/rounding.cc",
4584 "src/xnnpack/math-stubs.h",
4585 "src/xnnpack/AlignedAllocator.h",
4586 ] + MICROKERNEL_BENCHMARK_HDRS,
4587 deps = MICROKERNEL_BENCHMARK_DEPS,
4588)
4589
Marat Dukhan08c4a432019-10-03 09:29:21 -07004590########################### Benchmarks for operators ###########################
4591
4592xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004593 name = "average_pooling_bench",
4594 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004595 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004596 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004597 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004598)
4599
4600xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004601 name = "bankers_rounding_bench",
4602 srcs = ["bench/bankers-rounding.cc"],
4603 copts = xnnpack_optional_tflite_copts(),
4604 tags = ["nowin32"],
4605 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4606)
4607
4608xnnpack_benchmark(
4609 name = "ceiling_bench",
4610 srcs = ["bench/ceiling.cc"],
4611 copts = xnnpack_optional_tflite_copts(),
4612 tags = ["nowin32"],
4613 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4614)
4615
4616xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004617 name = "channel_shuffle_bench",
4618 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004619 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004620)
4621
4622xnnpack_benchmark(
4623 name = "convolution_bench",
4624 srcs = ["bench/convolution.cc"],
4625 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004626 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004627 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004628)
4629
4630xnnpack_benchmark(
4631 name = "deconvolution_bench",
4632 srcs = ["bench/deconvolution.cc"],
4633 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004634 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004635 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004636)
4637
4638xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004639 name = "floor_bench",
4640 srcs = ["bench/floor.cc"],
4641 copts = xnnpack_optional_tflite_copts(),
4642 tags = ["nowin32"],
4643 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4644)
4645
4646xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004647 name = "global_average_pooling_bench",
4648 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004649 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004650)
4651
4652xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004653 name = "hardswish_bench",
4654 srcs = ["bench/hardswish.cc"],
4655 copts = xnnpack_optional_tflite_copts(),
4656 tags = ["nowin32"],
4657 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4658)
4659
4660xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004661 name = "max_pooling_bench",
4662 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004663 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004664)
4665
4666xnnpack_benchmark(
4667 name = "sigmoid_bench",
4668 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004669 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004670 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004671 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004672)
4673
4674xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004675 name = "prelu_bench",
4676 srcs = ["bench/prelu.cc"],
4677 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004678 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004679 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004680)
4681
4682xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004683 name = "softmax_bench",
4684 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004685 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004686 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004687 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004688)
4689
Marat Dukhan87727142020-06-24 15:24:10 -07004690xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004691 name = "square_root_bench",
4692 srcs = ["bench/square-root.cc"],
4693 copts = xnnpack_optional_tflite_copts(),
4694 tags = ["nowin32"],
4695 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4696)
4697
4698xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004699 name = "truncation_bench",
4700 srcs = ["bench/truncation.cc"],
4701 deps = OPERATOR_BENCHMARK_DEPS,
4702)
4703
Marat Dukhanc068bb62019-10-04 13:24:39 -07004704############################# End-to-end benchmarks ############################
4705
4706cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004707 name = "fp32_mobilenet_v1",
4708 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004709 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004710 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004711 linkstatic = True,
4712 deps = [
4713 ":XNNPACK",
4714 "@pthreadpool",
4715 ],
4716)
4717
4718cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004719 name = "fp16_mobilenet_v1",
4720 srcs = ["models/fp16-mobilenet-v1.cc"],
4721 hdrs = ["models/models.h"],
4722 copts = xnnpack_std_cxxopts(),
4723 linkstatic = True,
4724 deps = [
4725 ":XNNPACK",
4726 "@FP16",
4727 "@pthreadpool",
4728 ],
4729)
4730
4731cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004732 name = "qs8_mobilenet_v1",
4733 srcs = ["models/qs8-mobilenet-v1.cc"],
4734 hdrs = ["models/models.h"],
4735 copts = xnnpack_std_cxxopts(),
4736 linkstatic = True,
4737 deps = [
4738 ":XNNPACK",
4739 "@pthreadpool",
4740 ],
4741)
4742
4743cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004744 name = "qs8_mobilenet_v2",
4745 srcs = ["models/qs8-mobilenet-v2.cc"],
4746 hdrs = ["models/models.h"],
4747 copts = xnnpack_std_cxxopts(),
4748 linkstatic = True,
4749 deps = [
4750 ":XNNPACK",
4751 "@pthreadpool",
4752 ],
4753)
4754
4755cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004756 name = "fp32_mobilenet_v2",
4757 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004758 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004759 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004760 linkstatic = True,
4761 deps = [
4762 ":XNNPACK",
4763 "@pthreadpool",
4764 ],
4765)
4766
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004767cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004768 name = "fp16_mobilenet_v2",
4769 srcs = ["models/fp16-mobilenet-v2.cc"],
4770 hdrs = ["models/models.h"],
4771 copts = xnnpack_std_cxxopts(),
4772 linkstatic = True,
4773 deps = [
4774 ":XNNPACK",
4775 "@FP16",
4776 "@pthreadpool",
4777 ],
4778)
4779
4780cc_library(
4781 name = "fp32_mobilenet_v3_large",
4782 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004783 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004784 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004785 linkstatic = True,
4786 deps = [
4787 ":XNNPACK",
4788 "@pthreadpool",
4789 ],
4790)
4791
4792cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004793 name = "fp16_mobilenet_v3_large",
4794 srcs = ["models/fp16-mobilenet-v3-large.cc"],
4795 hdrs = ["models/models.h"],
4796 copts = xnnpack_std_cxxopts(),
4797 linkstatic = True,
4798 deps = [
4799 ":XNNPACK",
4800 "@FP16",
4801 "@pthreadpool",
4802 ],
4803)
4804
4805cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004806 name = "fp32_mobilenet_v3_small",
4807 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004808 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004809 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004810 linkstatic = True,
4811 deps = [
4812 ":XNNPACK",
4813 "@pthreadpool",
4814 ],
4815)
4816
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004817cc_library(
4818 name = "fp16_mobilenet_v3_small",
4819 srcs = ["models/fp16-mobilenet-v3-small.cc"],
4820 hdrs = ["models/models.h"],
4821 copts = xnnpack_std_cxxopts(),
4822 linkstatic = True,
4823 deps = [
4824 ":XNNPACK",
4825 "@FP16",
4826 "@pthreadpool",
4827 ],
4828)
4829
Marat Dukhanc068bb62019-10-04 13:24:39 -07004830xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004831 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004832 srcs = [
4833 "bench/f32-dwconv-e2e.cc",
4834 "bench/end2end.h",
4835 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004836 deps = MICROKERNEL_BENCHMARK_DEPS + [
4837 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004838 ":fp32_mobilenet_v1",
4839 ":fp32_mobilenet_v2",
4840 ":fp32_mobilenet_v3_large",
4841 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004842 ],
4843)
4844
4845xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004846 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004847 srcs = [
4848 "bench/f32-gemm-e2e.cc",
4849 "bench/end2end.h",
4850 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004851 deps = MICROKERNEL_BENCHMARK_DEPS + [
4852 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004853 ":fp32_mobilenet_v1",
4854 ":fp32_mobilenet_v2",
4855 ":fp32_mobilenet_v3_large",
4856 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004857 ],
4858)
4859
4860xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004861 name = "end2end_bench",
4862 srcs = ["bench/end2end.cc"],
4863 deps = [
