blob: ff2069d51b98994669a147e4ad276534c38bedfe [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",
Artsiom Ablavatski97918102020-10-27 15:52:59 -070068 "src/operators/resize-bilinear-nchw.c",
Marat Dukhane8265432020-04-28 18:42:59 -070069 "src/operators/resize-bilinear-nhwc.c",
70 "src/operators/sigmoid-nc.c",
71 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070072 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070073 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070074]
75
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070076SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070077 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070078 "src/subgraph/add2.c",
79 "src/subgraph/argmax-pooling-2d.c",
80 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070081 "src/subgraph/bankers-rounding.c",
82 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070083 "src/subgraph/clamp.c",
84 "src/subgraph/convolution-2d.c",
85 "src/subgraph/deconvolution-2d.c",
86 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070087 "src/subgraph/divide.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -070089 "src/subgraph/fully-connected.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070090 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070091 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070092 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070093 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070094 "src/subgraph/maximum2.c",
95 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070096 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070097 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070098 "src/subgraph/prelu.c",
99 "src/subgraph/sigmoid.c",
100 "src/subgraph/softmax.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700101 "src/subgraph/square-root.c",
102 "src/subgraph/square.c",
103 "src/subgraph/squared-difference.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700104 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700105 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700106 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700107 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700108 "src/subgraph/unpooling-2d.c",
109]
110
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800111TABLE_SRCS = [
112 "src/tables/exp2-k-over-64.c",
113 "src/tables/exp2-k-over-2048.c",
Marat Dukhan1f256fc2020-09-24 21:27:14 -0700114 "src/tables/exp2minus-k-over-64.c",
115 "src/tables/exp2minus-k-over-2048.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800116]
117
Marat Dukhan08c4a432019-10-03 09:29:21 -0700118SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800119 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800120 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800121 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700122 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
123 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700124 "src/f32-clamp/gen/scalar-x1.c",
125 "src/f32-clamp/gen/scalar-x2.c",
126 "src/f32-clamp/gen/scalar-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700127 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700128 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700129 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700130 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700131 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
132 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
133 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700135 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
136 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
137 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700138 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700139 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
140 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
141 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700142 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700143 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
144 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
145 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700146 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700147 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
148 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
149 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700150 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700151 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
152 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
153 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700154 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c",
155 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c",
156 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700157 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700158 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700159 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c",
160 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c",
161 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c",
162 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c",
163 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700164 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c",
165 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c",
166 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700167 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700168 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700169 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c",
170 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c",
171 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700172 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c",
173 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c",
174 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c",
175 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700176 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700177 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c",
178 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700179 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700180 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700181 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c",
Marat Dukhan29c0c332020-10-28 22:11:00 -0700182 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c",
183 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c",
184 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c",
185 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c",
186 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c",
187 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c",
188 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c",
189 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c",
190 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c",
191 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700192 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700193 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
194 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700195 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
196 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
197 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700198 "src/f32-gemm/gen/1x4-minmax-scalar.c",
199 "src/f32-gemm/gen/1x4-relu-scalar.c",
200 "src/f32-gemm/gen/1x4-scalar.c",
201 "src/f32-gemm/gen/2x4-minmax-scalar.c",
202 "src/f32-gemm/gen/2x4-relu-scalar.c",
203 "src/f32-gemm/gen/2x4-scalar.c",
204 "src/f32-gemm/gen/4x2-minmax-scalar.c",
205 "src/f32-gemm/gen/4x2-relu-scalar.c",
206 "src/f32-gemm/gen/4x2-scalar.c",
207 "src/f32-gemm/gen/4x4-minmax-scalar.c",
208 "src/f32-gemm/gen/4x4-relu-scalar.c",
209 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700210 "src/f32-hswish/gen/hswish-scalar-x1.c",
211 "src/f32-hswish/gen/hswish-scalar-x2.c",
212 "src/f32-hswish/gen/hswish-scalar-x4.c",
XNNPACK Team21432672020-10-19 19:58:48 -0700213 "src/f32-ibilinear-chw/gen/scalar-p1.c",
214 "src/f32-ibilinear-chw/gen/scalar-p2.c",
215 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700216 "src/f32-ibilinear/gen/scalar-c1.c",
217 "src/f32-ibilinear/gen/scalar-c2.c",
218 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700219 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700220 "src/f32-igemm/gen/1x4-relu-scalar.c",
221 "src/f32-igemm/gen/1x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700222 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700223 "src/f32-igemm/gen/2x4-relu-scalar.c",
224 "src/f32-igemm/gen/2x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700225 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700226 "src/f32-igemm/gen/4x2-relu-scalar.c",
227 "src/f32-igemm/gen/4x2-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700228 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700229 "src/f32-igemm/gen/4x4-relu-scalar.c",
230 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700231 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
232 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
233 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700234 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
235 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
236 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
237 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800238 "src/f32-prelu/gen/scalar-2x1.c",
239 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800240 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800241 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700242 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800243 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
244 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700245 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800246 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800247 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700248 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800249 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
250 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700251 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700252 "src/f32-relu/gen/scalar-x1.c",
253 "src/f32-relu/gen/scalar-x2.c",
254 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700255 "src/f32-relu/gen/scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700256 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800257 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
258 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
259 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700260 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
261 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
262 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800263 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
264 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
265 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700266 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
267 "src/f32-spmm/gen/1x1-minmax-scalar.c",
268 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
269 "src/f32-spmm/gen/2x1-minmax-scalar.c",
270 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
271 "src/f32-spmm/gen/4x1-minmax-scalar.c",
272 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
273 "src/f32-spmm/gen/8x1-minmax-scalar.c",
274 "src/f32-spmm/gen/8x2-minmax-scalar.c",
275 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700276 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
277 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
278 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700279 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700280 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
281 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
282 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700283 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700284 "src/f32-vbinary/gen/vadd-scalar-x1.c",
285 "src/f32-vbinary/gen/vadd-scalar-x2.c",
286 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700287 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700288 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
289 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
290 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700291 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700292 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
293 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
294 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700295 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700296 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
297 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
298 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700299 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700300 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
301 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
302 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700303 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700304 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
305 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
306 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700307 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700308 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
309 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
310 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700311 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700312 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
313 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
314 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700315 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700316 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
317 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
318 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700319 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700320 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
321 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
322 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700323 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800324 "src/f32-vbinary/gen/vmax-scalar-x1.c",
325 "src/f32-vbinary/gen/vmax-scalar-x2.c",
326 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700327 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800328 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
329 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
330 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700331 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800332 "src/f32-vbinary/gen/vmin-scalar-x1.c",
333 "src/f32-vbinary/gen/vmin-scalar-x2.c",
334 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700335 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800336 "src/f32-vbinary/gen/vminc-scalar-x1.c",
337 "src/f32-vbinary/gen/vminc-scalar-x2.c",
338 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700339 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700340 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
341 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
342 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700343 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700344 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
345 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
346 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700347 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700348 "src/f32-vbinary/gen/vmul-scalar-x1.c",
349 "src/f32-vbinary/gen/vmul-scalar-x2.c",
350 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700351 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700352 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
353 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
354 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700355 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700356 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
357 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
358 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700359 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700360 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
361 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
362 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700363 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700364 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
365 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
366 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700367 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700368 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
369 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
370 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700371 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700372 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
373 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
374 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700375 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700376 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
377 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
378 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700379 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700380 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
381 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
382 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700383 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700384 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
385 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
386 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700387 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700388 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
389 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
390 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700391 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700392 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
393 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
394 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700395 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700396 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
397 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
398 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700399 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700400 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
401 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
402 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700403 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700404 "src/f32-vbinary/gen/vsub-scalar-x1.c",
405 "src/f32-vbinary/gen/vsub-scalar-x2.c",
406 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700407 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700408 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
409 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
410 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700411 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700412 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
413 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
414 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700415 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700416 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
417 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
418 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700419 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700420 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
421 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
422 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700423 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
424 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
425 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700426 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
427 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
428 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700429 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
430 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
431 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
432 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
433 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
434 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
435 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
436 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
437 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700438 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
439 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
440 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700441 "src/f32-vunary/gen/vabs-scalar-x1.c",
442 "src/f32-vunary/gen/vabs-scalar-x2.c",
443 "src/f32-vunary/gen/vabs-scalar-x4.c",
444 "src/f32-vunary/gen/vneg-scalar-x1.c",
445 "src/f32-vunary/gen/vneg-scalar-x2.c",
446 "src/f32-vunary/gen/vneg-scalar-x4.c",
447 "src/f32-vunary/gen/vsqr-scalar-x1.c",
448 "src/f32-vunary/gen/vsqr-scalar-x2.c",
449 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800450 "src/math/expminus-scalar-lut64-p2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700451 "src/math/expminus-scalar-lut2048-p1.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800452 "src/math/expminus-scalar-p5.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700453 "src/math/roundd-scalar-addsub.c",
454 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700455 "src/math/roundd-scalar-floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700456 "src/math/roundne-scalar-addsub.c",
457 "src/math/roundne-scalar-nearbyint.c",
458 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700459 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700460 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700461 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700462 "src/math/roundz-scalar-addsub.c",
463 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700464 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700465 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700466 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700467 "src/math/sigmoid-scalar-rr2-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700468 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
469 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700470 "src/qs8-requantization/fp32-scalar-lrintf.c",
471 "src/qs8-requantization/fp32-scalar-magic.c",
472 "src/qs8-requantization/precise-scalar-signed64.c",
473 "src/qs8-requantization/precise-scalar-unsigned32.c",
474 "src/qs8-requantization/precise-scalar-unsigned64.c",
475 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700476 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
477 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
478 "src/qu8-dwconv/up1x9-minmax-scalar.c",
479 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
480 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700481 "src/qu8-gemm/2x2-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700482 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
483 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700484 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700485 "src/qu8-requantization/fp32-scalar-lrintf.c",
486 "src/qu8-requantization/fp32-scalar-magic.c",
487 "src/qu8-requantization/precise-scalar-signed64.c",
488 "src/qu8-requantization/precise-scalar-unsigned32.c",
489 "src/qu8-requantization/precise-scalar-unsigned64.c",
490 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700491 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700492 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700493 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700494 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700495 "src/u8-rmax/scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700496 "src/x8-lut/scalar.c",
497 "src/x8-zip/x2-scalar.c",
498 "src/x8-zip/x3-scalar.c",
499 "src/x8-zip/x4-scalar.c",
500 "src/x8-zip/xm-scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700501 "src/x32-fill/scalar-float.c",
502 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700503 "src/x32-packx/x2-scalar.c",
504 "src/x32-packx/x3-scalar.c",
505 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700506 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700507 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700508 "src/x32-unpool/scalar.c",
509 "src/x32-zip/x2-scalar.c",
510 "src/x32-zip/x3-scalar.c",
511 "src/x32-zip/x4-scalar.c",
512 "src/x32-zip/xm-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700513]
514
Marat Dukhan436ebe62019-12-04 15:10:12 -0800515WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700516 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
517 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700518 "src/f32-clamp/gen/wasm-x1.c",
519 "src/f32-clamp/gen/wasm-x2.c",
520 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700521 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
522 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700523 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
524 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700525 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
526 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700527 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
528 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700529 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
530 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700531 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
532 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700533 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
534 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700535 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
536 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700537 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
538 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700539 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
540 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700541 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
542 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700543 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
544 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700545 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
546 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700547 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
548 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
549 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
550 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700551 "src/f32-gemm/gen/1x4-relu-wasm.c",
552 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700553 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700554 "src/f32-gemm/gen/2x4-relu-wasm.c",
555 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700556 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700557 "src/f32-gemm/gen/4x2-relu-wasm.c",
558 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700559 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700560 "src/f32-gemm/gen/4x4-relu-wasm.c",
561 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700562 "src/f32-hswish/gen/hswish-wasm-x1.c",
563 "src/f32-hswish/gen/hswish-wasm-x2.c",
564 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700565 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700566 "src/f32-igemm/gen/1x4-relu-wasm.c",
567 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700568 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700569 "src/f32-igemm/gen/2x4-relu-wasm.c",
570 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700571 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700572 "src/f32-igemm/gen/4x2-relu-wasm.c",
573 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700574 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700575 "src/f32-igemm/gen/4x4-relu-wasm.c",
576 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700577 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
578 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
579 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700580 "src/f32-prelu/gen/wasm-2x1.c",
581 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700582 "src/f32-relu/gen/wasm-x1.c",
583 "src/f32-relu/gen/wasm-x2.c",
584 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700585 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700586 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
587 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
588 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700589 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700590 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
591 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
592 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700593 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700594 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
595 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
596 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
597 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700598 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
599 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
600 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700601 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700602 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
603 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
604 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
605 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700606 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
607 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
608 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700609 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700610 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
611 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
612 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
613 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700614 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
615 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
616 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700617 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800618 "src/f32-vbinary/gen/vmax-wasm-x1.c",
619 "src/f32-vbinary/gen/vmax-wasm-x2.c",
620 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700621 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800622 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
623 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
624 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700625 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800626 "src/f32-vbinary/gen/vmin-wasm-x1.c",
627 "src/f32-vbinary/gen/vmin-wasm-x2.c",
628 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700629 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800630 "src/f32-vbinary/gen/vminc-wasm-x1.c",
631 "src/f32-vbinary/gen/vminc-wasm-x2.c",
632 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700633 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700634 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
635 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
636 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700637 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700638 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
639 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
640 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700641 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700642 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
643 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
644 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
645 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700646 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
647 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
648 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700649 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700650 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
651 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
652 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
653 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700654 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
655 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
656 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700657 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700658 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
659 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
660 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
661 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700662 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
663 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
664 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700665 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700666 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
667 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
668 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
669 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700670 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
671 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
672 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700673 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700674 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
675 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
676 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
677 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700678 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
679 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
680 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700681 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700682 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
683 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
684 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700685 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
686 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
687 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800688]
689
Marat Dukhan290055c2020-06-09 12:24:29 -0700690WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700691 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
692 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
693 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700694 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
695 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
696 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
697 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700698 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
699 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
700 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
701 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700702 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700703 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700704 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700705 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700706 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
707 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700708 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700709 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700710 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700711 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
712 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
713 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
714 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
715 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
716 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
717 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700718 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700719 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700720 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700721 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
722 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700723 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700724 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700725 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700726 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
727 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
728 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
729 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
730 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
731 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
732 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
733 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700734 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
735 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
736 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
737 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700738 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
739 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
740 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
741 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
742 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
743 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
744 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
745 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
746 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
747 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
748 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
749 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
750 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
751 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
752 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
753 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
754 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
755 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
756 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
757 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
758 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
759 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
760 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
761 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
762 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
763 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
764 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
765 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
766 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
767 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
768 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
769 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