4864 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004865 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004866 ":fp16_mobilenet_v1",
4867 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004868 ":fp16_mobilenet_v3_large",
4869 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004870 ":fp32_mobilenet_v1",
4871 ":fp32_mobilenet_v2",
4872 ":fp32_mobilenet_v3_large",
4873 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004874 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07004875 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004876 "@pthreadpool",
4877 ],
4878)
4879
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004880#################### Accuracy evaluation for math functions ####################
4881
4882xnnpack_benchmark(
4883 name = "f32_exp_eval",
4884 srcs = [
4885 "eval/f32-exp.cc",
4886 "src/xnnpack/AlignedAllocator.h",
4887 ] + ACCURACY_EVAL_HDRS,
4888 deps = ACCURACY_EVAL_DEPS,
4889)
4890
Marat Dukhan515c9772019-10-17 18:07:57 -07004891xnnpack_benchmark(
4892 name = "f32_expminus_eval",
4893 srcs = [
4894 "eval/f32-expminus.cc",
4895 "src/xnnpack/AlignedAllocator.h",
4896 ] + ACCURACY_EVAL_HDRS,
4897 deps = ACCURACY_EVAL_DEPS,
4898)
4899
Marat Dukhan98ba4412019-10-23 02:14:28 -07004900xnnpack_benchmark(
4901 name = "f32_extexp_eval",
4902 srcs = [
4903 "eval/f32-extexp.cc",
4904 "src/xnnpack/AlignedAllocator.h",
4905 ] + ACCURACY_EVAL_HDRS,
4906 deps = ACCURACY_EVAL_DEPS,
4907)
4908
Marat Dukhan8853b822020-05-07 12:19:01 -07004909xnnpack_unit_test(
4910 name = "f32_roundne_eval",
4911 srcs = [
4912 "eval/f32-roundne.cc",
4913 "src/xnnpack/AlignedAllocator.h",
4914 "src/xnnpack/math-stubs.h",
4915 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004916 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004917 deps = MICROKERNEL_TEST_DEPS,
4918)
4919
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004920xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004921 name = "f32_roundd_eval",
4922 srcs = [
4923 "eval/f32-roundd.cc",
4924 "src/xnnpack/AlignedAllocator.h",
4925 "src/xnnpack/math-stubs.h",
4926 ] + MICROKERNEL_TEST_HDRS,
4927 automatic = False,
4928 deps = MICROKERNEL_TEST_DEPS,
4929)
4930
4931xnnpack_unit_test(
4932 name = "f32_roundu_eval",
4933 srcs = [
4934 "eval/f32-roundu.cc",
4935 "src/xnnpack/AlignedAllocator.h",
4936 "src/xnnpack/math-stubs.h",
4937 ] + MICROKERNEL_TEST_HDRS,
4938 automatic = False,
4939 deps = MICROKERNEL_TEST_DEPS,
4940)
4941
4942xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004943 name = "f32_roundz_eval",
4944 srcs = [
4945 "eval/f32-roundz.cc",
4946 "src/xnnpack/AlignedAllocator.h",
4947 "src/xnnpack/math-stubs.h",
4948 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004949 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004950 deps = MICROKERNEL_TEST_DEPS,
4951)
4952
Marat Dukhan346a9e52019-11-15 09:06:30 -08004953xnnpack_benchmark(
4954 name = "f32_sigmoid_eval",
4955 srcs = [
4956 "eval/f32-sigmoid.cc",
4957 "src/xnnpack/AlignedAllocator.h",
4958 ] + ACCURACY_EVAL_HDRS,
4959 deps = ACCURACY_EVAL_DEPS,
4960)
4961
Marat Dukhan84000762020-06-29 18:38:43 -07004962xnnpack_benchmark(
4963 name = "f32_sqrt_eval",
4964 srcs = [
4965 "eval/f32-sqrt.cc",
4966 "src/xnnpack/AlignedAllocator.h",
4967 ] + ACCURACY_EVAL_HDRS,
4968 deps = ACCURACY_EVAL_DEPS + [
4969 ":bench_utils",
4970 ],
4971)
4972
Marat Dukhan08c4a432019-10-03 09:29:21 -07004973######################### Unit tests for micro-kernels #########################
4974
4975xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004976 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004977 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004978 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004979 "test/gemm-microkernel-tester.h",
4980 "src/xnnpack/AlignedAllocator.h",
4981 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004982 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004983)
4984
4985xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004986 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004987 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004988 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004989 "test/spmm-microkernel-tester.h",
4990 "src/xnnpack/AlignedAllocator.h",
4991 ] + MICROKERNEL_TEST_HDRS,
4992 deps = MICROKERNEL_TEST_DEPS,
4993)
4994
4995xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004996 name = "f32_argmaxpool_test",
4997 srcs = [
4998 "test/f32-argmaxpool.cc",
4999 "test/argmaxpool-microkernel-tester.h",
5000 "src/xnnpack/AlignedAllocator.h",
5001 ] + MICROKERNEL_TEST_HDRS,
5002 deps = MICROKERNEL_TEST_DEPS,
5003)
5004
5005xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005006 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005007 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005008 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005009 "test/avgpool-microkernel-tester.h",
5010 "src/xnnpack/AlignedAllocator.h",
5011 ] + MICROKERNEL_TEST_HDRS,
5012 deps = MICROKERNEL_TEST_DEPS,
5013)
5014
5015xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005016 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005017 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005018 "test/f32-ibilinear.cc",
5019 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005020 "src/xnnpack/AlignedAllocator.h",
5021 ] + MICROKERNEL_TEST_HDRS,
5022 deps = MICROKERNEL_TEST_DEPS,
5023)
5024
5025xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005026 name = "f16_clamp_test",
5027 srcs = [
5028 "test/f16-clamp.cc",
5029 "test/clamp-microkernel-tester.h",
5030 ] + MICROKERNEL_TEST_HDRS,
5031 deps = MICROKERNEL_TEST_DEPS,
5032)
5033
5034xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005035 name = "f32_clamp_test",
5036 srcs = [
5037 "test/f32-clamp.cc",
5038 "test/clamp-microkernel-tester.h",
5039 ] + MICROKERNEL_TEST_HDRS,
5040 deps = MICROKERNEL_TEST_DEPS,
5041)
5042
5043xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005044 name = "f32_igemm_test",
5045 srcs = [
5046 "test/f32-igemm.cc",
5047 "test/gemm-microkernel-tester.h",
5048 "src/xnnpack/AlignedAllocator.h",
5049 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005050 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005051)
5052
5053xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005054 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005055 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005056 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005057 "test/gemm-microkernel-tester.h",
5058 "src/xnnpack/AlignedAllocator.h",
5059 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005060 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005061)
5062
5063xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005064 name = "f32_igemm_minmax_test",
5065 srcs = [
5066 "test/f32-igemm-minmax.cc",
5067 "test/gemm-microkernel-tester.h",
5068 "src/xnnpack/AlignedAllocator.h",
5069 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005070 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005071)
5072
5073xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005074 name = "f16_igemm_minmax_test",
5075 srcs = [
5076 "test/f16-igemm-minmax.cc",
5077 "test/gemm-microkernel-tester.h",
5078 "src/xnnpack/AlignedAllocator.h",
5079 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005080 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005081)
5082
5083xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005084 name = "f32_conv_hwc_test",
5085 srcs = [
5086 "test/f32-conv-hwc.cc",
5087 "test/conv-hwc-microkernel-tester.h",
5088 "src/xnnpack/AlignedAllocator.h",
5089 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005090 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005091)
5092
5093xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005094 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005095 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005096 "test/f32-conv-hwc2chw.cc",
5097 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005098 "src/xnnpack/AlignedAllocator.h",
5099 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005100 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005101)
5102
5103xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005104 name = "f32_dwconv_test",
5105 srcs = [
5106 "test/f32-dwconv.cc",
5107 "test/dwconv-microkernel-tester.h",
5108 "src/xnnpack/AlignedAllocator.h",
5109 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005110 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005111)
5112
5113xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005114 name = "f16_dwconv_minmax_test",
5115 srcs = [
5116 "test/f16-dwconv-minmax.cc",
5117 "test/dwconv-microkernel-tester.h",
5118 "src/xnnpack/AlignedAllocator.h",
5119 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005120 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005121)
5122
5123xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005124 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005125 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005126 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005127 "test/dwconv-microkernel-tester.h",
5128 "src/xnnpack/AlignedAllocator.h",
5129 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005130 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005131)
5132
5133xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005134 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005135 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005136 "test/f32-dwconv-chw.cc",
5137 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005138 "src/xnnpack/AlignedAllocator.h",
5139 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005140 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005141)
5142
5143xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005144 name = "f16_gavgpool_minmax_test",
5145 srcs = [
5146 "test/f16-gavgpool-minmax.cc",
5147 "test/gavgpool-microkernel-tester.h",
5148 "src/xnnpack/AlignedAllocator.h",
5149 ] + MICROKERNEL_TEST_HDRS,
5150 deps = MICROKERNEL_TEST_DEPS,
5151)
5152
5153xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005154 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005155 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005156 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005157 "test/gavgpool-microkernel-tester.h",
5158 "src/xnnpack/AlignedAllocator.h",
5159 ] + MICROKERNEL_TEST_HDRS,
5160 deps = MICROKERNEL_TEST_DEPS,
5161)
5162
5163xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005164 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005165 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005166 "test/f32-gavgpool-cw.cc",
5167 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005168 "src/xnnpack/AlignedAllocator.h",
5169 ] + MICROKERNEL_TEST_HDRS,
5170 deps = MICROKERNEL_TEST_DEPS,
5171)
5172
5173xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005174 name = "f32_gemm_test",
5175 srcs = [
5176 "test/f32-gemm.cc",
5177 "test/gemm-microkernel-tester.h",
5178 "src/xnnpack/AlignedAllocator.h",
5179 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005180 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005181)
5182
5183xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005184 name = "f32_gemm_relu_test",
5185 srcs = [
5186 "test/f32-gemm-relu.cc",
5187 "test/gemm-microkernel-tester.h",
5188 "src/xnnpack/AlignedAllocator.h",
5189 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005190 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005191)
5192
5193xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005194 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005195 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005196 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005197 "test/gemm-microkernel-tester.h",
5198 "src/xnnpack/AlignedAllocator.h",
5199 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005200 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005201)
5202
5203xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005204 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005205 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005206 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005207 "test/gemm-microkernel-tester.h",
5208 "src/xnnpack/AlignedAllocator.h",
5209 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005210 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005211)
5212
5213xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005214 name = "f16_hswish_test",
5215 srcs = [
5216 "test/f16-hswish.cc",
5217 "test/hswish-microkernel-tester.h",
5218 ] + MICROKERNEL_TEST_HDRS,
5219 deps = MICROKERNEL_TEST_DEPS,
5220)
5221
5222xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005223 name = "f32_hswish_test",
5224 srcs = [
5225 "test/f32-hswish.cc",
5226 "test/hswish-microkernel-tester.h",
5227 ] + MICROKERNEL_TEST_HDRS,
5228 deps = MICROKERNEL_TEST_DEPS,
5229)
5230
5231xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005232 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005233 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005234 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005235 "test/maxpool-microkernel-tester.h",
5236 ] + MICROKERNEL_TEST_HDRS,
5237 deps = MICROKERNEL_TEST_DEPS,
5238)
5239
5240xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005241 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005242 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005243 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005244 "test/avgpool-microkernel-tester.h",
5245 "src/xnnpack/AlignedAllocator.h",
5246 ] + MICROKERNEL_TEST_HDRS,
5247 deps = MICROKERNEL_TEST_DEPS,
5248)
5249
5250xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005251 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005252 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005253 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005254 "test/gemm-microkernel-tester.h",
5255 "src/xnnpack/AlignedAllocator.h",
5256 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005257 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005258)
5259
5260xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005261 name = "f16_prelu_test",
5262 srcs = [
5263 "test/f16-prelu.cc",
5264 "test/prelu-microkernel-tester.h",
5265 "src/xnnpack/AlignedAllocator.h",
5266 ] + MICROKERNEL_TEST_HDRS,
5267 deps = MICROKERNEL_TEST_DEPS,
5268)
5269
5270xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005271 name = "f32_prelu_test",
5272 srcs = [
5273 "test/f32-prelu.cc",
5274 "test/prelu-microkernel-tester.h",
5275 "src/xnnpack/AlignedAllocator.h",
5276 ] + MICROKERNEL_TEST_HDRS,
5277 deps = MICROKERNEL_TEST_DEPS,
5278)
5279
5280xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005281 name = "f32_raddexpminusmax_test",
5282 srcs = [
5283 "test/f32-raddexpminusmax.cc",
5284 "test/raddexpminusmax-microkernel-tester.h",
5285 ] + MICROKERNEL_TEST_HDRS,
5286 deps = MICROKERNEL_TEST_DEPS,
5287)
5288
5289xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005290 name = "f32_raddextexp_test",
5291 srcs = [
5292 "test/f32-raddextexp.cc",
5293 "test/raddextexp-microkernel-tester.h",
5294 ] + MICROKERNEL_TEST_HDRS,
5295 deps = MICROKERNEL_TEST_DEPS,
5296)
5297
5298xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005299 name = "f32_raddstoreexpminusmax_test",
5300 srcs = [
5301 "test/f32-raddstoreexpminusmax.cc",
5302 "test/raddstoreexpminusmax-microkernel-tester.h",
5303 ] + MICROKERNEL_TEST_HDRS,
5304 deps = MICROKERNEL_TEST_DEPS,
5305)
5306
5307xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005308 name = "f32_relu_test",
5309 srcs = [
5310 "test/f32-relu.cc",
5311 "test/vunary-microkernel-tester.h",
5312 ] + MICROKERNEL_TEST_HDRS,
5313 deps = MICROKERNEL_TEST_DEPS,
5314)
5315
5316xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005317 name = "f32_rmax_test",
5318 srcs = [
5319 "test/f32-rmax.cc",
5320 "test/rmax-microkernel-tester.h",
5321 ] + MICROKERNEL_TEST_HDRS,
5322 deps = MICROKERNEL_TEST_DEPS,
5323)
5324
5325xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005326 name = "f32_sigmoid_test",
5327 srcs = [
5328 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005329 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005330 ] + MICROKERNEL_TEST_HDRS,
5331 deps = MICROKERNEL_TEST_DEPS,
5332)
5333
5334xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005335 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005336 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005337 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005338 "test/spmm-microkernel-tester.h",
5339 "src/xnnpack/AlignedAllocator.h",
5340 ] + MICROKERNEL_TEST_HDRS,
5341 deps = MICROKERNEL_TEST_DEPS,
5342)
5343
5344xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005345 name = "f16_vadd_minmax_test",
5346 srcs = [
5347 "test/f16-vadd-minmax.cc",
5348 "test/vbinary-microkernel-tester.h",
5349 ] + MICROKERNEL_TEST_HDRS,
5350 deps = MICROKERNEL_TEST_DEPS,
5351)
5352
5353xnnpack_unit_test(
5354 name = "f16_vaddc_minmax_test",
5355 srcs = [
5356 "test/f16-vaddc-minmax.cc",
5357 "test/vbinaryc-microkernel-tester.h",
5358 ] + MICROKERNEL_TEST_HDRS,
5359 deps = MICROKERNEL_TEST_DEPS,
5360)
5361
5362xnnpack_unit_test(
5363 name = "f16_vdiv_minmax_test",
5364 srcs = [
5365 "test/f16-vdiv-minmax.cc",
5366 "test/vbinary-microkernel-tester.h",
5367 ] + MICROKERNEL_TEST_HDRS,
5368 deps = MICROKERNEL_TEST_DEPS,
5369)
5370
5371xnnpack_unit_test(
5372 name = "f16_vdivc_minmax_test",
5373 srcs = [
5374 "test/f16-vdivc-minmax.cc",
5375 "test/vbinaryc-microkernel-tester.h",
5376 ] + MICROKERNEL_TEST_HDRS,
5377 deps = MICROKERNEL_TEST_DEPS,
5378)
5379
5380xnnpack_unit_test(
5381 name = "f16_vrdivc_minmax_test",
5382 srcs = [
5383 "test/f16-vrdivc-minmax.cc",
5384 "test/vbinaryc-microkernel-tester.h",
5385 ] + MICROKERNEL_TEST_HDRS,
5386 deps = MICROKERNEL_TEST_DEPS,
5387)
5388
5389xnnpack_unit_test(
5390 name = "f16_vmax_test",
5391 srcs = [
5392 "test/f16-vmax.cc",
5393 "test/vbinary-microkernel-tester.h",
5394 ] + MICROKERNEL_TEST_HDRS,
5395 deps = MICROKERNEL_TEST_DEPS,
5396)
5397
5398xnnpack_unit_test(
5399 name = "f16_vmaxc_test",
5400 srcs = [
5401 "test/f16-vmaxc.cc",
5402 "test/vbinaryc-microkernel-tester.h",