770 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
771 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
772 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
773 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
774 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
775 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
776 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
777 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
778 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
779 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
780 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
781 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
782 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
783 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
784 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
785 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
786 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
787 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
788 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
789 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
790 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
791 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
792 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
793 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
794 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
795 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
796 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
797 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
798 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
799 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
800 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
801 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
802 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
803 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
804 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
805 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
806 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
807 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700808 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
809 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
810 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
XNNPACK Team965272b2020-10-23 21:10:15 -0700811 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
812 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700813 "src/f32-ibilinear/gen/wasmsimd-c4.c",
814 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700815 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
816 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
817 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
818 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
819 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
820 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
821 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
822 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
823 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
824 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
825 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
826 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
827 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
828 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
829 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
830 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
831 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
832 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
833 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
834 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
835 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
836 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
837 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
838 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
839 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
840 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
841 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
842 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
843 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
844 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
845 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
846 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
847 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
848 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
849 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
850 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
851 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
852 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
853 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
854 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700855 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
856 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700857 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
858 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
859 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
860 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700861 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
862 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700863 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
864 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
865 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700866 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
867 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700868 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
869 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
870 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
871 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
872 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
873 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
874 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700875 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
876 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700877 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
878 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
879 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
880 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700881 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700882 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700883 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700884 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
885 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700886 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700887 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
888 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700889 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700890 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
891 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700892 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700893 "src/f32-relu/gen/wasmsimd-x4.c",
894 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700895 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700896 "src/f32-rmax/wasmsimd-arm.c",
897 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700898 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
899 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
900 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
901 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
902 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
903 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700904 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
905 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
906 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
907 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
908 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
909 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
910 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-unroll2.c",
911 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-unroll4.c",
912 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
913 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-unroll2.c",
914 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-unroll4.c",
915 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
916 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-unroll2.c",
917 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-unroll4.c",
918 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
919 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-unroll2.c",
920 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-unroll4.c",
921 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
922 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-unroll2.c",
923 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-unroll4.c",
924 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
925 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-unroll2.c",
926 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-unroll4.c",
927 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Frank Barchard846c0c62020-10-26 15:01:39 -0700928 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-unroll2.c",
929 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-unroll4.c",
930 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
931 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-unroll2.c",
932 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-unroll4.c",
933 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700934 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
935 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700936 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700937 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
938 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700939 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700940 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
941 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700942 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700943 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
944 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700945 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700946 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
947 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700948 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700949 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
950 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700951 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700952 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
953 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700954 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700955 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
956 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700957 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700958 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
959 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700960 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700961 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
962 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700963 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700964 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
965 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700966 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700967 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
968 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700969 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700970 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
971 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700972 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700973 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
974 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700975 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700976 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
977 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700978 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700979 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
980 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700981 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700982 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
983 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700984 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700985 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
986 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700987 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700988 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
989 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700990 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700991 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
992 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700993 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700994 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
995 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700996 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700997 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
998 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700999 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001000 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
1001 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001002 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001003 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
1004 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001005 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001006 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
1007 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001008 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001009 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
1010 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001011 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001012 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
1013 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001014 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001015 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
1016 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001017 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001018 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
1019 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001020 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001021 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
1022 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001023 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001024 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
1025 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001026 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001027 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
1028 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001029 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001030 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
1031 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001032 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001033 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
1034 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001035 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001036 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
1037 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001038 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001039 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
1040 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001041 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001042 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1043 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001044 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001045 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1046 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001047 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001048 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1049 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001050 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001051 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1052 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001053 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001054 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1055 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001056 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001057 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1058 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001059 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001060 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1061 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001062 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001063 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1064 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001065 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001066 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1067 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001068 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001069 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1070 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001071 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001072 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1073 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001074 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001075 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1076 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001077 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001078 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1079 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001080 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001081 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1082 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001083 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001084 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1085 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1086 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1087 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001088 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001089 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001090 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001091 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001092 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1093 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1094 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1095 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001096 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1097 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1098 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1099 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1100 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1101 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1102 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1103 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1104 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1105 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001106 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1107 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001108 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1109 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1110 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1111 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1112 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1113 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001114 "src/math/roundd-wasmsimd-addsub.c",
1115 "src/math/roundd-wasmsimd-cvt.c",
1116 "src/math/roundne-wasmsimd-addsub.c",
1117 "src/math/roundu-wasmsimd-addsub.c",
1118 "src/math/roundu-wasmsimd-cvt.c",
1119 "src/math/roundz-wasmsimd-addsub.c",
1120 "src/math/roundz-wasmsimd-cvt.c",
1121 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1122 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001123 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1124 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1125 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001126 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1127 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1128 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001129 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1130 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1131 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001132 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001133 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001134 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001135 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1136 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001137 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001138 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1139 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001140 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1141 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001142 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001143 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001144 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001145 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001146 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001147 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001148 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001149 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1150 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1151 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1152 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1153 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1154 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1155 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1156 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001157 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001158 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001159 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001160 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001161 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001162 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001163 "src/x32-zip/x2-wasmsimd.c",
1164 "src/x32-zip/x3-wasmsimd.c",
1165 "src/x32-zip/x4-wasmsimd.c",
1166 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001167]
1168
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001169PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -07001170 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Marat Dukhanbf715f92020-10-23 20:17:00 -07001171 "src/f32-dwconv2d-chw/3x3p1-psimd-1x4-acc3.c",
1172 "src/f32-dwconv2d-chw/3x3s2p1-psimd-1x4-acc3.c",
1173 "src/f32-dwconv2d-chw/5x5p2-psimd-3x4.c",
1174 "src/f32-dwconv2d-chw/5x5s2p2-psimd-1x4-acc2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001175]
1176
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001177PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001178 "src/qs8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001179 "src/qs8-requantization/precise-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001180 "src/qu8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001181 "src/qu8-requantization/precise-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001182]
1183
Marat Dukhan08c4a432019-10-03 09:29:21 -07001184# ISA-specific micro-kernels
1185NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001186 "src/f32-argmaxpool/4x-neon-c4.c",
1187 "src/f32-argmaxpool/9p8x-neon-c4.c",
1188 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001189 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1190 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001191 "src/f32-clamp/gen/neon-x4.c",
1192 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001193 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001194 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001195 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001196 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001197 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001198 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001199 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001200 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001201 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001202 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001203 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001204 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001205 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001206 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1207 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1208 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1209 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1210 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001211 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001212 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001213 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1214 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1215 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001216 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001217 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001218 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1219 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1220 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1221 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1222 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001223 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1224 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1225 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001226 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001227 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001228 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1229 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1230 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001231 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1232 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1233 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1234 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001235 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001236 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1237 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001238 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001239 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001240 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001241 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001242 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1243 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001244 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001245 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1246 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001247 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001248 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1249 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001250 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001251 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1252 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1253 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1254 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1255 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001256 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1257 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001258 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1259 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001260 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1261 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001262 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1263 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1264 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1265 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1266 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1267 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1268 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1269 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1270 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1271 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1272 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1273 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1274 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1275 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1276 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1277 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001278 "src/f32-hswish/gen/hswish-neon-x4.c",
1279 "src/f32-hswish/gen/hswish-neon-x8.c",
1280 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001281 "src/f32-ibilinear/gen/neon-c4.c",
1282 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001283 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001284 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001285 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001286 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1287 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001288 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001289 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1290 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1291 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1292 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001293 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1294 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001295 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1296 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001297 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1298 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001299 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1300 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1301 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001302 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1303 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001304 "src/f32-prelu/gen/neon-1x4.c",
1305 "src/f32-prelu/gen/neon-1x8.c",
1306 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001307 "src/f32-prelu/gen/neon-2x4.c",
1308 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001309 "src/f32-prelu/gen/neon-2x16.c",
1310 "src/f32-prelu/gen/neon-4x4.c",
1311 "src/f32-prelu/gen/neon-4x8.c",
1312 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001313 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001314 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001315 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001316 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1317 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001318 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001319 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1320 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001321 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001322 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1323 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001324 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1325 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1326 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1327 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1328 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1329 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1330 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1331 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1332 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1333 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1334 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1335 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1336 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001337 "src/f32-relu/gen/neon-x4.c",
1338 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001339 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001340 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001341 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1342 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1343 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1344 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1345 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1346 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1347 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1348 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1349 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1350 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1351 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1352 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001353 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1354 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1355 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1356 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1357 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1358 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001359 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1360 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1361 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1362 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001363 "src/f32-vbinary/gen/vmax-neon-x4.c",
1364 "src/f32-vbinary/gen/vmax-neon-x8.c",
1365 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1366 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1367 "src/f32-vbinary/gen/vmin-neon-x4.c",
1368 "src/f32-vbinary/gen/vmin-neon-x8.c",
1369 "src/f32-vbinary/gen/vminc-neon-x4.c",
1370 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001371 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1372 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1373 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1374 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1375 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1376 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001377 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1378 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1379 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1380 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001381 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1382 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1383 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1384 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001385 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1386 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001387 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1388 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001389 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1390 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001391 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1392 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1393 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1394 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1395 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1396 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001397 "src/f32-vunary/gen/vabs-neon-x4.c",
1398 "src/f32-vunary/gen/vabs-neon-x8.c",
1399 "src/f32-vunary/gen/vneg-neon-x4.c",
1400 "src/f32-vunary/gen/vneg-neon-x8.c",
1401 "src/f32-vunary/gen/vsqr-neon-x4.c",
1402 "src/f32-vunary/gen/vsqr-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001403 "src/math/roundd-neon-addsub.c",
1404 "src/math/roundd-neon-cvt.c",
1405 "src/math/roundne-neon-addsub.c",
1406 "src/math/roundu-neon-addsub.c",
1407 "src/math/roundu-neon-cvt.c",
1408 "src/math/roundz-neon-addsub.c",
1409 "src/math/roundz-neon-cvt.c",
1410 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
1411 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1412 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1413 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1414 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1415 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1416 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1417 "src/math/sqrt-neon-nr1rsqrts.c",
1418 "src/math/sqrt-neon-nr2rsqrts.c",
1419 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001420 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1421 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1422 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1423 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001424 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1425 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1426 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1427 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001428 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1429 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1430 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1431 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001432 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001433 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001434 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001435 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1436 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001437 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001438 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001439 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001440 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001441 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001442 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001443 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1444 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1445 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1446 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1447 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1448 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1449 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1450 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001451 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1452 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1453 "src/qu8-dwconv/up8x9-minmax-neon.c",
1454 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1455 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1456 "src/qu8-gemm/4x8-minmax-neon.c",
1457 "src/qu8-gemm/8x8-minmax-neon.c",
1458 "src/qu8-igemm/4x8-minmax-neon.c",
1459 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001460 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001461 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001462 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001463 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001464 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001465 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001466 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001467 "src/x8-zip/x2-neon.c",
1468 "src/x8-zip/x3-neon.c",
1469 "src/x8-zip/x4-neon.c",
1470 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001471 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001472 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001473 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001474 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001475 "src/x32-zip/x2-neon.c",
1476 "src/x32-zip/x3-neon.c",
1477 "src/x32-zip/x4-neon.c",
1478 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001479]
1480
1481NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001482 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1483 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1484 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1485 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1486 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1487 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1488 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1489 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1490 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1491 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1492 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1493 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1494 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1495 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1496 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1497 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1498 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1499 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1500 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1501 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1502 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1503 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1504 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1505 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1506 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1507 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1508 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1509 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1510 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1511 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001512 "src/f32-ibilinear/gen/neonfma-c4.c",
1513 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001514 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001515 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001516 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001517 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1518 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001519 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1520 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001521 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1522 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001523 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1524 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001525 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001526 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001527 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001528 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1529 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001530 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001531 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1532 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001533 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001534 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1535 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001536 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1537 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1538 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1539 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1540 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1541 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1542 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1543 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1544 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1545 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1546 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1547 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1548 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001549 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1550 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1551 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1552 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1553 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1554 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001555 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1556 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1557 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1558 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1559 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1560 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001561 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1562 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1563 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1564 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1565 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1566 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001567 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1568 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1569 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1570 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1571 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1572 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001573 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1574 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1575 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1576 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1577 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1578 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001579 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1580 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1581 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1582 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1583 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1584 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001585 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1586 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1587 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1588 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1589 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1590 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1591 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1592 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1593 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1594 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1595 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1596 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1597 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1598 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1599 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1600 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1601 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1602 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001603 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1604 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001605 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1606 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1607 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1608 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1609 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1610 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1611 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1612 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1613 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1614 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1615 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1616 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1617 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1618 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1619 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1620 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1621 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1622 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1623 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1624 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001625 "src/math/exp-neonfma-lut64-p2.c",
1626 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001627 "src/math/expminus-neonfma-lut64-p2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001628 "src/math/expminus-neonfma-lut2048-p1.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001629 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001630 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1631 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1632 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001633 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1634 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1635 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001636 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1637 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1638 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001639 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1640 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1641 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001642 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1643 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1644 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001645 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1646 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1647 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001648 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001649 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001650 "src/math/sqrt-neonfma-nr2fma.c",
1651 "src/math/sqrt-neonfma-nr2fma1adj.c",
1652 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001653]
1654
1655AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001656 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001657 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001658 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001659 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001660 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001661 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001662 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001663 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001664 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001665 "src/f32-dwconv2d-chw/5x5s2p2-neonfma-1x4-acc2.c",
1666 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
1667 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
1668 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001669 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001670 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001671 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
1672 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
1673 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
1674 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
1675 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001676 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
1677 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
1678 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001679 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001680 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001681 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
1682 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
1683 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001684 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
1685 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
1686 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
1687 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001688 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001689 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
1690 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001691 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001692 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001693 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001694 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001695 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
1696 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001697 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1698 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1699 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1700 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1701 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1702 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1703 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1704 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1705 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1706 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1707 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1708 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1709 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1710 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1711 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1712 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1713 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1714 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1715 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1716 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001717 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1718 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1719 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1720 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1721 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1722 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1723 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1724 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1725 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1726 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1727 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1728 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001729 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1730 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1731 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1732 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1733 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1734 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001735 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1736 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1737 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1738 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1739 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1740 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1741 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1742 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1743 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1744 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001745 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1746 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1747 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1748 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1749 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1750 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1751 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1752 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001753 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
1754 "src/f32-spmm/gen/32x1-minmax-neonfma-unroll2.c",
1755 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
1756 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
1757 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001758 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1759 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1760 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1761 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1762 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1763 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001764 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1765 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001766 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001767 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001768 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001769 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001770 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001771 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001772]
1773
Marat Dukhan8853b822020-05-07 12:19:01 -07001774NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001775 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1776 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001777 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1778 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1779 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1780 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1781 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1782 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001783 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001784 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001785 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001786 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001787]
1788
Marat Dukhan08c4a432019-10-03 09:29:21 -07001789AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001790 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001791 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001792 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1793 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1794 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1795 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001796 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1797 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1798 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1799 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1800 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1801 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1802 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1803 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001804 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1805 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001806 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1807 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1808 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1809 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1810 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1811 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1812 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
1813 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1814 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1815 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1816 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1817 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1818 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1819 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1820 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1821 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001822 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001823 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1824 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1825 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1826 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1827 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1828 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1829 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1830 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1831 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001832 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001833 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001834 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001835 "src/f16-relu/gen/neonfp16arith-x16.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001836 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001837 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001838 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001839 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001840 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001841 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001842 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001843 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1844 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1845 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1846 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1847 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1848 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1849 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1850 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1851 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1852 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1853 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1854 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1855 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1856 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1857 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1858 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1859 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1860 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1861 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1862 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1863 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1864 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1865 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1866 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1867 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1868 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
1869 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1870 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1871 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1872 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
1873 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001874]
1875
Benoit Jacoba9644732020-08-13 12:48:55 -07001876NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07001877 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001878 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001879 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001880 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001881 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001882 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001883 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001884 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001885 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001886 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001887 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001888 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001889 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001890 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001891 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001892 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001893 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001894 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001895]
1896
Marat Dukhan08c4a432019-10-03 09:29:21 -07001897SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001898 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1899 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001900 "src/f32-clamp/gen/sse-x4.c",
1901 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001902 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1903 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001904 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1905 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1906 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1907 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001908 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1909 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001910 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1911 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1912 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1913 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001914 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1915 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001916 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
1917 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
1918 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07001919 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001920 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07001921 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
1922 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
1923 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
1924 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
1925 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07001926 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
1927 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
1928 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001929 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07001930 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001931 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
1932 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
1933 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001934 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001935 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1936 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001937 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1938 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1939 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1940 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1941 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1942 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001943 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1944 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1945 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1946 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1947 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1948 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1949 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001950 "src/f32-hswish/gen/hswish-sse-x4.c",
1951 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001952 "src/f32-ibilinear/gen/sse-c4.c",
1953 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001954 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1955 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1956 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1957 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1958 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1959 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1960 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001961 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1962 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1963 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001964 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001965 "src/f32-prelu/gen/sse-2x4.c",
1966 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001967 "src/f32-relu/gen/sse-x4.c",
1968 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001969 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001970 "src/f32-spmm/gen/4x1-minmax-sse.c",
1971 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001972 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001973 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001974 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1975 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1976 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1977 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1978 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1979 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1980 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1981 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001982 "src/f32-vbinary/gen/vmax-sse-x4.c",
1983 "src/f32-vbinary/gen/vmax-sse-x8.c",
1984 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1985 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1986 "src/f32-vbinary/gen/vmin-sse-x4.c",
1987 "src/f32-vbinary/gen/vmin-sse-x8.c",
1988 "src/f32-vbinary/gen/vminc-sse-x4.c",
1989 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001990 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1991 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1992 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1993 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1994 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1995 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1996 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1997 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001998 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1999 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2000 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2001 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002002 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2003 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2004 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2005 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002006 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2007 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002008 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2009 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002010 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2011 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002012 "src/f32-vunary/gen/vabs-sse-x4.c",
2013 "src/f32-vunary/gen/vabs-sse-x8.c",
2014 "src/f32-vunary/gen/vneg-sse-x4.c",
2015 "src/f32-vunary/gen/vneg-sse-x8.c",
2016 "src/f32-vunary/gen/vsqr-sse-x4.c",
2017 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002018 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002019 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002020 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002021 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002022 "src/math/sqrt-sse-hh1mac.c",
2023 "src/math/sqrt-sse-nr1mac.c",
2024 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002025 "src/x32-fill/sse.c",
2026 "src/x32-packx/x4-sse.c",
2027 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002028]
2029
2030SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002031 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002032 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002033 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002034 "src/f32-prelu/gen/sse2-2x4.c",
2035 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002036 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002037 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002038 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002039 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2040 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002041 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002042 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2043 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002044 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002045 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2046 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002047 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002048 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
2049 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
2050 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
2051 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
2052 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
2053 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002054 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002055 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002056 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002057 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002058 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
2059 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002060 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2061 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002062 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2063 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002064 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2065 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2066 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2067 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2068 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2069 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
2070 "src/math/exp-sse2-lut64-p2.c",
2071 "src/math/exp-sse2-p5.c",
2072 "src/math/expminus-sse2-p5.c",
2073 "src/math/roundd-sse2-cvt.c",
2074 "src/math/roundne-sse2-cvt.c",
2075 "src/math/roundu-sse2-cvt.c",
2076 "src/math/roundz-sse2-cvt.c",
2077 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2078 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2079 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2080 "src/math/sigmoid-sse2-rr2-p5-div.c",
2081 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2082 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002083 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2084 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2085 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002086 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2087 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2088 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002089 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2090 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2091 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002092 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002093 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002094 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002095 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002096 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002097 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002098 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2099 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002100 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002101 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2102 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002103 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002104 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2105 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2106 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002107 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002108 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002109 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002110 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002111 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002112 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002113 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002114 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002115 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2116 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002117 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002118 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002119 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002120 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2121 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2122 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2123 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002124 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2125 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2126 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2127 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002128 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2129 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002130 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2131 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2132 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2133 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2134 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002135 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002136 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002137 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002138 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002139 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002140 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002141 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002142 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002143 "src/x8-zip/x2-sse2.c",
2144 "src/x8-zip/x3-sse2.c",
2145 "src/x8-zip/x4-sse2.c",
2146 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002147 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002148 "src/x32-zip/x2-sse2.c",
2149 "src/x32-zip/x3-sse2.c",
2150 "src/x32-zip/x4-sse2.c",
2151 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002152]
2153