5403 ] + MICROKERNEL_TEST_HDRS,
5404 deps = MICROKERNEL_TEST_DEPS,
5405)
5406
5407xnnpack_unit_test(
5408 name = "f16_vmin_test",
5409 srcs = [
5410 "test/f16-vmin.cc",
5411 "test/vbinary-microkernel-tester.h",
5412 ] + MICROKERNEL_TEST_HDRS,
5413 deps = MICROKERNEL_TEST_DEPS,
5414)
5415
5416xnnpack_unit_test(
5417 name = "f16_vminc_test",
5418 srcs = [
5419 "test/f16-vminc.cc",
5420 "test/vbinaryc-microkernel-tester.h",
5421 ] + MICROKERNEL_TEST_HDRS,
5422 deps = MICROKERNEL_TEST_DEPS,
5423)
5424
5425xnnpack_unit_test(
5426 name = "f16_vmul_minmax_test",
5427 srcs = [
5428 "test/f16-vmul-minmax.cc",
5429 "test/vbinary-microkernel-tester.h",
5430 ] + MICROKERNEL_TEST_HDRS,
5431 deps = MICROKERNEL_TEST_DEPS,
5432)
5433
5434xnnpack_unit_test(
5435 name = "f16_vmulc_minmax_test",
5436 srcs = [
5437 "test/f16-vmulc-minmax.cc",
5438 "test/vbinaryc-microkernel-tester.h",
5439 ] + MICROKERNEL_TEST_HDRS,
5440 deps = MICROKERNEL_TEST_DEPS,
5441)
5442
5443xnnpack_unit_test(
5444 name = "f16_vsub_minmax_test",
5445 srcs = [
5446 "test/f16-vsub-minmax.cc",
5447 "test/vbinary-microkernel-tester.h",
5448 ] + MICROKERNEL_TEST_HDRS,
5449 deps = MICROKERNEL_TEST_DEPS,
5450)
5451
5452xnnpack_unit_test(
5453 name = "f16_vsubc_minmax_test",
5454 srcs = [
5455 "test/f16-vsubc-minmax.cc",
5456 "test/vbinaryc-microkernel-tester.h",
5457 ] + MICROKERNEL_TEST_HDRS,
5458 deps = MICROKERNEL_TEST_DEPS,
5459)
5460
5461xnnpack_unit_test(
5462 name = "f16_vrsubc_minmax_test",
5463 srcs = [
5464 "test/f16-vrsubc-minmax.cc",
5465 "test/vbinaryc-microkernel-tester.h",
5466 ] + MICROKERNEL_TEST_HDRS,
5467 deps = MICROKERNEL_TEST_DEPS,
5468)
5469
5470xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005471 name = "f32_vabs_test",
5472 srcs = [
5473 "test/f32-vabs.cc",
5474 "test/vunary-microkernel-tester.h",
5475 ] + MICROKERNEL_TEST_HDRS,
5476 deps = MICROKERNEL_TEST_DEPS,
5477)
5478
5479xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005480 name = "f32_vadd_test",
5481 srcs = [
5482 "test/f32-vadd.cc",
5483 "test/vbinary-microkernel-tester.h",
5484 ] + MICROKERNEL_TEST_HDRS,
5485 deps = MICROKERNEL_TEST_DEPS,
5486)
5487
5488xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005489 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005490 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005491 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005492 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005493 ] + MICROKERNEL_TEST_HDRS,
5494 deps = MICROKERNEL_TEST_DEPS,
5495)
5496
5497xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005498 name = "f32_vadd_relu_test",
5499 srcs = [
5500 "test/f32-vadd-relu.cc",
5501 "test/vbinary-microkernel-tester.h",
5502 ] + MICROKERNEL_TEST_HDRS,
5503 deps = MICROKERNEL_TEST_DEPS,
5504)
5505
5506xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005507 name = "f32_vaddc_test",
5508 srcs = [
5509 "test/f32-vaddc.cc",
5510 "test/vbinaryc-microkernel-tester.h",
5511 ] + MICROKERNEL_TEST_HDRS,
5512 deps = MICROKERNEL_TEST_DEPS,
5513)
5514
5515xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005516 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005517 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005518 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005519 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005520 ] + MICROKERNEL_TEST_HDRS,
5521 deps = MICROKERNEL_TEST_DEPS,
5522)
5523
5524xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005525 name = "f32_vaddc_relu_test",
5526 srcs = [
5527 "test/f32-vaddc-relu.cc",
5528 "test/vbinaryc-microkernel-tester.h",
5529 ] + MICROKERNEL_TEST_HDRS,
5530 deps = MICROKERNEL_TEST_DEPS,
5531)
5532
5533xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005534 name = "f32_vdiv_test",
5535 srcs = [
5536 "test/f32-vdiv.cc",
5537 "test/vbinary-microkernel-tester.h",
5538 ] + MICROKERNEL_TEST_HDRS,
5539 deps = MICROKERNEL_TEST_DEPS,
5540)
5541
5542xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005543 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005544 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005545 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005546 "test/vbinary-microkernel-tester.h",
5547 ] + MICROKERNEL_TEST_HDRS,
5548 deps = MICROKERNEL_TEST_DEPS,
5549)
5550
5551xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005552 name = "f32_vdiv_relu_test",
5553 srcs = [
5554 "test/f32-vdiv-relu.cc",
5555 "test/vbinary-microkernel-tester.h",
5556 ] + MICROKERNEL_TEST_HDRS,
5557 deps = MICROKERNEL_TEST_DEPS,
5558)
5559
5560xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005561 name = "f32_vdivc_test",
5562 srcs = [
5563 "test/f32-vdivc.cc",
5564 "test/vbinaryc-microkernel-tester.h",
5565 ] + MICROKERNEL_TEST_HDRS,
5566 deps = MICROKERNEL_TEST_DEPS,
5567)
5568
5569xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005570 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005571 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005572 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005573 "test/vbinaryc-microkernel-tester.h",
5574 ] + MICROKERNEL_TEST_HDRS,
5575 deps = MICROKERNEL_TEST_DEPS,
5576)
5577
5578xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005579 name = "f32_vdivc_relu_test",
5580 srcs = [
5581 "test/f32-vdivc-relu.cc",
5582 "test/vbinaryc-microkernel-tester.h",
5583 ] + MICROKERNEL_TEST_HDRS,
5584 deps = MICROKERNEL_TEST_DEPS,
5585)
5586
5587xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005588 name = "f32_vrdivc_test",
5589 srcs = [
5590 "test/f32-vrdivc.cc",
5591 "test/vbinaryc-microkernel-tester.h",
5592 ] + MICROKERNEL_TEST_HDRS,
5593 deps = MICROKERNEL_TEST_DEPS,
5594)
5595
5596xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005597 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005598 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005599 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005600 "test/vbinaryc-microkernel-tester.h",
5601 ] + MICROKERNEL_TEST_HDRS,
5602 deps = MICROKERNEL_TEST_DEPS,
5603)
5604
5605xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005606 name = "f32_vrdivc_relu_test",
5607 srcs = [
5608 "test/f32-vrdivc-relu.cc",
5609 "test/vbinaryc-microkernel-tester.h",
5610 ] + MICROKERNEL_TEST_HDRS,
5611 deps = MICROKERNEL_TEST_DEPS,
5612)
5613
5614xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005615 name = "f32_vmax_test",
5616 srcs = [
5617 "test/f32-vmax.cc",
5618 "test/vbinary-microkernel-tester.h",
5619 ] + MICROKERNEL_TEST_HDRS,
5620 deps = MICROKERNEL_TEST_DEPS,
5621)
5622
5623xnnpack_unit_test(
5624 name = "f32_vmaxc_test",
5625 srcs = [
5626 "test/f32-vmaxc.cc",
5627 "test/vbinaryc-microkernel-tester.h",
5628 ] + MICROKERNEL_TEST_HDRS,
5629 deps = MICROKERNEL_TEST_DEPS,
5630)
5631
5632xnnpack_unit_test(
5633 name = "f32_vmin_test",
5634 srcs = [
5635 "test/f32-vmin.cc",
5636 "test/vbinary-microkernel-tester.h",
5637 ] + MICROKERNEL_TEST_HDRS,
5638 deps = MICROKERNEL_TEST_DEPS,
5639)
5640
5641xnnpack_unit_test(
5642 name = "f32_vminc_test",
5643 srcs = [
5644 "test/f32-vminc.cc",
5645 "test/vbinaryc-microkernel-tester.h",
5646 ] + MICROKERNEL_TEST_HDRS,
5647 deps = MICROKERNEL_TEST_DEPS,
5648)
5649
5650xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005651 name = "f32_vmul_test",
5652 srcs = [
5653 "test/f32-vmul.cc",
5654 "test/vbinary-microkernel-tester.h",
5655 ] + MICROKERNEL_TEST_HDRS,
5656 deps = MICROKERNEL_TEST_DEPS,
5657)
5658
5659xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005660 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005661 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005662 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005663 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005664 ] + MICROKERNEL_TEST_HDRS,
5665 deps = MICROKERNEL_TEST_DEPS,
5666)
5667
5668xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005669 name = "f32_vmul_relu_test",
5670 srcs = [
5671 "test/f32-vmul-relu.cc",
5672 "test/vbinary-microkernel-tester.h",
5673 ] + MICROKERNEL_TEST_HDRS,
5674 deps = MICROKERNEL_TEST_DEPS,
5675)
5676
5677xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005678 name = "f32_vmulc_test",
5679 srcs = [
5680 "test/f32-vmulc.cc",
5681 "test/vbinaryc-microkernel-tester.h",
5682 ] + MICROKERNEL_TEST_HDRS,
5683 deps = MICROKERNEL_TEST_DEPS,
5684)
5685
5686xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005687 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005688 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005689 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005690 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005691 ] + MICROKERNEL_TEST_HDRS,
5692 deps = MICROKERNEL_TEST_DEPS,
5693)
5694
5695xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005696 name = "f32_vmulc_relu_test",
5697 srcs = [
5698 "test/f32-vmulc-relu.cc",
5699 "test/vbinaryc-microkernel-tester.h",
5700 ] + MICROKERNEL_TEST_HDRS,
5701 deps = MICROKERNEL_TEST_DEPS,
5702)
5703
5704xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005705 name = "f16_vmulcaddc_minmax_test",
5706 srcs = [
5707 "test/f16-vmulcaddc-minmax.cc",
5708 "test/vmulcaddc-microkernel-tester.h",