2154SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002155 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2156 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2157 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002158 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002159 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002160 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2161 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2162 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2163 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2164 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002165 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2166 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2167 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002168 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2169 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2170 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002171 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2172 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2173 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002174 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002175 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002176 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002177 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002178 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002179 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002180 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2181 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002182 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002183 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2184 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002185 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002186 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2187 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2188 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002189 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002190 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002191 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002192 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002193 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002194 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002195 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002196 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002197 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2198 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002199 "src/qs8-requantization/precise-ssse3.c",
2200 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002201 "src/qu8-requantization/precise-ssse3.c",
2202 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002203]
2204
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002205SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002206 "src/f32-prelu/gen/sse41-2x4.c",
2207 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002208 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2209 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2210 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2211 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2212 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2213 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002214 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002215 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002216 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002217 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002218 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2219 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002220 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2221 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002222 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2223 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002224 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2225 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2226 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2227 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2228 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2229 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2230 "src/math/roundd-sse41.c",
2231 "src/math/roundne-sse41.c",
2232 "src/math/roundu-sse41.c",
2233 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002234 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2235 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2236 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002237 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2238 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2239 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002240 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2241 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2242 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002243 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002244 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002245 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002246 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002247 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002248 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002249 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2250 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002251 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002252 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2253 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002254 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002255 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2256 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2257 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002258 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002259 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002260 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002261 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002262 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002263 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002264 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002265 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002266 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2267 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002268 "src/qs8-requantization/fp32-sse4.c",
2269 "src/qs8-requantization/precise-sse4.c",
2270 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002271 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2272 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2273 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2274 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002275 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2276 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2277 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2278 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002279 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2280 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2281 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2282 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002283 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2284 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2285 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2286 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002287 "src/qu8-requantization/precise-sse4.c",
2288 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002289]
2290
Marat Dukhan08c4a432019-10-03 09:29:21 -07002291AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002292 "src/f32-clamp/gen/avx-x8.c",
2293 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002294 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2295 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002296 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2297 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002298 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2299 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002300 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2301 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2302 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2303 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2304 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2305 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002306 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002307 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2308 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002309 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002310 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002311 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002312 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002313 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2314 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2315 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2316 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2317 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2318 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2319 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2320 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2321 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2322 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2323 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002324 "src/f32-hswish/gen/hswish-avx-x8.c",
2325 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002326 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002327 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2328 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002329 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002330 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002331 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002332 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002333 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2334 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002335 "src/f32-prelu/gen/avx-2x8.c",
2336 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002337 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002338 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002339 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002340 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2341 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2342 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2343 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2344 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2345 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2346 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2347 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2348 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2349 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2350 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2351 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2352 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2353 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2354 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2355 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2356 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2357 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2358 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2359 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002360 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2361 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2362 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2363 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2364 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2365 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2366 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2367 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002368 "src/f32-vbinary/gen/vmax-avx-x8.c",
2369 "src/f32-vbinary/gen/vmax-avx-x16.c",
2370 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2371 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2372 "src/f32-vbinary/gen/vmin-avx-x8.c",
2373 "src/f32-vbinary/gen/vmin-avx-x16.c",
2374 "src/f32-vbinary/gen/vminc-avx-x8.c",
2375 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002376 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2377 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2378 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2379 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2380 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2381 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2382 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2383 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002384 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2385 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2386 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2387 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002388 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2389 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2390 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2391 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002392 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2393 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002394 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2395 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002396 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2397 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2398 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2399 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2400 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2401 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2402 "src/f32-vscale/avx-unroll32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002403 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2404 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002405 "src/f32-vunary/gen/vabs-avx-x8.c",
2406 "src/f32-vunary/gen/vabs-avx-x16.c",
2407 "src/f32-vunary/gen/vneg-avx-x8.c",
2408 "src/f32-vunary/gen/vneg-avx-x16.c",
2409 "src/f32-vunary/gen/vsqr-avx-x8.c",
2410 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhand08abcb2020-09-21 00:26:25 -07002411 "src/math/exp-avx-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002412 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002413 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002414 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2415 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002416]
2417
Marat Dukhan1566fee2020-08-02 21:55:41 -07002418XOP_UKERNELS = [
2419 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002420 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002421 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002422 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002423 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002424 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002425 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2426 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002427 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002428 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2429 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002430 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002431 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2432 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2433 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002434 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002435 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002436 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002437 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002438 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002439 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002440 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002441 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002442 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2443 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002444 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2445 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2446 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2447 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2448 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2449 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2450 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2451 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002452]
2453
Marat Dukhanfda12b82019-11-21 12:27:59 -08002454FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002455 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2456 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002457 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2458 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002459 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2460 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002461 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2462 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2463 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2464 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2465 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2466 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002467 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002468 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2469 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2470 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2471 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002472 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002473 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2474 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002475 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002476 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2477 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002478 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2479 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2480 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002481 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2482 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2483 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2484 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2485 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2486 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2487 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2488 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2489 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2490 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2491 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2492 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2493 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2494 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002495 "src/f32-hswish/gen/hswish-fma3-x8.c",
2496 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002497 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002498 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2499 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2500 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2501 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002502 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002503 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2504 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002505 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002506 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2507 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002508 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2509 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2510 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002511 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2512 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2513 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2514 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2515 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2516 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2517 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2518 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002519 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002520 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002521 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002522]
2523
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002524AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002525 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2526 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002527 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002528 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002529 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002530 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2531 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002532 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002533 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2534 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2535 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002536 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002537 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2538 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002539 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002540 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002541 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002542 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2543 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002544 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002545 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2546 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2547 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002548 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002549 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2550 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002551 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002552 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002553 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002554 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2555 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002556 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002557 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2558 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2559 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002560 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002561 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2562 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2563 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2564 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2565 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2566 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2567 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2568 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2569 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2570 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2571 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2572 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2573 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2574 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2575 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2576 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2577 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2578 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2579 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2580 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2581 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2582 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2583 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2584 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2585 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2586 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2587 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2588 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2589 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2590 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002591 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2592 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2593 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2594 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2595 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2596 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2597 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2598 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2599 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2600 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2601 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2602 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2603 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2604 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2605 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2606 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2607 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2608 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2609 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2610 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2611 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2612 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2613 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2614 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002615 "src/math/exp-avx2-lut8-p3-perm.c",
2616 "src/math/exp-avx2-lut8-p4-perm.c",
2617 "src/math/exp-avx2-p5.c",
2618 "src/math/expminus-avx2-p5.c",
2619 "src/math/extexp-avx2-p5.c",
2620 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2621 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2622 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2623 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2624 "src/math/sigmoid-avx2-rr1-p5-div.c",
2625 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
2626 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2627 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2628 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2629 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2630 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2631 "src/math/sigmoid-avx2-rr2-p5-div.c",
2632 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2633 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002634 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002635 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002636 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2637 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002638 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002639 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002640 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002641 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002642 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002643 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002644 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002645 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002646 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2647 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2648 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002649 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2650 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2651 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2652 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2653 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2654 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2655 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2656 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002657]
2658
Marat Dukhan08c4a432019-10-03 09:29:21 -07002659AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002660 "src/f32-clamp/gen/avx512f-x16.c",
2661 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002662 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2663 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002664 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2665 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002666 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2667 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002668 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2669 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2670 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2671 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2672 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2673 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002674 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2675 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2676 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2677 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2678 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2679 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002680 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2681 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2682 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2683 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2684 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2685 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002686 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2687 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002688 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2689 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2690 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2691 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2692 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2693 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002694 "src/f32-prelu/gen/avx512f-2x16.c",
2695 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002696 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2697 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002698 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002699 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002700 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002701 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2702 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002703 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002704 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2705 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2706 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002707 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002708 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2709 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002710 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002711 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002712 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002713 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2714 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002715 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002716 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2717 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2718 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002719 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002720 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2721 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002722 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002723 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002724 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002725 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2726 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002727 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002728 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2729 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2730 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002731 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002732 "src/f32-relu/gen/avx512f-x16.c",
2733 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002734 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002735 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2736 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2737 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2738 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2739 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2740 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2741 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2742 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2743 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2744 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2745 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2746 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2747 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2748 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2749 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2750 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002751 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2752 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2753 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2754 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2755 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2756 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2757 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2758 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2759 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2760 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2761 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2762 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2763 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2764 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2765 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2766 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002767 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2768 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2769 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2770 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2771 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2772 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2773 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2774 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2775 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2776 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2777 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2778 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2779 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2780 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2781 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2782 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002783 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2784 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2785 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2786 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2787 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2788 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2789 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2790 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002791 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2792 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2793 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2794 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2795 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2796 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2797 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2798 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002799 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2800 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2801 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2802 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2803 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2804 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2805 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2806 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002807 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2808 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2809 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2810 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002811 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2812 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2813 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2814 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002815 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2816 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002817 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2818 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
2819 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2820 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2821 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2822 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2823 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2824 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002825 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002826 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2827 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2828 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2829 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2830 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2831 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2832 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2833 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2834 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2835 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2836 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2837 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2838 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2839 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2840 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2841 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2842 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2843 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2844 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2845 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2846 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2847 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2848 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2849 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002850 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2851 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2852 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2853 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2854 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2855 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2856 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2857 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002858 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2859 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2860 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2861 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2862 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2863 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002864 "src/math/exp-avx512f-lut16-p3-perm-scalef.c",
2865 "src/math/exp-avx512f-lut16-p3-perm.c",
2866 "src/math/exp-avx512f-lut32-p2-perm2-scalef.c",
2867 "src/math/exp-avx512f-lut32-p2-perm2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002868 "src/math/exp-avx512f-p5-scalef.c",
2869 "src/math/exp-avx512f-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002870 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002871 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002872 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002873 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002874 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002875 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002876 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002877 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002878 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002879 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
2880 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
2881 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
2882 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
2883 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2884 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2885 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2886 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2887 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
2888 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002889 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002890 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002891 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
2892 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
2893 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
2894 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002895 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002896 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002897 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002898]
2899
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002900AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002901 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2902 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002903 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2904 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2905 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2906 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2907 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2908 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2909 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2910 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2911]
2912
Frank Barchardbcedc082020-08-17 18:00:51 -07002913WASM32_ASM_UKERNELS = [
2914 "src/f32-relu/wasm_shr_x1.S",
2915 "src/f32-relu/wasm_shr_x2.S",
2916 "src/f32-relu/wasm_shr_x4.S",
2917]
2918
Marat Dukhan08c4a432019-10-03 09:29:21 -07002919AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002920 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002921 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002922 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2923 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002924 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002925 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002926 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002927 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002928 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2929 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002930 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2931 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2932 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2933 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002934]
2935
2936AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002937 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002938 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002939 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002940 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002941 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002942 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002943 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002944 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2945 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
2946 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2947 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2948 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
2949 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
2950 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002951 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2952 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002953 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2954 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2955 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002956 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
2957 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002958 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2959 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2960 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2961 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002962 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002963 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2964 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002965 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2966 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2967 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2968 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002969 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002970 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002971 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002972 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002973 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2974 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2975 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2976 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2977 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
2978 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2979 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2980 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2981 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2982 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2983 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2984 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2985 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2986 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2987 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2988 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2989 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2990 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2991 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2992 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
2993 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
2994 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002995 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002996 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002997 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2998 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002999 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3000 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3001 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3002 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3003 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3004 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003005 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3006 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3007 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3008 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003009 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3010 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003011 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07003012 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07003013 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003014]
3015
Marat Dukhan1b354632020-03-23 12:50:22 -07003016INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003017 "src/xnnpack/argmaxpool.h",
3018 "src/xnnpack/avgpool.h",
3019 "src/xnnpack/clamp.h",
3020 "src/xnnpack/common.h",
3021 "src/xnnpack/conv.h",
3022 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003023 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003024 "src/xnnpack/gavgpool.h",
3025 "src/xnnpack/gemm.h",
3026 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003027 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003028 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003029 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003030 "src/xnnpack/lut.h",
3031 "src/xnnpack/math.h",
3032 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003033 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003034 "src/xnnpack/packx.h",
3035 "src/xnnpack/pad.h",
3036 "src/xnnpack/params.h",
3037 "src/xnnpack/pavgpool.h",
3038 "src/xnnpack/ppmm.h",
3039 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003040 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003041 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003042 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003043 "src/xnnpack/rmax.h",
3044 "src/xnnpack/scalar-utils.h",
3045 "src/xnnpack/spmm.h",
3046 "src/xnnpack/unpool.h",
3047 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003048 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003049 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003050 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003051 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003052 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003053 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003054 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003055]
3056
3057INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003058 "include/xnnpack.h",
3059 "src/xnnpack/allocator.h",
3060 "src/xnnpack/compute.h",
3061 "src/xnnpack/im2col.h",
3062 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003063 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003064 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003065 "src/xnnpack/operator.h",
3066 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003067 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003068 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003069 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003070 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003071]
3072
Marat Dukhan1b354632020-03-23 12:50:22 -07003073ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003074 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003075]
3076
Marat Dukhan1b354632020-03-23 12:50:22 -07003077MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003078 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003079 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003080]
3081
Marat Dukhan1b354632020-03-23 12:50:22 -07003082MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003083 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003084 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003085 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003086 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003087]
3088
3089OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003090 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003091 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003092]
3093
3094WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003095 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003096 "src/xnnpack/operator.h",
3097 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003098]
3099
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003100LOGGING_COPTS = select({
3101 # No logging in optimized mode
3102 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3103 # Full logging in debug mode
3104 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3105 # Error-only logging in default (fastbuild) mode
3106 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3107})
3108
Marat Dukhan3b59de22020-06-03 20:15:19 -07003109LOGGING_SRCS = select({
3110 # No logging in optimized mode
3111 ":optimized_build": [],
3112 "//conditions:default": [
3113 "src/operator-strings.c",
3114 "src/subgraph-strings.c",
3115 ],
3116})
3117
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003118LOGGING_HDRS = [
3119 "src/xnnpack/log.h",
3120]
3121
Marat Dukhan08c4a432019-10-03 09:29:21 -07003122xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003123 name = "tables",
3124 srcs = TABLE_SRCS,
3125 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003126 gcc_copts = xnnpack_gcc_std_copts(),
3127 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003128)
3129
3130xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003131 name = "scalar_ukernels",
3132 srcs = SCALAR_UKERNELS,
3133 hdrs = INTERNAL_HDRS,
3134 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003135 gcc_copts = xnnpack_gcc_std_copts(),
3136 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003137 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003138 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003139 "@FP16",
3140 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003141 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003142 ],
3143)
3144
3145xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003146 name = "scalar_ukernels_test_mode",
3147 srcs = SCALAR_UKERNELS,
3148 hdrs = INTERNAL_HDRS,
3149 aarch32_copts = ["-marm"],
3150 copts = [
3151 "-UNDEBUG",
3152 "-DXNN_TEST_MODE=1",
3153 ],
3154 gcc_copts = xnnpack_gcc_std_copts(),
3155 msvc_copts = xnnpack_msvc_std_copts(),
3156 deps = [
3157 ":tables",
3158 "@FP16",
3159 "@FXdiv",
3160 "@pthreadpool",
3161 ],
3162)
3163
3164xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003165 name = "wasm_ukernels",
3166 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003167 gcc_copts = xnnpack_gcc_std_copts(),
3168 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003169 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003170 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003171 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003172 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003173 "@FP16",
3174 "@FXdiv",
3175 "@pthreadpool",
3176 ],
3177)
3178
3179xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003180 name = "wasm_ukernels_test_mode",
3181 hdrs = INTERNAL_HDRS,
3182 copts = [
3183 "-UNDEBUG",
3184 "-DXNN_TEST_MODE=1",
3185 ],
3186 gcc_copts = xnnpack_gcc_std_copts(),
3187 msvc_copts = xnnpack_msvc_std_copts(),
3188 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003189 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003190 deps = [
3191 ":tables",
3192 "@FP16",
3193 "@FXdiv",
3194 "@pthreadpool",
3195 ],
3196)
3197
3198xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003199 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003200 hdrs = INTERNAL_HDRS,
3201 aarch32_copts = [
3202 "-marm",
3203 "-mfpu=neon",
3204 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003205 gcc_copts = xnnpack_gcc_std_copts(),
3206 gcc_x86_copts = ["-msse2"],
3207 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003208 optimized_copts = [
3209 "-O3",
3210 "-ffast-math",
3211 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003212 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003213 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003214 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003215 "@FP16",
3216 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003217 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003218 ],
3219)
3220
3221xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003222 name = "psimd_fastmath_ukernels_test_mode",
3223 hdrs = INTERNAL_HDRS,
3224 aarch32_copts = [
3225 "-marm",
3226 "-mfpu=neon",
3227 ],
3228 copts = [
3229 "-UNDEBUG",
3230 "-DXNN_TEST_MODE=1",
3231 ],
3232 gcc_copts = xnnpack_gcc_std_copts(),
3233 gcc_x86_copts = ["-msse2"],
3234 msvc_copts = xnnpack_msvc_std_copts(),
3235 optimized_copts = [
3236 "-O3",
3237 "-ffast-math",
3238 ],
3239 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
3240 deps = [
3241 ":tables",
3242 "@FP16",
3243 "@psimd",
3244 "@pthreadpool",
3245 ],
3246)
3247
3248xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003249 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003250 hdrs = INTERNAL_HDRS,
3251 aarch32_copts = [
3252 "-marm",
3253 "-mfpu=neon",
3254 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003255 gcc_copts = xnnpack_gcc_std_copts(),
3256 gcc_x86_copts = ["-msse2"],
3257 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003258 optimized_copts = [
3259 "-O3",
3260 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003261 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003262 deps = [
3263 ":tables",
3264 "@FP16",
3265 "@psimd",
3266 "@pthreadpool",
3267 ],
3268)
3269
3270xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003271 name = "psimd_accmath_ukernels_test_mode",
3272 hdrs = INTERNAL_HDRS,
3273 aarch32_copts = [
3274 "-marm",
3275 "-mfpu=neon",
3276 ],
3277 copts = [
3278 "-UNDEBUG",
3279 "-DXNN_TEST_MODE=1",
3280 ],
3281 gcc_copts = xnnpack_gcc_std_copts(),
3282 gcc_x86_copts = ["-msse2"],
3283 msvc_copts = xnnpack_msvc_std_copts(),
3284 optimized_copts = [
3285 "-O3",
3286 ],
3287 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3288 deps = [
3289 ":tables",
3290 "@FP16",
3291 "@psimd",
3292 "@pthreadpool",
3293 ],
3294)
3295
3296xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003297 name = "neon_ukernels",
3298 hdrs = INTERNAL_HDRS,
3299 aarch32_copts = [
3300 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003301 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003302 "-mfpu=neon",
3303 ],
3304 aarch32_srcs = NEON_UKERNELS,
3305 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003306 gcc_copts = xnnpack_gcc_std_copts(),
3307 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003308 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003309 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003310 "@FP16",
3311 "@pthreadpool",
3312 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003313)
3314
3315xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003316 name = "neon_ukernels_test_mode",
3317 hdrs = INTERNAL_HDRS,
3318 aarch32_copts = [
3319 "-marm",
3320 "-march=armv7-a",
3321 "-mfpu=neon",
3322 ],
3323 aarch32_srcs = NEON_UKERNELS,
3324 aarch64_srcs = NEON_UKERNELS,
3325 copts = [
3326 "-UNDEBUG",
3327 "-DXNN_TEST_MODE=1",
3328 ],
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(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003339 name = "neonfma_ukernels",
3340 hdrs = INTERNAL_HDRS,
3341 aarch32_copts = [
3342 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003343 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003344 "-mfpu=neon-vfpv4",
3345 ],
3346 aarch32_srcs = NEONFMA_UKERNELS,
3347 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003348 apple_aarch32_copts = [
3349 "-mcpu=swift",
3350 "-mtune=generic",
3351 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003352 gcc_copts = xnnpack_gcc_std_copts(),
3353 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003354 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003355 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003356 "@FP16",
3357 "@pthreadpool",
3358 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003359)
3360
3361xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003362 name = "neonfma_ukernels_test_mode",
3363 hdrs = INTERNAL_HDRS,
3364 aarch32_copts = [
3365 "-marm",
3366 "-march=armv7-a",
3367 "-mfpu=neon-vfpv4",
3368 ],
3369 aarch32_srcs = NEONFMA_UKERNELS,
3370 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003371 apple_aarch32_copts = [
3372 "-mcpu=swift",
3373 "-mtune=generic",
3374 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003375 copts = [
3376 "-UNDEBUG",
3377 "-DXNN_TEST_MODE=1",
3378 ],
3379 gcc_copts = xnnpack_gcc_std_copts(),
3380 msvc_copts = xnnpack_msvc_std_copts(),
3381 deps = [
3382 ":tables",
3383 "@FP16",
3384 "@pthreadpool",
3385 ],
3386)
3387
3388xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003389 name = "neonv8_ukernels",
3390 hdrs = INTERNAL_HDRS,
3391 aarch32_copts = [
3392 "-marm",
3393 "-march=armv8-a",
3394 "-mfpu=neon-fp-armv8",
3395 ],
3396 aarch32_srcs = NEONV8_UKERNELS,
3397 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003398 apple_aarch32_copts = [
3399 "-mcpu=cyclone",
3400 "-mtune=generic",
3401 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003402 gcc_copts = xnnpack_gcc_std_copts(),
3403 msvc_copts = xnnpack_msvc_std_copts(),
3404 deps = [
3405 ":tables",
3406 "@FP16",
3407 "@pthreadpool",
3408 ],
3409)
3410
3411xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003412 name = "neonv8_ukernels_test_mode",
3413 hdrs = INTERNAL_HDRS,
3414 aarch32_copts = [
3415 "-marm",
3416 "-march=armv8-a",
3417 "-mfpu=neon-fp-armv8",
3418 ],
3419 aarch32_srcs = NEONV8_UKERNELS,
3420 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003421 apple_aarch32_copts = [
3422 "-mcpu=cyclone",
3423 "-mtune=generic",
3424 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003425 copts = [
3426 "-UNDEBUG",
3427 "-DXNN_TEST_MODE=1",
3428 ],
3429 gcc_copts = xnnpack_gcc_std_copts(),
3430 msvc_copts = xnnpack_msvc_std_copts(),
3431 deps = [
3432 ":tables",
3433 "@FP16",
3434 "@pthreadpool",
3435 ],
3436)
3437
3438xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003439 name = "neonfp16arith_ukernels",
3440 hdrs = INTERNAL_HDRS,
3441 aarch64_copts = ["-march=armv8.2-a+fp16"],
3442 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003443 gcc_copts = xnnpack_gcc_std_copts(),
3444 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003445 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003446 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003447 "@FP16",
3448 "@pthreadpool",
3449 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003450)
3451
3452xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003453 name = "neonfp16arith_ukernels_test_mode",
3454 hdrs = INTERNAL_HDRS,
3455 aarch64_copts = ["-march=armv8.2-a+fp16"],
3456 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3457 copts = [
3458 "-UNDEBUG",
3459 "-DXNN_TEST_MODE=1",
3460 ],
3461 gcc_copts = xnnpack_gcc_std_copts(),
3462 msvc_copts = xnnpack_msvc_std_copts(),
3463 deps = [
3464 ":tables",
3465 "@FP16",
3466 "@pthreadpool",
3467 ],
3468)
3469
3470xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003471 name = "neondot_ukernels",
3472 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003473 aarch32_copts = [
3474 "-marm",
3475 "-march=armv8.2-a+dotprod",
3476 "-mfpu=neon-fp-armv8",
3477 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003478 aarch32_srcs = NEONDOT_UKERNELS,
3479 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3480 aarch64_srcs = NEONDOT_UKERNELS,
3481 gcc_copts = xnnpack_gcc_std_copts(),
3482 msvc_copts = xnnpack_msvc_std_copts(),
3483 deps = [
3484 ":tables",
3485 "@FP16",
3486 "@pthreadpool",
3487 ],
3488)
3489
3490xnnpack_cc_library(
3491 name = "neondot_ukernels_test_mode",
3492 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003493 aarch32_copts = [
3494 "-marm",
3495 "-march=armv8.2-a+dotprod",
3496 "-mfpu=neon-fp-armv8",
3497 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003498 aarch32_srcs = NEONDOT_UKERNELS,
3499 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3500 aarch64_srcs = NEONDOT_UKERNELS,
3501 copts = [
3502 "-UNDEBUG",
3503 "-DXNN_TEST_MODE=1",
3504 ],
3505 gcc_copts = xnnpack_gcc_std_copts(),
3506 msvc_copts = xnnpack_msvc_std_copts(),
3507 deps = [
3508 ":tables",
3509 "@FP16",
3510 "@pthreadpool",
3511 ],
3512)
3513
3514xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003515 name = "sse2_ukernels",
3516 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003517 gcc_copts = xnnpack_gcc_std_copts(),
3518 gcc_x86_copts = ["-msse2"],
3519 msvc_copts = xnnpack_msvc_std_copts(),
3520 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003521 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003522 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003523 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003524 "@FP16",
3525 "@pthreadpool",
3526 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003527)
3528
3529xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003530 name = "sse2_ukernels_test_mode",
3531 hdrs = INTERNAL_HDRS,
3532 copts = [
3533 "-UNDEBUG",
3534 "-DXNN_TEST_MODE=1",
3535 ],
3536 gcc_copts = xnnpack_gcc_std_copts(),
3537 gcc_x86_copts = ["-msse2"],
3538 msvc_copts = xnnpack_msvc_std_copts(),
3539 msvc_x86_32_copts = ["/arch:SSE2"],
3540 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3541 deps = [
3542 ":tables",
3543 "@FP16",
3544 "@pthreadpool",
3545 ],
3546)
3547
3548xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003549 name = "ssse3_ukernels",
3550 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003551 gcc_copts = xnnpack_gcc_std_copts(),
3552 gcc_x86_copts = ["-mssse3"],
3553 msvc_copts = xnnpack_msvc_std_copts(),
3554 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003555 x86_srcs = SSSE3_UKERNELS,
3556 deps = [
3557 ":tables",
3558 "@FP16",
3559 "@pthreadpool",
3560 ],
3561)
3562
3563xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003564 name = "ssse3_ukernels_test_mode",
3565 hdrs = INTERNAL_HDRS,
3566 copts = [
3567 "-UNDEBUG",
3568 "-DXNN_TEST_MODE=1",
3569 ],
3570 gcc_copts = xnnpack_gcc_std_copts(),
3571 gcc_x86_copts = ["-mssse3"],
3572 msvc_copts = xnnpack_msvc_std_copts(),
3573 msvc_x86_32_copts = ["/arch:SSE2"],
3574 x86_srcs = SSSE3_UKERNELS,
3575 deps = [
3576 ":tables",
3577 "@FP16",
3578 "@pthreadpool",
3579 ],
3580)
3581
3582xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003583 name = "sse41_ukernels",
3584 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003585 gcc_copts = xnnpack_gcc_std_copts(),
3586 gcc_x86_copts = ["-msse4.1"],
3587 msvc_copts = xnnpack_msvc_std_copts(),
3588 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003589 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003590 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003591 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003592 "@FP16",
3593 "@pthreadpool",
3594 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003595)
3596
3597xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003598 name = "sse41_ukernels_test_mode",
3599 hdrs = INTERNAL_HDRS,
3600 copts = [
3601 "-UNDEBUG",
3602 "-DXNN_TEST_MODE=1",
3603 ],
3604 gcc_copts = xnnpack_gcc_std_copts(),
3605 gcc_x86_copts = ["-msse4.1"],
3606 msvc_copts = xnnpack_msvc_std_copts(),
3607 msvc_x86_32_copts = ["/arch:SSE2"],
3608 x86_srcs = SSE41_UKERNELS,
3609 deps = [
3610 ":tables",
3611 "@FP16",
3612 "@pthreadpool",
3613 ],
3614)
3615
3616xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003617 name = "avx_ukernels",
3618 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003619 gcc_copts = xnnpack_gcc_std_copts(),
3620 gcc_x86_copts = ["-mavx"],
3621 msvc_copts = xnnpack_msvc_std_copts(),
3622 msvc_x86_32_copts = ["/arch:AVX"],
3623 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003624 x86_srcs = AVX_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 = "avx_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 = ["-mavx"],
3641 msvc_copts = xnnpack_msvc_std_copts(),
3642 msvc_x86_32_copts = ["/arch:AVX"],
3643 msvc_x86_64_copts = ["/arch:AVX"],
3644 x86_srcs = AVX_UKERNELS,
3645 deps = [
3646 ":tables",
3647 "@FP16",
3648 "@pthreadpool",
3649 ],
3650)
3651
3652xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003653 name = "xop_ukernels",
3654 hdrs = INTERNAL_HDRS,
3655 gcc_copts = xnnpack_gcc_std_copts(),
3656 gcc_x86_copts = ["-mxop"],
3657 msvc_copts = xnnpack_msvc_std_copts(),
3658 msvc_x86_32_copts = ["/arch:AVX"],
3659 msvc_x86_64_copts = ["/arch:AVX"],
3660 x86_srcs = XOP_UKERNELS,
3661 deps = [
3662 ":tables",
3663 "@FP16",
3664 "@pthreadpool",
3665 ],
3666)
3667
3668xnnpack_cc_library(
3669 name = "xop_ukernels_test_mode",
3670 hdrs = INTERNAL_HDRS,
3671 copts = [
3672 "-UNDEBUG",
3673 "-DXNN_TEST_MODE=1",
3674 ],
3675 gcc_copts = xnnpack_gcc_std_copts(),
3676 gcc_x86_copts = ["-mxop"],
3677 msvc_copts = xnnpack_msvc_std_copts(),
3678 msvc_x86_32_copts = ["/arch:AVX"],
3679 msvc_x86_64_copts = ["/arch:AVX"],
3680 x86_srcs = XOP_UKERNELS,
3681 deps = [
3682 ":tables",
3683 "@FP16",
3684 "@pthreadpool",
3685 ],
3686)
3687
3688xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003689 name = "fma3_ukernels",
3690 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003691 gcc_copts = xnnpack_gcc_std_copts(),
3692 gcc_x86_copts = ["-mfma"],
3693 msvc_copts = xnnpack_msvc_std_copts(),
3694 msvc_x86_32_copts = ["/arch:AVX"],
3695 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003696 x86_srcs = FMA3_UKERNELS,
3697 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003698 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003699 "@FP16",
3700 "@pthreadpool",
3701 ],
3702)
3703
3704xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003705 name = "fma3_ukernels_test_mode",
3706 hdrs = INTERNAL_HDRS,
3707 copts = [
3708 "-UNDEBUG",
3709 "-DXNN_TEST_MODE=1",
3710 ],
3711 gcc_copts = xnnpack_gcc_std_copts(),
3712 gcc_x86_copts = ["-mfma"],
3713 msvc_copts = xnnpack_msvc_std_copts(),
3714 msvc_x86_32_copts = ["/arch:AVX"],
3715 msvc_x86_64_copts = ["/arch:AVX"],
3716 x86_srcs = FMA3_UKERNELS,
3717 deps = [
3718 ":tables",
3719 "@FP16",
3720 "@pthreadpool",
3721 ],
3722)
3723
3724xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003725 name = "avx2_ukernels",
3726 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003727 gcc_copts = xnnpack_gcc_std_copts(),
3728 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003729 "-mfma",
3730 "-mavx2",
3731 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003732 msvc_copts = xnnpack_msvc_std_copts(),
3733 msvc_x86_32_copts = ["/arch:AVX2"],
3734 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003735 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003736 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003737 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003738 "@FP16",
3739 "@pthreadpool",
3740 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003741)
3742
3743xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003744 name = "avx2_ukernels_test_mode",
3745 hdrs = INTERNAL_HDRS,
3746 copts = [
3747 "-UNDEBUG",
3748 "-DXNN_TEST_MODE=1",
3749 ],
3750 gcc_copts = xnnpack_gcc_std_copts(),
3751 gcc_x86_copts = [
3752 "-mfma",
3753 "-mavx2",
3754 ],
3755 msvc_copts = xnnpack_msvc_std_copts(),
3756 msvc_x86_32_copts = ["/arch:AVX2"],
3757 msvc_x86_64_copts = ["/arch:AVX2"],
3758 x86_srcs = AVX2_UKERNELS,
3759 deps = [
3760 ":tables",
3761 "@FP16",
3762 "@pthreadpool",
3763 ],
3764)
3765
3766xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003767 name = "avx512f_ukernels",
3768 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003769 gcc_copts = xnnpack_gcc_std_copts(),
3770 gcc_x86_copts = ["-mavx512f"],
3771 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3772 msvc_copts = xnnpack_msvc_std_copts(),
3773 msvc_x86_32_copts = ["/arch:AVX512"],
3774 msvc_x86_64_copts = ["/arch:AVX512"],
3775 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003776 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003777 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003778 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003779 "@FP16",
3780 "@pthreadpool",
3781 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003782)
3783
3784xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003785 name = "avx512f_ukernels_test_mode",
3786 hdrs = INTERNAL_HDRS,
3787 copts = [
3788 "-UNDEBUG",
3789 "-DXNN_TEST_MODE=1",
3790 ],
3791 gcc_copts = xnnpack_gcc_std_copts(),
3792 gcc_x86_copts = ["-mavx512f"],
3793 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3794 msvc_copts = xnnpack_msvc_std_copts(),
3795 msvc_x86_32_copts = ["/arch:AVX512"],
3796 msvc_x86_64_copts = ["/arch:AVX512"],
3797 msys_copts = ["-fno-asynchronous-unwind-tables"],
3798 x86_srcs = AVX512F_UKERNELS,
3799 deps = [
3800 ":tables",
3801 "@FP16",
3802 "@pthreadpool",
3803 ],
3804)
3805
3806xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003807 name = "avx512skx_ukernels",
3808 hdrs = INTERNAL_HDRS,
3809 gcc_copts = xnnpack_gcc_std_copts(),
3810 gcc_x86_copts = [
3811 "-mavx512f",
3812 "-mavx512cd",
3813 "-mavx512bw",
3814 "-mavx512dq",
3815 "-mavx512vl",
3816 ],
3817 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3818 msvc_copts = xnnpack_msvc_std_copts(),
3819 msvc_x86_32_copts = ["/arch:AVX512"],
3820 msvc_x86_64_copts = ["/arch:AVX512"],
3821 msys_copts = ["-fno-asynchronous-unwind-tables"],
3822 x86_srcs = AVX512SKX_UKERNELS,
3823 deps = [
3824 ":tables",
3825 "@FP16",
3826 "@pthreadpool",
3827 ],
3828)
3829
3830xnnpack_cc_library(
3831 name = "avx512skx_ukernels_test_mode",
3832 hdrs = INTERNAL_HDRS,
3833 copts = [
3834 "-UNDEBUG",
3835 "-DXNN_TEST_MODE=1",
3836 ],
3837 gcc_copts = xnnpack_gcc_std_copts(),
3838 gcc_x86_copts = [
3839 "-mavx512f",
3840 "-mavx512cd",
3841 "-mavx512bw",
3842 "-mavx512dq",
3843 "-mavx512vl",
3844 ],
3845 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3846 msvc_copts = xnnpack_msvc_std_copts(),
3847 msvc_x86_32_copts = ["/arch:AVX512"],
3848 msvc_x86_64_copts = ["/arch:AVX512"],
3849 msys_copts = ["-fno-asynchronous-unwind-tables"],
3850 x86_srcs = AVX512SKX_UKERNELS,
3851 deps = [
3852 ":tables",
3853 "@FP16",
3854 "@pthreadpool",
3855 ],
3856)
3857
3858xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003859 name = "asm_ukernels",
3860 hdrs = ["src/xnnpack/assembly.h"],
3861 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07003862 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003863 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003864 wasm_srcs = WASM32_ASM_UKERNELS,
3865 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003866)
3867
Marat Dukhan3b59de22020-06-03 20:15:19 -07003868xnnpack_cc_library(
3869 name = "logging_utils",
3870 srcs = LOGGING_SRCS,
3871 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3872 copts = LOGGING_COPTS + [
3873 "-Isrc",
3874 "-Iinclude",
3875 ] + select({
3876 ":debug_build": [],
3877 "//conditions:default": xnnpack_min_size_copts(),
3878 }),
3879 gcc_copts = xnnpack_gcc_std_copts(),
3880 msvc_copts = xnnpack_msvc_std_copts(),
3881 visibility = xnnpack_visibility(),
3882 deps = [
3883 "@FP16",
3884 "@clog",
3885 "@pthreadpool",
3886 ],
3887)
3888
Marat Dukhan08c4a432019-10-03 09:29:21 -07003889xnnpack_aggregate_library(
3890 name = "ukernels",
3891 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003892 ":neon_ukernels",
3893 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003894 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003895 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003896 ":asm_ukernels",
3897 ],
3898 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003899 ":neon_ukernels",
3900 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003901 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003902 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003903 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003904 ":asm_ukernels",
3905 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003906 generic_deps = [
3907 ":scalar_ukernels",
3908 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003909 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003910 ":psimd_fastmath_ukernels",
3911 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003912 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003913 wasm_deps = [
3914 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003915 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003916 ],
3917 wasmsimd_deps = [
3918 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003919 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003920 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003921 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003922 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003923 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003924 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003925 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003926 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003927 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003928 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003929 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003930 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003931 ],
3932)
3933
Marat Dukhan33fcf782020-05-24 14:27:15 -07003934xnnpack_aggregate_library(
3935 name = "ukernels_test_mode",
3936 aarch32_deps = [
3937 ":neon_ukernels_test_mode",
3938 ":neonfma_ukernels_test_mode",
3939 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003940 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003941 ":asm_ukernels",
3942 ],
3943 aarch64_deps = [
3944 ":neon_ukernels_test_mode",
3945 ":neonfma_ukernels_test_mode",
3946 ":neonv8_ukernels_test_mode",
3947 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003948 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003949 ":asm_ukernels",
3950 ],
3951 generic_deps = [
3952 ":scalar_ukernels_test_mode",
3953 ],
3954 psimd_deps = [
3955 ":psimd_fastmath_ukernels_test_mode",
3956 ":psimd_accmath_ukernels_test_mode",
3957 ],
3958 wasm_deps = [
3959 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003960 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003961 ],
3962 wasmsimd_deps = [
3963 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003964 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003965 ],
3966 x86_deps = [
3967 ":sse2_ukernels_test_mode",
3968 ":ssse3_ukernels_test_mode",
3969 ":sse41_ukernels_test_mode",
3970 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003971 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003972 ":fma3_ukernels_test_mode",
3973 ":avx2_ukernels_test_mode",
3974 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003975 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003976 ],
3977)
3978
Marat Dukhan86df3842020-06-17 14:03:31 -07003979# This target must be compatible with Android NDK r17 and compile with Clang 6.