5709 "src/xnnpack/AlignedAllocator.h",
5710 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005711 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005712)
5713
5714xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005715 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005716 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005717 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005718 "test/vmulcaddc-microkernel-tester.h",
5719 "src/xnnpack/AlignedAllocator.h",
5720 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005721 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005722)
5723
5724xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005725 name = "f32_vlrelu_test",
5726 srcs = [
5727 "test/f32-vlrelu.cc",
5728 "test/vunary-microkernel-tester.h",
5729 ] + MICROKERNEL_TEST_HDRS,
5730 deps = MICROKERNEL_TEST_DEPS,
5731)
5732
5733xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005734 name = "f32_vneg_test",
5735 srcs = [
5736 "test/f32-vneg.cc",
5737 "test/vunary-microkernel-tester.h",
5738 ] + MICROKERNEL_TEST_HDRS,
5739 deps = MICROKERNEL_TEST_DEPS,
5740)
5741
5742xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005743 name = "f32_vrndne_test",
5744 srcs = [
5745 "test/f32-vrndne.cc",
5746 "test/vunary-microkernel-tester.h",
5747 ] + MICROKERNEL_TEST_HDRS,
5748 deps = MICROKERNEL_TEST_DEPS,
5749)
5750
5751xnnpack_unit_test(
5752 name = "f32_vrndz_test",
5753 srcs = [
5754 "test/f32-vrndz.cc",
5755 "test/vunary-microkernel-tester.h",
5756 ] + MICROKERNEL_TEST_HDRS,
5757 deps = MICROKERNEL_TEST_DEPS,
5758)
5759
5760xnnpack_unit_test(
5761 name = "f32_vrndu_test",
5762 srcs = [
5763 "test/f32-vrndu.cc",
5764 "test/vunary-microkernel-tester.h",
5765 ] + MICROKERNEL_TEST_HDRS,
5766 deps = MICROKERNEL_TEST_DEPS,
5767)
5768
5769xnnpack_unit_test(
5770 name = "f32_vrndd_test",
5771 srcs = [
5772 "test/f32-vrndd.cc",
5773 "test/vunary-microkernel-tester.h",
5774 ] + MICROKERNEL_TEST_HDRS,
5775 deps = MICROKERNEL_TEST_DEPS,
5776)
5777
5778xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005779 name = "f32_vscale_test",
5780 srcs = [
5781 "test/f32-vscale.cc",
5782 "test/vscale-microkernel-tester.h",
5783 ] + MICROKERNEL_TEST_HDRS,
5784 deps = MICROKERNEL_TEST_DEPS,
5785)
5786
5787xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005788 name = "f32_vscaleexpminusmax_test",
5789 srcs = [
5790 "test/f32-vscaleexpminusmax.cc",
5791 "test/vscaleexpminusmax-microkernel-tester.h",
5792 ] + MICROKERNEL_TEST_HDRS,
5793 deps = MICROKERNEL_TEST_DEPS,
5794)
5795
5796xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005797 name = "f32_vscaleextexp_test",
5798 srcs = [
5799 "test/f32-vscaleextexp.cc",
5800 "test/vscaleextexp-microkernel-tester.h",
5801 ] + MICROKERNEL_TEST_HDRS,
5802 deps = MICROKERNEL_TEST_DEPS,
5803)
5804
5805xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005806 name = "f32_vsqr_test",
5807 srcs = [
5808 "test/f32-vsqr.cc",
5809 "test/vunary-microkernel-tester.h",
5810 ] + MICROKERNEL_TEST_HDRS,
5811 deps = MICROKERNEL_TEST_DEPS,
5812)
5813
5814xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005815 name = "f32_vsqrdiff_test",
5816 srcs = [
5817 "test/f32-vsqrdiff.cc",
5818 "test/vbinary-microkernel-tester.h",
5819 ] + MICROKERNEL_TEST_HDRS,
5820 deps = MICROKERNEL_TEST_DEPS,
5821)
5822
5823xnnpack_unit_test(
5824 name = "f32_vsqrdiffc_test",
5825 srcs = [
5826 "test/f32-vsqrdiffc.cc",
5827 "test/vbinaryc-microkernel-tester.h",
5828 ] + MICROKERNEL_TEST_HDRS,
5829 deps = MICROKERNEL_TEST_DEPS,
5830)
5831
5832xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005833 name = "f32_vsqrt_test",
5834 srcs = [
5835 "test/f32-vsqrt.cc",
5836 "test/vunary-microkernel-tester.h",
5837 ] + MICROKERNEL_TEST_HDRS,
5838 deps = MICROKERNEL_TEST_DEPS,
5839)
5840
5841xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005842 name = "f32_vsub_test",
5843 srcs = [
5844 "test/f32-vsub.cc",
5845 "test/vbinary-microkernel-tester.h",
5846 ] + MICROKERNEL_TEST_HDRS,
5847 deps = MICROKERNEL_TEST_DEPS,
5848)
5849
5850xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005851 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005852 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005853 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005854 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005855 ] + MICROKERNEL_TEST_HDRS,
5856 deps = MICROKERNEL_TEST_DEPS,
5857)
5858
5859xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005860 name = "f32_vsub_relu_test",
5861 srcs = [
5862 "test/f32-vsub-relu.cc",
5863 "test/vbinary-microkernel-tester.h",
5864 ] + MICROKERNEL_TEST_HDRS,
5865 deps = MICROKERNEL_TEST_DEPS,
5866)
5867
5868xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005869 name = "f32_vsubc_test",
5870 srcs = [
5871 "test/f32-vsubc.cc",
5872 "test/vbinaryc-microkernel-tester.h",
5873 ] + MICROKERNEL_TEST_HDRS,
5874 deps = MICROKERNEL_TEST_DEPS,
5875)
5876
5877xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005878 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005879 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005880 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005881 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005882 ] + MICROKERNEL_TEST_HDRS,
5883 deps = MICROKERNEL_TEST_DEPS,
5884)
5885
5886xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005887 name = "f32_vsubc_relu_test",
5888 srcs = [
5889 "test/f32-vsubc-relu.cc",
5890 "test/vbinaryc-microkernel-tester.h",
5891 ] + MICROKERNEL_TEST_HDRS,
5892 deps = MICROKERNEL_TEST_DEPS,
5893)
5894
5895xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005896 name = "f32_vrsubc_test",
5897 srcs = [
5898 "test/f32-vrsubc.cc",
5899 "test/vbinaryc-microkernel-tester.h",
5900 ] + MICROKERNEL_TEST_HDRS,
5901 deps = MICROKERNEL_TEST_DEPS,
5902)
5903
5904xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005905 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005906 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005907 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005908 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005909 ] + MICROKERNEL_TEST_HDRS,
5910 deps = MICROKERNEL_TEST_DEPS,
5911)
5912
5913xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005914 name = "f32_vrsubc_relu_test",
5915 srcs = [
5916 "test/f32-vrsubc-relu.cc",
5917 "test/vbinaryc-microkernel-tester.h",
5918 ] + MICROKERNEL_TEST_HDRS,
5919 deps = MICROKERNEL_TEST_DEPS,
5920)
5921
5922xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005923 name = "qs8_dwconv_minmax_test",
5924 srcs = [
5925 "test/qs8-dwconv-minmax.cc",
5926 "test/dwconv-microkernel-tester.h",
5927 "src/xnnpack/AlignedAllocator.h",
5928 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5929 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5930)
5931
5932xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005933 name = "qs8_gavgpool_minmax_test",
5934 srcs = [
5935 "test/qs8-gavgpool-minmax.cc",
5936 "test/gavgpool-microkernel-tester.h",
5937 "src/xnnpack/AlignedAllocator.h",
5938 ] + MICROKERNEL_TEST_HDRS,
5939 deps = MICROKERNEL_TEST_DEPS,
5940)
5941
5942xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005943 name = "qs8_gemm_minmax_test",
5944 srcs = [
5945 "test/qs8-gemm-minmax.cc",
5946 "test/gemm-microkernel-tester.h",
5947 "src/xnnpack/AlignedAllocator.h",
5948 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5949 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5950)
5951
5952xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005953 name = "qs8_igemm_minmax_test",
5954 srcs = [
5955 "test/qs8-igemm-minmax.cc",
5956 "test/gemm-microkernel-tester.h",
5957 "src/xnnpack/AlignedAllocator.h",
5958 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5959 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5960)
5961
5962xnnpack_unit_test(
5963 name = "qs8_requantization_test",
5964 srcs = [
5965 "src/xnnpack/requantization-stubs.h",
5966 "test/qs8-requantization.cc",
5967 "test/requantization-tester.h",
5968 ] + MICROKERNEL_TEST_HDRS,
5969 deps = MICROKERNEL_TEST_DEPS,
5970)
5971
5972xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07005973 name = "qs8_vadd_minmax_test",
5974 srcs = [
5975 "test/qs8-vadd-minmax.cc",
5976 "test/vadd-microkernel-tester.h",
5977 ] + MICROKERNEL_TEST_HDRS,
5978 deps = MICROKERNEL_TEST_DEPS,
5979)
5980
5981xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07005982 name = "qs8_vaddc_minmax_test",
5983 srcs = [
5984 "test/qs8-vaddc-minmax.cc",
5985 "test/vaddc-microkernel-tester.h",
5986 ] + MICROKERNEL_TEST_HDRS,
5987 deps = MICROKERNEL_TEST_DEPS,
5988)
5989
5990xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005991 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005992 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005993 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005994 "test/avgpool-microkernel-tester.h",
5995 "src/xnnpack/AlignedAllocator.h",
5996 ] + MICROKERNEL_TEST_HDRS,
5997 deps = MICROKERNEL_TEST_DEPS,
5998)
5999
6000xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006001 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006002 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006003 "test/qu8-dwconv-minmax.cc",