3980# It is similar to :ukernels target, but excludes microkernels which require
3981# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003982xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003983 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003984 aarch32_deps = [
3985 ":neon_ukernels",
3986 ":neonfma_ukernels",
3987 ":neonv8_ukernels",
3988 ":asm_ukernels",
3989 ],
3990 aarch64_deps = [
3991 ":neon_ukernels",
3992 ":neonfma_ukernels",
3993 ":neonv8_ukernels",
3994 ":asm_ukernels",
3995 ],
3996 generic_deps = [
3997 ":scalar_ukernels",
3998 ],
3999 psimd_deps = [
4000 ":psimd_fastmath_ukernels",
4001 ":psimd_accmath_ukernels",
4002 ],
4003 wasm_deps = [
4004 ":wasm_ukernels",
4005 ],
4006 wasmsimd_deps = [
4007 ":wasm_ukernels",
4008 ],
4009 x86_deps = [
4010 ":sse2_ukernels",
4011 ":ssse3_ukernels",
4012 ":sse41_ukernels",
4013 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004014 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004015 ":fma3_ukernels",
4016 ":avx2_ukernels",
4017 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004018 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004019 ],
4020)
4021
Marat Dukhan08c4a432019-10-03 09:29:21 -07004022xnnpack_cc_library(
4023 name = "im2col",
4024 srcs = ["src/im2col.c"],
4025 hdrs = [
4026 "src/xnnpack/common.h",
4027 "src/xnnpack/im2col.h",
4028 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004029 gcc_copts = xnnpack_gcc_std_copts(),
4030 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004031)
4032
4033xnnpack_cc_library(
4034 name = "indirection",
4035 srcs = ["src/indirection.c"],
4036 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004037 gcc_copts = xnnpack_gcc_std_copts(),
4038 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004039 deps = [
4040 "@FP16",
4041 "@FXdiv",
4042 "@pthreadpool",
4043 ],
4044)
4045
4046xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004047 name = "indirection_test_mode",
4048 srcs = ["src/indirection.c"],
4049 hdrs = INTERNAL_HDRS,
4050 copts = [
4051 "-UNDEBUG",
4052 "-DXNN_TEST_MODE=1",
4053 ],
4054 gcc_copts = xnnpack_gcc_std_copts(),
4055 msvc_copts = xnnpack_msvc_std_copts(),
4056 deps = [
4057 "@FP16",
4058 "@FXdiv",
4059 "@pthreadpool",
4060 ],
4061)
4062
4063xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004064 name = "packing",
4065 srcs = ["src/packing.c"],
4066 hdrs = INTERNAL_HDRS,
4067 gcc_copts = xnnpack_gcc_std_copts(),
4068 msvc_copts = xnnpack_msvc_std_copts(),
4069 deps = [
4070 "@FP16",
4071 "@FXdiv",
4072 "@pthreadpool",
4073 ],
4074)
4075
4076xnnpack_cc_library(
4077 name = "packing_test_mode",
4078 srcs = ["src/packing.c"],
4079 hdrs = INTERNAL_HDRS,
4080 copts = [
4081 "-UNDEBUG",
4082 "-DXNN_TEST_MODE=1",
4083 ],
4084 gcc_copts = xnnpack_gcc_std_copts(),
4085 msvc_copts = xnnpack_msvc_std_copts(),
4086 deps = [
4087 "@FP16",
4088 "@FXdiv",
4089 "@pthreadpool",
4090 ],
4091)
4092
4093xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004094 name = "operator_run",
4095 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004096 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004097 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004098 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4099 "//conditions:default": [],
4100 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004101 gcc_copts = xnnpack_gcc_std_copts(),
4102 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004103 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004104 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004105 "@FP16",
4106 "@FXdiv",
4107 "@clog",
4108 "@pthreadpool",
4109 ],
4110)
4111
Chao Mei6ddfc602020-05-13 22:29:36 -07004112xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004113 name = "operator_run_test_mode",
4114 srcs = ["src/operator-run.c"],
4115 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4116 copts = LOGGING_COPTS + [
4117 "-UNDEBUG",
4118 "-DXNN_TEST_MODE=1",
4119 ] + select({
4120 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4121 "//conditions:default": [],
4122 }),
4123 gcc_copts = xnnpack_gcc_std_copts(),
4124 msvc_copts = xnnpack_msvc_std_copts(),
4125 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004126 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004127 "@FP16",
4128 "@FXdiv",
4129 "@clog",
4130 "@pthreadpool",
4131 ],
4132)
4133
4134xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004135 name = "memory_planner",
4136 srcs = ["src/memory-planner.c"],
4137 hdrs = INTERNAL_HDRS,
4138 defines = select({
4139 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4140 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4141 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4142 }),
4143 gcc_copts = xnnpack_gcc_std_copts(),
4144 msvc_copts = xnnpack_msvc_std_copts(),
4145 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004146 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004147 "@pthreadpool",
4148 ],
4149)
4150
Marat Dukhan33fcf782020-05-24 14:27:15 -07004151xnnpack_cc_library(
4152 name = "memory_planner_test_mode",
4153 srcs = ["src/memory-planner.c"],
4154 hdrs = INTERNAL_HDRS,
4155 copts = [
4156 "-UNDEBUG",
4157 "-DXNN_TEST_MODE=1",
4158 ],
4159 defines = select({
4160 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4161 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4162 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4163 }),
4164 gcc_copts = xnnpack_gcc_std_copts(),
4165 msvc_copts = xnnpack_msvc_std_copts(),
4166 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004167 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004168 "@pthreadpool",
4169 ],
4170)
4171
Marat Dukhan08c4a432019-10-03 09:29:21 -07004172cc_library(
4173 name = "enable_assembly",
4174 defines = select({
4175 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4176 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004177 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004178 }),
4179)
4180
Marat Dukhan9de90e02020-06-18 16:04:12 -07004181cc_library(
4182 name = "enable_sparse",
4183 defines = select({
4184 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4185 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07004186 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004187 }),
4188)
4189
Marat Dukhancf056b22019-10-07 10:26:29 -07004190xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004191 name = "operators",
4192 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004193 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004194 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004195 ],
4196 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004197 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004198 "-Isrc",
4199 "-Iinclude",
4200 ] + select({
4201 ":debug_build": [],
4202 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004203 }) + select({
4204 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4205 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004206 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004207 gcc_copts = xnnpack_gcc_std_copts(),
4208 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004209 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004210 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004211 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004212 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004213 "@FP16",
4214 "@FXdiv",
4215 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004216 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004217 ],
4218)
4219
Marat Dukhan10a38082020-04-17 03:58:35 -07004220xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004221 name = "operators_test_mode",
4222 srcs = OPERATOR_SRCS + [
4223 "src/memory.c",
4224 "src/operator-delete.c",
4225 ],
4226 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4227 copts = LOGGING_COPTS + [
4228 "-Isrc",
4229 "-Iinclude",
4230 "-UNDEBUG",
4231 "-DXNN_TEST_MODE=1",
4232 ] + select({
4233 ":debug_build": [],
4234 "//conditions:default": xnnpack_min_size_copts(),
4235 }) + select({
4236 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4237 "//conditions:default": [],
4238 }),
4239 gcc_copts = xnnpack_gcc_std_copts(),
4240 msvc_copts = xnnpack_msvc_std_copts(),
4241 deps = [
4242 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004243 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004244 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004245 "@FP16",
4246 "@FXdiv",
4247 "@clog",
4248 "@pthreadpool",
4249 ],
4250)
4251
4252xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004253 name = "XNNPACK",
4254 srcs = [
4255 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004256 "src/runtime.c",
4257 "src/subgraph.c",
4258 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004259 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004260 hdrs = ["include/xnnpack.h"],
4261 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004262 "-Isrc",
4263 "-Iinclude",
4264 ] + select({
4265 ":debug_build": [],
4266 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004267 }) + select({
4268 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4269 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004270 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004271 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004272 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004273 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004274 visibility = xnnpack_visibility(),
4275 deps = [
4276 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004277 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004278 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004279 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004280 ":operator_run",
4281 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004282 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004283 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004284 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004285 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004286 ] + select({
4287 ":emscripten": [],
4288 "//conditions:default": ["@cpuinfo"],
4289 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004290)
4291
Marat Dukhan10a38082020-04-17 03:58:35 -07004292xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004293 name = "XNNPACK_test_mode",
4294 srcs = [
4295 "src/init.c",
4296 "src/runtime.c",
4297 "src/subgraph.c",
4298 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004299 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004300 hdrs = ["include/xnnpack.h"],
4301 copts = LOGGING_COPTS + [
4302 "-Isrc",
4303 "-Iinclude",
4304 "-UNDEBUG",
4305 "-DXNN_TEST_MODE=1",
4306 ] + select({
4307 ":debug_build": [],
4308 "//conditions:default": xnnpack_min_size_copts(),
4309 }) + select({
4310 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4311 "//conditions:default": [],
4312 }),
4313 gcc_copts = xnnpack_gcc_std_copts(),
4314 includes = ["include"],
4315 msvc_copts = xnnpack_msvc_std_copts(),
4316 visibility = xnnpack_visibility(),
4317 deps = [
4318 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004319 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004320 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004321 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004322 ":operator_run_test_mode",
4323 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004324 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004325 "@clog",
4326 "@FP16",
4327 "@pthreadpool",
4328 ] + select({
4329 ":emscripten": [],
4330 "//conditions:default": ["@cpuinfo"],
4331 }),
4332)
4333
4334xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004335 name = "xnnpack_f32",
4336 srcs = [
4337 "src/init.c",
4338 "src/runtime.c",
4339 "src/subgraph.c",
4340 "src/tensor.c",
4341 ] + SUBGRAPH_SRCS,
4342 hdrs = ["include/xnnpack.h"],
4343 copts = LOGGING_COPTS + [
4344 "-Isrc",
4345 "-Iinclude",
4346 ] + select({
4347 ":debug_build": [],
4348 "//conditions:default": xnnpack_min_size_copts(),
4349 }) + select({
4350 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4351 "//conditions:default": [],
4352 }),
4353 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004354 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004355 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004356 "XNN_NO_U8_OPERATORS",
4357 "XNN_NO_X8_OPERATORS",
4358 "XNN_NO_F16_OPERATORS",
4359 "XNN_NO_X16_OPERATORS",
4360 ],
4361 gcc_copts = xnnpack_gcc_std_copts(),
4362 includes = ["include"],
4363 msvc_copts = xnnpack_msvc_std_copts(),
4364 visibility = xnnpack_visibility(),
4365 deps = [
4366 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004367 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004368 ":logging_utils",
4369 ":memory_planner",
4370 ":operator_run",
4371 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004372 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004373 "@clog",
4374 "@FP16",
4375 "@pthreadpool",
4376 ] + select({
4377 ":emscripten": [],
4378 "//conditions:default": ["@cpuinfo"],
4379 }),
4380)
4381
4382xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004383 name = "xnnpack_operators_nhwc_f32",
4384 srcs = [
4385 "src/init.c",
4386 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004387 hdrs = ["include/xnnpack.h"],
4388 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004389 "-Isrc",
4390 "-Iinclude",
4391 ] + select({
4392 ":debug_build": [],
4393 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004394 }) + select({
4395 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4396 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004397 }),
4398 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004399 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004400 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004401 "XNN_NO_U8_OPERATORS",
4402 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004403 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004404 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004405 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004406 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004407 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004408 visibility = xnnpack_visibility(),
4409 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004410 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004411 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004412 ":operator_run",
4413 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004414 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004415 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004416 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004417 ] + select({
4418 ":emscripten": [],
4419 "//conditions:default": ["@cpuinfo"],
4420 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004421)
4422
Marat Dukhancf056b22019-10-07 10:26:29 -07004423xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004424 name = "bench_utils",
4425 srcs = ["bench/utils.cc"],
4426 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004427 deps = [
4428 "@com_google_benchmark//:benchmark",
4429 "@cpuinfo",
4430 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004431)
4432
Frank Barchard7e955972019-10-11 10:34:25 -07004433######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004434
4435xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004436 name = "qs8_gemm_bench",
4437 srcs = [
4438 "bench/gemm.h",
4439 "bench/qs8-gemm.cc",
4440 "src/xnnpack/AlignedAllocator.h",
4441 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004442 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4443 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004444)
4445
4446xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004447 name = "qs8_requantization_bench",
4448 srcs = [
4449 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004450 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004451 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004452 ] + MICROKERNEL_BENCHMARK_HDRS,
4453 deps = MICROKERNEL_BENCHMARK_DEPS,
4454)
4455
4456xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004457 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004458 srcs = [
4459 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004460 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004461 "src/xnnpack/AlignedAllocator.h",