6004 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006005 "src/xnnpack/AlignedAllocator.h",
6006 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006007 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006008)
6009
6010xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006011 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006012 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006013 "test/qu8-igemm-minmax.cc",
6014 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006015 "src/xnnpack/AlignedAllocator.h",
6016 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006017 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006018)
6019
6020xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006021 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006022 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006023 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006024 "test/gavgpool-microkernel-tester.h",
6025 "src/xnnpack/AlignedAllocator.h",
6026 ] + MICROKERNEL_TEST_HDRS,
6027 deps = MICROKERNEL_TEST_DEPS,
6028)
6029
6030xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006031 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006032 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006033 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006034 "test/gemm-microkernel-tester.h",
6035 "src/xnnpack/AlignedAllocator.h",
6036 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006037 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006038)
6039
6040xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006041 name = "qu8_requantization_test",
6042 srcs = [
6043 "src/xnnpack/requantization-stubs.h",
6044 "test/qu8-requantization.cc",
6045 "test/requantization-tester.h",
6046 ] + MICROKERNEL_TEST_HDRS,
6047 deps = MICROKERNEL_TEST_DEPS,
6048)
6049
6050xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006051 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006052 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006053 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006054 "test/vadd-microkernel-tester.h",
6055 ] + MICROKERNEL_TEST_HDRS,
6056 deps = MICROKERNEL_TEST_DEPS,
6057)
6058
6059xnnpack_unit_test(
6060 name = "u8_clamp_test",
6061 srcs = [
6062 "test/u8-clamp.cc",
6063 "test/clamp-microkernel-tester.h",
6064 ] + MICROKERNEL_TEST_HDRS,
6065 deps = MICROKERNEL_TEST_DEPS,
6066)
6067
6068xnnpack_unit_test(
6069 name = "u8_lut32norm_test",
6070 srcs = [
6071 "test/u8-lut32norm.cc",
6072 "test/lut-norm-microkernel-tester.h",
6073 ] + MICROKERNEL_TEST_HDRS,
6074 deps = MICROKERNEL_TEST_DEPS,
6075)
6076
6077xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006078 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006079 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006080 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006081 "test/maxpool-microkernel-tester.h",
6082 ] + MICROKERNEL_TEST_HDRS,
6083 deps = MICROKERNEL_TEST_DEPS,
6084)
6085
6086xnnpack_unit_test(
6087 name = "u8_rmax_test",
6088 srcs = [
6089 "test/u8-rmax.cc",
6090 "test/rmax-microkernel-tester.h",
6091 ] + MICROKERNEL_TEST_HDRS,
6092 deps = MICROKERNEL_TEST_DEPS,
6093)
6094
6095xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006096 name = "x32_fill_test",
6097 srcs = [
6098 "test/x32-fill.cc",
6099 "test/fill-microkernel-tester.h",
6100 ] + MICROKERNEL_TEST_HDRS,
6101 deps = MICROKERNEL_TEST_DEPS,
6102)
6103
6104xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006105 name = "x32_packx_test",
6106 srcs = [
6107 "test/x32-packx.cc",
6108 "test/pack-microkernel-tester.h",
6109 "src/xnnpack/AlignedAllocator.h",
6110 ] + MICROKERNEL_TEST_HDRS,
6111 deps = MICROKERNEL_TEST_DEPS,
6112)
6113
6114xnnpack_unit_test(
6115 name = "x32_pad_test",
6116 srcs = [
6117 "test/x32-pad.cc",
6118 "test/pad-microkernel-tester.h",
6119 ] + MICROKERNEL_TEST_HDRS,
6120 deps = MICROKERNEL_TEST_DEPS,
6121)
6122
6123xnnpack_unit_test(
6124 name = "x32_unpool_test",
6125 srcs = [
6126 "test/x32-unpool.cc",
6127 "test/unpool-microkernel-tester.h",
6128 ] + MICROKERNEL_TEST_HDRS,
6129 deps = MICROKERNEL_TEST_DEPS,
6130)
6131
6132xnnpack_unit_test(
6133 name = "x32_zip_test",
6134 srcs = [
6135 "test/x32-zip.cc",
6136 "test/zip-microkernel-tester.h",
6137 ] + MICROKERNEL_TEST_HDRS,
6138 deps = MICROKERNEL_TEST_DEPS,
6139)
6140
6141xnnpack_unit_test(
6142 name = "x8_lut_test",
6143 srcs = [
6144 "test/x8-lut.cc",
6145 "test/lut-microkernel-tester.h",
6146 ] + MICROKERNEL_TEST_HDRS,
6147 deps = MICROKERNEL_TEST_DEPS,
6148)
6149
6150xnnpack_unit_test(
6151 name = "x8_zip_test",
6152 srcs = [
6153 "test/x8-zip.cc",
6154 "test/zip-microkernel-tester.h",
6155 ] + MICROKERNEL_TEST_HDRS,
6156 deps = MICROKERNEL_TEST_DEPS,
6157)
6158
Marat Dukhan20c3b922020-03-10 03:45:06 -07006159########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006160
6161xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006162 name = "operator_size_test",
6163 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006164 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006165)
6166
Marat Dukhan20c3b922020-03-10 03:45:06 -07006167xnnpack_binary(
6168 name = "subgraph_size_test",
6169 srcs = ["test/subgraph-size.c"],
6170 deps = [":XNNPACK"],
6171)
6172
6173########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006174
6175xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006176 name = "abs_nc_test",
6177 srcs = [
6178 "test/abs-nc.cc",
6179 "test/abs-operator-tester.h",
6180 ],
6181 deps = OPERATOR_TEST_DEPS,
6182)
6183
6184xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006185 name = "add_nd_test",
6186 srcs = [
6187 "test/add-nd.cc",
6188 "test/binary-elementwise-operator-tester.h",
6189 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006190 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006191)
6192
6193xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006194 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006195 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006196 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006197 "test/argmax-pooling-operator-tester.h",
6198 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006199 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006200)
6201
6202xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006203 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006204 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006205 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006206 "test/average-pooling-operator-tester.h",
6207 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006208 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006209)
6210
6211xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006212 name = "bankers_rounding_nc_test",
6213 srcs = [
6214 "test/bankers-rounding-nc.cc",
6215 "test/bankers-rounding-operator-tester.h",
6216 ],
6217 deps = OPERATOR_TEST_DEPS,
6218)
6219
6220xnnpack_unit_test(
6221 name = "ceiling_nc_test",
6222 srcs = [
6223 "test/ceiling-nc.cc",
6224 "test/ceiling-operator-tester.h",
6225 ],
6226 deps = OPERATOR_TEST_DEPS,
6227)
6228
6229xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006230 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006231 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006232 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006233 "test/channel-shuffle-operator-tester.h",
6234 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006235 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006236)
6237
6238xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006239 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006240 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006241 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006242 "test/clamp-operator-tester.h",
6243 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006244 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006245)
6246
6247xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006248 name = "constant_pad_nd_test",
6249 srcs = [
6250 "test/constant-pad-nd.cc",
6251 "test/constant-pad-operator-tester.h",
6252 ],
6253 deps = OPERATOR_TEST_DEPS,
6254)
6255
6256xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006257 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006258 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006259 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006260 "test/convolution-operator-tester.h",
6261 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006262 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006263)
6264
6265xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006266 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006267 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006268 "test/convolution-nchw.cc",
6269 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006270 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006271 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006272)
6273
6274xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006275 name = "copy_nc_test",
6276 srcs = [
6277 "test/copy-nc.cc",
6278 "test/copy-operator-tester.h",
6279 ],
6280 deps = OPERATOR_TEST_DEPS,
6281)
6282
6283xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006284 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006285 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006286 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006287 "test/deconvolution-operator-tester.h",
6288 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006289 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006290)
6291
6292xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006293 name = "divide_nd_test",
6294 srcs = [
6295 "test/binary-elementwise-operator-tester.h",
6296 "test/divide-nd.cc",
6297 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006298 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006299)
6300
6301xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006302 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006303 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006304 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006305 "test/fully-connected-operator-tester.h",
6306 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006307 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006308)
6309
6310xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006311 name = "floor_nc_test",
6312 srcs = [
6313 "test/floor-nc.cc",
6314 "test/floor-operator-tester.h",
6315 ],
6316 deps = OPERATOR_TEST_DEPS,
6317)
6318
6319xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006320 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006321 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006322 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006323 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006324 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006325 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006326)
6327
6328xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006329 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006330 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006331 "test/global-average-pooling-ncw.cc",
6332 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006333 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006334 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006335)
6336
6337xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006338 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006339 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006340 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006341 "test/hardswish-operator-tester.h",
6342 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006343 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006344)
6345
6346xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006347 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006348 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006349 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006350 "test/leaky-relu-operator-tester.h",
6351 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006352 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006353)
6354
6355xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006356 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006357 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006358 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006359 "test/max-pooling-operator-tester.h",
6360 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006361 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006362)
6363
6364xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006365 name = "maximum_nd_test",
6366 srcs = [
6367 "test/binary-elementwise-operator-tester.h",
6368 "test/maximum-nd.cc",
6369 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006370 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006371)
6372
6373xnnpack_unit_test(
6374 name = "minimum_nd_test",
6375 srcs = [
6376 "test/binary-elementwise-operator-tester.h",
6377 "test/minimum-nd.cc",
6378 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006379 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006380)
6381
6382xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006383 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006384 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006385 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006386 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006387 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006388 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006389)
6390
6391xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006392 name = "negate_nc_test",
6393 srcs = [
6394 "test/negate-nc.cc",
6395 "test/negate-operator-tester.h",
6396 ],
6397 deps = OPERATOR_TEST_DEPS,
6398)
6399
6400xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006401 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006402 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006403 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006404 "test/prelu-operator-tester.h",
6405 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006406 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006407)
6408
6409xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006410 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006411 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006412 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006413 "test/resize-bilinear-operator-tester.h",
6414 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006415 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006416)
6417
6418xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006419 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006420 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006421 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006422 "test/sigmoid-operator-tester.h",
6423 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006424 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006425)
6426
6427xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006428 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006429 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006430 "test/softmax-nc.cc",
6431 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006432 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006433 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006434)
6435
6436xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006437 name = "square_nc_test",
6438 srcs = [
6439 "test/square-nc.cc",
6440 "test/square-operator-tester.h",
6441 ],
6442 deps = OPERATOR_TEST_DEPS,
6443)
6444
6445xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006446 name = "square_root_nc_test",
6447 srcs = [
6448 "test/square-root-nc.cc",
6449 "test/square-root-operator-tester.h",
6450 ],
6451 deps = OPERATOR_TEST_DEPS,
6452)
6453
6454xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006455 name = "squared_difference_nd_test",
6456 srcs = [
6457 "test/binary-elementwise-operator-tester.h",
6458 "test/squared-difference-nd.cc",
6459 ],
6460 deps = OPERATOR_TEST_DEPS,
6461)
6462
6463xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006464 name = "subtract_nd_test",
6465 srcs = [
6466 "test/binary-elementwise-operator-tester.h",
6467 "test/subtract-nd.cc",
6468 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006469 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006470)
6471
6472xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006473 name = "truncation_nc_test",
6474 srcs = [
6475 "test/truncation-nc.cc",
6476 "test/truncation-operator-tester.h",
6477 ],
6478 deps = OPERATOR_TEST_DEPS,
6479)
6480
6481xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006482 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006483 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006484 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006485 "test/unpooling-operator-tester.h",
6486 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006487 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006488)
6489
Chao Mei6ddfc602020-05-13 22:29:36 -07006490############################### Misc unit tests ###############################
6491
6492xnnpack_unit_test(
6493 name = "memory_planner_test",
6494 srcs = [
6495 "test/memory-planner-test.cc",
6496 ],
6497 deps = [
6498 ":XNNPACK",
6499 ":memory_planner",
6500 ],
6501)
6502
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07006503xnnpack_unit_test(
6504 name = "subgraph_nchw_test",
6505 srcs = [
6506 "src/xnnpack/subgraph.h",
6507 "test/subgraph-nchw.cc",
6508 "test/subgraph-tester.h",
6509 ],
6510 deps = [
6511 ":XNNPACK",
6512 ],
6513)
6514
Marat Dukhan08c4a432019-10-03 09:29:21 -07006515############################# Build configurations #############################
6516
Marat Dukhanb8642352019-10-30 15:43:02 -07006517# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006518config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006519 name = "xnn_enable_assembly_explicit_true",
6520 define_values = {"xnn_enable_assembly": "true"},
6521)
6522
6523# Disables usage of assembly kernels.