4462 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004463 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004464 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004465)
4466
4467xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004468 name = "qu8_requantization_bench",
4469 srcs = [
4470 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004471 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004472 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004473 ] + MICROKERNEL_BENCHMARK_HDRS,
4474 deps = MICROKERNEL_BENCHMARK_DEPS,
4475)
4476
4477xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004478 name = "f16_igemm_bench",
4479 srcs = [
4480 "bench/f16-igemm.cc",
4481 "bench/conv.h",
4482 "bench/google/conv.h",
4483 "src/xnnpack/AlignedAllocator.h",
4484 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004485 deps = MICROKERNEL_BENCHMARK_DEPS + [
4486 ":indirection",
4487 ":packing",
4488 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004489)
4490
4491xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004492 name = "f16_gemm_bench",
4493 srcs = [
4494 "bench/f16-gemm.cc",
4495 "bench/gemm.h",
4496 "src/xnnpack/AlignedAllocator.h",
4497 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004498 deps = MICROKERNEL_BENCHMARK_DEPS + [
4499 ":packing",
4500 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004501)
4502
4503xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004504 name = "f16_spmm_bench",
4505 srcs = [
4506 "bench/f16-spmm.cc",
4507 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004508 "src/xnnpack/AlignedAllocator.h",
4509 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004510 deps = MICROKERNEL_BENCHMARK_DEPS,
4511)
4512
4513xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004514 name = "f32_igemm_bench",
4515 srcs = [
4516 "bench/f32-igemm.cc",
4517 "bench/conv.h",
4518 "src/xnnpack/AlignedAllocator.h",
4519 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004520 deps = MICROKERNEL_BENCHMARK_DEPS + [
4521 ":indirection",
4522 ":packing",
4523 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004524)
4525
4526xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004527 name = "f16_relu_bench",
4528 srcs = [
4529 "bench/f16-relu.cc",
4530 "src/xnnpack/AlignedAllocator.h",
4531 ] + MICROKERNEL_BENCHMARK_HDRS,
4532 deps = MICROKERNEL_BENCHMARK_DEPS,
4533)
4534
4535xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004536 name = "f32_conv_hwc_bench",
4537 srcs = [
4538 "bench/f32-conv-hwc.cc",
4539 "bench/dconv.h",
4540 "src/xnnpack/AlignedAllocator.h",
4541 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004542 deps = MICROKERNEL_BENCHMARK_DEPS + [
4543 ":packing",
4544 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004545)
4546
4547xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004548 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004549 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004550 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004551 "bench/dconv.h",
4552 "src/xnnpack/AlignedAllocator.h",
4553 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004554 deps = MICROKERNEL_BENCHMARK_DEPS + [
4555 ":packing",
4556 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004557)
4558
4559xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004560 name = "f16_dwconv_bench",
4561 srcs = [
4562 "bench/f16-dwconv.cc",
4563 "bench/dwconv.h",
4564 "bench/google/dwconv.h",
4565 "src/xnnpack/AlignedAllocator.h",
4566 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004567 deps = MICROKERNEL_BENCHMARK_DEPS + [
4568 ":indirection",
4569 ":packing",
4570 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004571)
4572
4573xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004574 name = "f32_dwconv_bench",
4575 srcs = [
4576 "bench/f32-dwconv.cc",
4577 "bench/dwconv.h",
4578 "src/xnnpack/AlignedAllocator.h",
4579 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004580 deps = MICROKERNEL_BENCHMARK_DEPS + [
4581 ":indirection",
4582 ":packing",
4583 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004584)
4585
4586xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07004587 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004588 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07004589 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004590 "bench/dwconv.h",
4591 "src/xnnpack/AlignedAllocator.h",
4592 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004593 deps = MICROKERNEL_BENCHMARK_DEPS + [
4594 ":indirection",
4595 ":packing",
4596 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004597)
4598
4599xnnpack_benchmark(
4600 name = "f32_gemm_bench",
4601 srcs = [
4602 "bench/f32-gemm.cc",
4603 "bench/gemm.h",
4604 "src/xnnpack/AlignedAllocator.h",
4605 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004606 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004607 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004608)
4609
4610xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004611 name = "f32_hswish_bench",
4612 srcs = [
4613 "bench/f32-hswish.cc",
4614 "src/xnnpack/AlignedAllocator.h",
4615 ] + MICROKERNEL_BENCHMARK_HDRS,
4616 deps = MICROKERNEL_BENCHMARK_DEPS,
4617)
4618
4619xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004620 name = "f32_raddexpminusmax_bench",
4621 srcs = [
4622 "bench/f32-raddexpminusmax.cc",
4623 "src/xnnpack/AlignedAllocator.h",
4624 ] + MICROKERNEL_BENCHMARK_HDRS,
4625 deps = MICROKERNEL_BENCHMARK_DEPS,
4626)
4627
4628xnnpack_benchmark(
4629 name = "f32_raddextexp_bench",
4630 srcs = [
4631 "bench/f32-raddextexp.cc",
4632 "src/xnnpack/AlignedAllocator.h",
4633 ] + MICROKERNEL_BENCHMARK_HDRS,
4634 deps = MICROKERNEL_BENCHMARK_DEPS,
4635)
4636
4637xnnpack_benchmark(
4638 name = "f32_raddstoreexpminusmax_bench",
4639 srcs = [
4640 "bench/f32-raddstoreexpminusmax.cc",
4641 "src/xnnpack/AlignedAllocator.h",
4642 ] + MICROKERNEL_BENCHMARK_HDRS,
4643 deps = MICROKERNEL_BENCHMARK_DEPS,
4644)
4645
4646xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004647 name = "f32_relu_bench",
4648 srcs = [
4649 "bench/f32-relu.cc",
4650 "src/xnnpack/AlignedAllocator.h",
4651 ] + MICROKERNEL_BENCHMARK_HDRS,
4652 deps = MICROKERNEL_BENCHMARK_DEPS,
4653)
4654
4655xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004656 name = "f32_rmax_bench",
4657 srcs = [
4658 "bench/f32-rmax.cc",
4659 "src/xnnpack/AlignedAllocator.h",
4660 ] + MICROKERNEL_BENCHMARK_HDRS,
4661 deps = MICROKERNEL_BENCHMARK_DEPS,
4662)
4663
4664xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004665 name = "f32_sigmoid_bench",
4666 srcs = [
4667 "bench/f32-sigmoid.cc",
4668 "src/xnnpack/AlignedAllocator.h",
4669 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004670 deps = MICROKERNEL_BENCHMARK_DEPS,
4671)
4672
4673xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004674 name = "f32_spmm_bench",
4675 srcs = [
4676 "bench/f32-spmm.cc",
4677 "bench/gemm.h",
4678 "src/xnnpack/AlignedAllocator.h",
4679 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004680 deps = MICROKERNEL_BENCHMARK_DEPS,
4681)
4682
4683xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004684 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004685 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004686 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004687 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004688 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004689 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004690)
4691
4692xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004693 name = "f32_vscaleexpminusmax_bench",
4694 srcs = [
4695 "bench/f32-vscaleexpminusmax.cc",
4696 "src/xnnpack/AlignedAllocator.h",
4697 ] + MICROKERNEL_BENCHMARK_HDRS,
4698 deps = MICROKERNEL_BENCHMARK_DEPS,
4699)
4700
4701xnnpack_benchmark(
4702 name = "f32_vscaleextexp_bench",
4703 srcs = [
4704 "bench/f32-vscaleextexp.cc",
4705 "src/xnnpack/AlignedAllocator.h",
4706 ] + MICROKERNEL_BENCHMARK_HDRS,
4707 deps = MICROKERNEL_BENCHMARK_DEPS,
4708)
4709
4710xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004711 name = "f32_vsqrt_bench",
4712 srcs = [
4713 "bench/f32-vsqrt.cc",
4714 "src/xnnpack/AlignedAllocator.h",
4715 ] + MICROKERNEL_BENCHMARK_HDRS,
4716 deps = MICROKERNEL_BENCHMARK_DEPS,
4717)
4718
4719xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004720 name = "f32_im2col_gemm_bench",
4721 srcs = [
4722 "bench/f32-im2col-gemm.cc",
4723 "bench/conv.h",
4724 "src/xnnpack/AlignedAllocator.h",
4725 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004726 deps = MICROKERNEL_BENCHMARK_DEPS + [
4727 ":im2col",
4728 ":packing",
4729 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004730)
4731
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004732xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004733 name = "rounding_bench",
4734 srcs = [
4735 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004736 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004737 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004738 ] + MICROKERNEL_BENCHMARK_HDRS,
4739 deps = MICROKERNEL_BENCHMARK_DEPS,
4740)
4741
Marat Dukhan08c4a432019-10-03 09:29:21 -07004742########################### Benchmarks for operators ###########################
4743
4744xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004745 name = "average_pooling_bench",
4746 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004747 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004748 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004749 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004750)
4751
4752xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004753 name = "bankers_rounding_bench",
4754 srcs = ["bench/bankers-rounding.cc"],
4755 copts = xnnpack_optional_tflite_copts(),
4756 tags = ["nowin32"],
4757 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4758)
4759
4760xnnpack_benchmark(
4761 name = "ceiling_bench",
4762 srcs = ["bench/ceiling.cc"],
4763 copts = xnnpack_optional_tflite_copts(),
4764 tags = ["nowin32"],
4765 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4766)
4767
4768xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004769 name = "channel_shuffle_bench",
4770 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004771 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004772)
4773
4774xnnpack_benchmark(
4775 name = "convolution_bench",
4776 srcs = ["bench/convolution.cc"],
4777 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004778 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004779 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004780)
4781
4782xnnpack_benchmark(
4783 name = "deconvolution_bench",
4784 srcs = ["bench/deconvolution.cc"],
4785 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004786 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004787 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004788)
4789
4790xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004791 name = "floor_bench",
4792 srcs = ["bench/floor.cc"],
4793 copts = xnnpack_optional_tflite_copts(),
4794 tags = ["nowin32"],
4795 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4796)
4797
4798xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004799 name = "global_average_pooling_bench",
4800 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004801 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004802)
4803
4804xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004805 name = "hardswish_bench",
4806 srcs = ["bench/hardswish.cc"],
4807 copts = xnnpack_optional_tflite_copts(),
4808 tags = ["nowin32"],
4809 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4810)
4811
4812xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004813 name = "max_pooling_bench",
4814 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004815 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004816)
4817
4818xnnpack_benchmark(
4819 name = "sigmoid_bench",
4820 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004821 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004822 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004823 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004824)
4825
4826xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004827 name = "prelu_bench",
4828 srcs = ["bench/prelu.cc"],
4829 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004830 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004831 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004832)
4833
4834xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004835 name = "softmax_bench",
4836 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004837 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004838 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004839 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004840)
4841
Marat Dukhan87727142020-06-24 15:24:10 -07004842xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004843 name = "square_root_bench",
4844 srcs = ["bench/square-root.cc"],
4845 copts = xnnpack_optional_tflite_copts(),
4846 tags = ["nowin32"],
4847 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4848)
4849
4850xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004851 name = "truncation_bench",
4852 srcs = ["bench/truncation.cc"],
4853 deps = OPERATOR_BENCHMARK_DEPS,
4854)
4855
Marat Dukhanc068bb62019-10-04 13:24:39 -07004856############################# End-to-end benchmarks ############################
4857
4858cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004859 name = "fp32_mobilenet_v1",
4860 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004861 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004862 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004863 linkstatic = True,
4864 deps = [
4865 ":XNNPACK",
4866 "@pthreadpool",
4867 ],
4868)
4869
4870cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004871 name = "fp16_mobilenet_v1",
4872 srcs = ["models/fp16-mobilenet-v1.cc"],
4873 hdrs = ["models/models.h"],
4874 copts = xnnpack_std_cxxopts(),
4875 linkstatic = True,
4876 deps = [
4877 ":XNNPACK",
4878 "@FP16",
4879 "@pthreadpool",
4880 ],
4881)
4882
4883cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004884 name = "qs8_mobilenet_v1",
4885 srcs = ["models/qs8-mobilenet-v1.cc"],
4886 hdrs = ["models/models.h"],
4887 copts = xnnpack_std_cxxopts(),
4888 linkstatic = True,
4889 deps = [
4890 ":XNNPACK",
4891 "@pthreadpool",
4892 ],
4893)
4894
4895cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004896 name = "qs8_mobilenet_v2",
4897 srcs = ["models/qs8-mobilenet-v2.cc"],
4898 hdrs = ["models/models.h"],
4899 copts = xnnpack_std_cxxopts(),
4900 linkstatic = True,
4901 deps = [
4902 ":XNNPACK",
4903 "@pthreadpool",
4904 ],
4905)
4906
4907cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004908 name = "fp32_mobilenet_v2",
4909 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004910 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004911 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004912 linkstatic = True,
4913 deps = [
4914 ":XNNPACK",
4915 "@pthreadpool",
4916 ],
4917)
4918
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004919cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004920 name = "fp16_mobilenet_v2",
4921 srcs = ["models/fp16-mobilenet-v2.cc"],
4922 hdrs = ["models/models.h"],
4923 copts = xnnpack_std_cxxopts(),
4924 linkstatic = True,
4925 deps = [
4926 ":XNNPACK",
4927 "@FP16",
4928 "@pthreadpool",
4929 ],
4930)
4931
4932cc_library(
4933 name = "fp32_mobilenet_v3_large",
4934 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004935 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004936 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004937 linkstatic = True,
4938 deps = [
4939 ":XNNPACK",
4940 "@pthreadpool",
4941 ],
4942)
4943
4944cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004945 name = "fp16_mobilenet_v3_large",
4946 srcs = ["models/fp16-mobilenet-v3-large.cc"],
4947 hdrs = ["models/models.h"],
4948 copts = xnnpack_std_cxxopts(),
4949 linkstatic = True,
4950 deps = [
4951 ":XNNPACK",
4952 "@FP16",
4953 "@pthreadpool",
4954 ],
4955)
4956
4957cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004958 name = "fp32_mobilenet_v3_small",