6524config_setting(
6525 name = "xnn_enable_assembly_explicit_false",
6526 define_values = {"xnn_enable_assembly": "false"},
6527)
6528
Marat Dukhan9de90e02020-06-18 16:04:12 -07006529# Enables usage of sparse inference.
6530config_setting(
6531 name = "xnn_enable_sparse_explicit_true",
6532 define_values = {"xnn_enable_sparse": "true"},
6533)
6534
6535# Disables usage of sparse inference.
6536config_setting(
6537 name = "xnn_enable_sparse_explicit_false",
6538 define_values = {"xnn_enable_sparse": "false"},
6539)
6540
Marat Dukhan05702cf2020-03-26 15:41:33 -07006541# Disables usage of HMP-aware optimizations.
6542config_setting(
6543 name = "xnn_enable_hmp_explicit_false",
6544 define_values = {"xnn_enable_hmp": "false"},
6545)
6546
Chao Mei6ddfc602020-05-13 22:29:36 -07006547# Enable usage of optimized memory allocation
6548config_setting(
6549 name = "xnn_enable_memopt_explicit_true",
6550 define_values = {"xnnpack_enable_memopt": "true"},
6551)
6552
6553# Disable usage of optimized memory allocation
6554config_setting(
6555 name = "xnn_enable_memopt_explicit_false",
6556 define_values = {"xnnpack_enable_memopt": "false"},
6557)
6558
Marat Dukhanb8642352019-10-30 15:43:02 -07006559# Builds with -c dbg
6560config_setting(
6561 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006562 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006563 "compilation_mode": "dbg",
6564 },
6565)
6566
6567# Builds with -c opt
6568config_setting(
6569 name = "optimized_build",
6570 values = {
6571 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006572 },
6573)
6574
6575config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006576 name = "linux_k8",
6577 values = {"cpu": "k8"},
6578)
6579
6580config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006581 name = "linux_arm",
6582 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006583)
6584
6585config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006586 name = "linux_armeabi",
6587 values = {"cpu": "armeabi"},
6588)
6589
6590config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006591 name = "linux_armhf",
6592 values = {"cpu": "armhf"},
6593)
6594
6595config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006596 name = "linux_armv7a",
6597 values = {"cpu": "armv7a"},
6598)
6599
6600config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006601 name = "linux_aarch64",
6602 values = {"cpu": "aarch64"},
6603)
6604
6605config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006606 name = "android",
6607 values = {"crosstool_top": "//external:android/crosstool"},
6608)
6609
6610config_setting(
6611 name = "android_armv7",
6612 values = {
6613 "crosstool_top": "//external:android/crosstool",
6614 "cpu": "armeabi-v7a",
6615 },
6616)
6617
6618config_setting(
6619 name = "android_arm64",
6620 values = {
6621 "crosstool_top": "//external:android/crosstool",
6622 "cpu": "arm64-v8a",
6623 },
6624)
6625
6626config_setting(
6627 name = "android_x86",
6628 values = {
6629 "crosstool_top": "//external:android/crosstool",
6630 "cpu": "x86",
6631 },
6632)
6633
6634config_setting(
6635 name = "android_x86_64",
6636 values = {
6637 "crosstool_top": "//external:android/crosstool",
6638 "cpu": "x86_64",
6639 },
6640)
6641
6642config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006643 name = "windows_x86_64",
6644 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006645)
6646
6647config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006648 name = "windows_x86_64_clang",
6649 values = {
6650 "compiler": "clang-cl",
6651 "cpu": "x64_windows",
6652 },
6653)
6654
6655config_setting(
6656 name = "windows_x86_64_mingw",
6657 values = {
6658 "compiler": "mingw-gcc",
6659 "cpu": "x64_windows",
6660 },
6661)
6662
6663config_setting(
6664 name = "windows_x86_64_msys",
6665 values = {
6666 "compiler": "msys-gcc",
6667 "cpu": "x64_windows",
6668 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006669)
6670
6671config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006672 name = "macos_x86_64",
6673 values = {
6674 "apple_platform_type": "macos",
6675 "cpu": "darwin",
6676 },
6677)
6678
6679config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006680 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006681 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006682)
6683
6684config_setting(
6685 name = "emscripten_wasm",
6686 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006687 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006688 "cpu": "wasm",
6689 },
6690)
6691
6692config_setting(
6693 name = "emscripten_wasmsimd",
6694 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006695 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006696 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006697 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006698 },
6699)
6700
6701config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006702 name = "ios_armv7",
6703 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006704 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006705 "cpu": "ios_armv7",
6706 },
6707)
6708
6709config_setting(
6710 name = "ios_arm64",
6711 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006712 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006713 "cpu": "ios_arm64",
6714 },
6715)
6716
6717config_setting(
6718 name = "ios_arm64e",
6719 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006720 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006721 "cpu": "ios_arm64e",
6722 },
6723)
6724
6725config_setting(
6726 name = "ios_x86",
6727 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006728 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006729 "cpu": "ios_i386",
6730 },
6731)
6732
6733config_setting(
6734 name = "ios_x86_64",
6735 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006736 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006737 "cpu": "ios_x86_64",
6738 },
6739)
6740
6741config_setting(
6742 name = "watchos_armv7k",
6743 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006744 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006745 "cpu": "watchos_armv7k",
6746 },
6747)
6748
6749config_setting(
6750 name = "watchos_arm64_32",
6751 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006752 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006753 "cpu": "watchos_arm64_32",
6754 },
6755)
6756
6757config_setting(
6758 name = "watchos_x86",
6759 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006760 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006761 "cpu": "watchos_i386",
6762 },
6763)
6764
6765config_setting(
6766 name = "watchos_x86_64",
6767 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006768 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006769 "cpu": "watchos_x86_64",
6770 },
6771)
6772
6773config_setting(
6774 name = "tvos_arm64",
6775 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006776 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006777 "cpu": "tvos_arm64",
6778 },
6779)
6780
6781config_setting(
6782 name = "tvos_x86_64",
6783 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006784 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006785 "cpu": "tvos_x86_64",
6786 },
6787)