4959 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004960 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004961 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004962 linkstatic = True,
4963 deps = [
4964 ":XNNPACK",
4965 "@pthreadpool",
4966 ],
4967)
4968
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004969cc_library(
4970 name = "fp16_mobilenet_v3_small",
4971 srcs = ["models/fp16-mobilenet-v3-small.cc"],
4972 hdrs = ["models/models.h"],
4973 copts = xnnpack_std_cxxopts(),
4974 linkstatic = True,
4975 deps = [
4976 ":XNNPACK",
4977 "@FP16",
4978 "@pthreadpool",
4979 ],
4980)
4981
Marat Dukhanc068bb62019-10-04 13:24:39 -07004982xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004983 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004984 srcs = [
4985 "bench/f32-dwconv-e2e.cc",
4986 "bench/end2end.h",
4987 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004988 deps = MICROKERNEL_BENCHMARK_DEPS + [
4989 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004990 ":fp32_mobilenet_v1",
4991 ":fp32_mobilenet_v2",
4992 ":fp32_mobilenet_v3_large",
4993 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004994 ],
4995)
4996
4997xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004998 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004999 srcs = [
5000 "bench/f32-gemm-e2e.cc",
5001 "bench/end2end.h",
5002 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005003 deps = MICROKERNEL_BENCHMARK_DEPS + [
5004 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005005 ":fp32_mobilenet_v1",
5006 ":fp32_mobilenet_v2",
5007 ":fp32_mobilenet_v3_large",
5008 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005009 ],
5010)
5011
5012xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005013 name = "end2end_bench",
5014 srcs = ["bench/end2end.cc"],
5015 deps = [
5016 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005017 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005018 ":fp16_mobilenet_v1",
5019 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005020 ":fp16_mobilenet_v3_large",
5021 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005022 ":fp32_mobilenet_v1",
5023 ":fp32_mobilenet_v2",
5024 ":fp32_mobilenet_v3_large",
5025 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005026 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005027 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005028 "@pthreadpool",
5029 ],
5030)
5031
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005032#################### Accuracy evaluation for math functions ####################
5033
5034xnnpack_benchmark(
5035 name = "f32_exp_eval",
5036 srcs = [
5037 "eval/f32-exp.cc",
5038 "src/xnnpack/AlignedAllocator.h",
5039 ] + ACCURACY_EVAL_HDRS,
5040 deps = ACCURACY_EVAL_DEPS,
5041)
5042
Marat Dukhan515c9772019-10-17 18:07:57 -07005043xnnpack_benchmark(
5044 name = "f32_expminus_eval",
5045 srcs = [
5046 "eval/f32-expminus.cc",
5047 "src/xnnpack/AlignedAllocator.h",
5048 ] + ACCURACY_EVAL_HDRS,
5049 deps = ACCURACY_EVAL_DEPS,
5050)
5051
Marat Dukhan98ba4412019-10-23 02:14:28 -07005052xnnpack_benchmark(
5053 name = "f32_extexp_eval",
5054 srcs = [
5055 "eval/f32-extexp.cc",
5056 "src/xnnpack/AlignedAllocator.h",
5057 ] + ACCURACY_EVAL_HDRS,
5058 deps = ACCURACY_EVAL_DEPS,
5059)
5060
Marat Dukhan8853b822020-05-07 12:19:01 -07005061xnnpack_unit_test(
5062 name = "f32_roundne_eval",
5063 srcs = [
5064 "eval/f32-roundne.cc",
5065 "src/xnnpack/AlignedAllocator.h",
5066 "src/xnnpack/math-stubs.h",
5067 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005068 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005069 deps = MICROKERNEL_TEST_DEPS,
5070)
5071
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005072xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005073 name = "f32_roundd_eval",
5074 srcs = [
5075 "eval/f32-roundd.cc",
5076 "src/xnnpack/AlignedAllocator.h",
5077 "src/xnnpack/math-stubs.h",
5078 ] + MICROKERNEL_TEST_HDRS,
5079 automatic = False,
5080 deps = MICROKERNEL_TEST_DEPS,
5081)
5082
5083xnnpack_unit_test(
5084 name = "f32_roundu_eval",
5085 srcs = [
5086 "eval/f32-roundu.cc",
5087 "src/xnnpack/AlignedAllocator.h",
5088 "src/xnnpack/math-stubs.h",
5089 ] + MICROKERNEL_TEST_HDRS,
5090 automatic = False,
5091 deps = MICROKERNEL_TEST_DEPS,
5092)
5093
5094xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005095 name = "f32_roundz_eval",
5096 srcs = [
5097 "eval/f32-roundz.cc",
5098 "src/xnnpack/AlignedAllocator.h",
5099 "src/xnnpack/math-stubs.h",
5100 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005101 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005102 deps = MICROKERNEL_TEST_DEPS,
5103)
5104
Marat Dukhan346a9e52019-11-15 09:06:30 -08005105xnnpack_benchmark(
5106 name = "f32_sigmoid_eval",
5107 srcs = [
5108 "eval/f32-sigmoid.cc",
5109 "src/xnnpack/AlignedAllocator.h",
5110 ] + ACCURACY_EVAL_HDRS,
5111 deps = ACCURACY_EVAL_DEPS,
5112)
5113
Marat Dukhan84000762020-06-29 18:38:43 -07005114xnnpack_benchmark(
5115 name = "f32_sqrt_eval",
5116 srcs = [
5117 "eval/f32-sqrt.cc",
5118 "src/xnnpack/AlignedAllocator.h",
5119 ] + ACCURACY_EVAL_HDRS,
5120 deps = ACCURACY_EVAL_DEPS + [
5121 ":bench_utils",
5122 ],
5123)
5124
Marat Dukhan08c4a432019-10-03 09:29:21 -07005125######################### Unit tests for micro-kernels #########################
5126
5127xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005128 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005129 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005130 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005131 "test/gemm-microkernel-tester.h",
5132 "src/xnnpack/AlignedAllocator.h",
5133 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005134 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005135)
5136
5137xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005138 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005139 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005140 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005141 "test/spmm-microkernel-tester.h",
5142 "src/xnnpack/AlignedAllocator.h",
5143 ] + MICROKERNEL_TEST_HDRS,
5144 deps = MICROKERNEL_TEST_DEPS,
5145)
5146
5147xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005148 name = "f32_argmaxpool_test",
5149 srcs = [
5150 "test/f32-argmaxpool.cc",
5151 "test/argmaxpool-microkernel-tester.h",
5152 "src/xnnpack/AlignedAllocator.h",
5153 ] + MICROKERNEL_TEST_HDRS,
5154 deps = MICROKERNEL_TEST_DEPS,
5155)
5156
5157xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005158 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005159 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005160 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005161 "test/avgpool-microkernel-tester.h",
5162 "src/xnnpack/AlignedAllocator.h",
5163 ] + MICROKERNEL_TEST_HDRS,
5164 deps = MICROKERNEL_TEST_DEPS,
5165)
5166
5167xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005168 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005169 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005170 "test/f32-ibilinear.cc",
5171 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005172 "src/xnnpack/AlignedAllocator.h",
5173 ] + MICROKERNEL_TEST_HDRS,
5174 deps = MICROKERNEL_TEST_DEPS,
5175)
5176
5177xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005178 name = "f32_ibilinear_chw_test",
5179 srcs = [
5180 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005181 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005182 "src/xnnpack/AlignedAllocator.h",
5183 ] + MICROKERNEL_TEST_HDRS,
5184 deps = MICROKERNEL_TEST_DEPS,
5185)
5186
5187xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005188 name = "f16_clamp_test",
5189 srcs = [
5190 "test/f16-clamp.cc",
5191 "test/clamp-microkernel-tester.h",
5192 ] + MICROKERNEL_TEST_HDRS,
5193 deps = MICROKERNEL_TEST_DEPS,
5194)
5195
5196xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005197 name = "f32_clamp_test",
5198 srcs = [
5199 "test/f32-clamp.cc",
5200 "test/clamp-microkernel-tester.h",
5201 ] + MICROKERNEL_TEST_HDRS,
5202 deps = MICROKERNEL_TEST_DEPS,
5203)
5204
5205xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005206 name = "f32_igemm_test",
5207 srcs = [
5208 "test/f32-igemm.cc",
5209 "test/gemm-microkernel-tester.h",
5210 "src/xnnpack/AlignedAllocator.h",
5211 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005212 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005213)
5214
5215xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005216 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005217 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005218 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005219 "test/gemm-microkernel-tester.h",
5220 "src/xnnpack/AlignedAllocator.h",
5221 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005222 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005223)
5224
5225xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005226 name = "f32_igemm_minmax_test",
5227 srcs = [
5228 "test/f32-igemm-minmax.cc",
5229 "test/gemm-microkernel-tester.h",
5230 "src/xnnpack/AlignedAllocator.h",
5231 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005232 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005233)
5234
5235xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005236 name = "f16_igemm_minmax_test",
5237 srcs = [
5238 "test/f16-igemm-minmax.cc",
5239 "test/gemm-microkernel-tester.h",
5240 "src/xnnpack/AlignedAllocator.h",
5241 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005242 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005243)
5244
5245xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005246 name = "f32_conv_hwc_test",
5247 srcs = [
5248 "test/f32-conv-hwc.cc",
5249 "test/conv-hwc-microkernel-tester.h",
5250 "src/xnnpack/AlignedAllocator.h",
5251 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005252 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005253)
5254
5255xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005256 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005257 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005258 "test/f32-conv-hwc2chw.cc",
5259 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005260 "src/xnnpack/AlignedAllocator.h",
5261 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005262 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005263)
5264
5265xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005266 name = "f32_dwconv_test",
5267 srcs = [
5268 "test/f32-dwconv.cc",
5269 "test/dwconv-microkernel-tester.h",
5270 "src/xnnpack/AlignedAllocator.h",
5271 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005272 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005273)
5274
5275xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005276 name = "f16_dwconv_minmax_test",
5277 srcs = [
5278 "test/f16-dwconv-minmax.cc",
5279 "test/dwconv-microkernel-tester.h",
5280 "src/xnnpack/AlignedAllocator.h",
5281 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005282 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005283)
5284
5285xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005286 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005287 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005288 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005289 "test/dwconv-microkernel-tester.h",
5290 "src/xnnpack/AlignedAllocator.h",
5291 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005292 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005293)
5294
5295xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005296 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005297 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005298 "test/f32-dwconv2d-chw.cc",
5299 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005300 "src/xnnpack/AlignedAllocator.h",
5301 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005302 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005303)
5304
5305xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005306 name = "f16_gavgpool_minmax_test",
5307 srcs = [
5308 "test/f16-gavgpool-minmax.cc",
5309 "test/gavgpool-microkernel-tester.h",
5310 "src/xnnpack/AlignedAllocator.h",
5311 ] + MICROKERNEL_TEST_HDRS,
5312 deps = MICROKERNEL_TEST_DEPS,
5313)
5314
5315xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005316 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005317 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005318 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005319 "test/gavgpool-microkernel-tester.h",
5320 "src/xnnpack/AlignedAllocator.h",
5321 ] + MICROKERNEL_TEST_HDRS,
5322 deps = MICROKERNEL_TEST_DEPS,
5323)
5324
5325xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005326 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005327 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005328 "test/f32-gavgpool-cw.cc",
5329 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005330 "src/xnnpack/AlignedAllocator.h",
5331 ] + MICROKERNEL_TEST_HDRS,
5332 deps = MICROKERNEL_TEST_DEPS,
5333)
5334
5335xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005336 name = "f32_gemm_test",
5337 srcs = [
5338 "test/f32-gemm.cc",
5339 "test/gemm-microkernel-tester.h",
5340 "src/xnnpack/AlignedAllocator.h",
5341 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005342 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005343)
5344
5345xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005346 name = "f32_gemm_relu_test",
5347 srcs = [
5348 "test/f32-gemm-relu.cc",
5349 "test/gemm-microkernel-tester.h",
5350 "src/xnnpack/AlignedAllocator.h",
5351 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005352 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005353)
5354
5355xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005356 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005357 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005358 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005359 "test/gemm-microkernel-tester.h",
5360 "src/xnnpack/AlignedAllocator.h",
5361 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005362 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005363)
5364
5365xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005366 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005367 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005368 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005369 "test/gemm-microkernel-tester.h",
5370 "src/xnnpack/AlignedAllocator.h",
5371 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005372 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005373)
5374
5375xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005376 name = "f16_hswish_test",
5377 srcs = [
5378 "test/f16-hswish.cc",
5379 "test/hswish-microkernel-tester.h",
5380 ] + MICROKERNEL_TEST_HDRS,
5381 deps = MICROKERNEL_TEST_DEPS,
5382)
5383
5384xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005385 name = "f32_hswish_test",
5386 srcs = [
5387 "test/f32-hswish.cc",
5388 "test/hswish-microkernel-tester.h",
5389 ] + MICROKERNEL_TEST_HDRS,
5390 deps = MICROKERNEL_TEST_DEPS,
5391)
5392
5393xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005394 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005395 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005396 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005397 "test/maxpool-microkernel-tester.h",
5398 ] + MICROKERNEL_TEST_HDRS,
5399 deps = MICROKERNEL_TEST_DEPS,
5400)
5401
5402xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005403 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005404 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005405 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005406 "test/avgpool-microkernel-tester.h",
5407 "src/xnnpack/AlignedAllocator.h",
5408 ] + MICROKERNEL_TEST_HDRS,
5409 deps = MICROKERNEL_TEST_DEPS,
5410)
5411
5412xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005413 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005414 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005415 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005416 "test/gemm-microkernel-tester.h",
5417 "src/xnnpack/AlignedAllocator.h",
5418 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005419 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005420)
5421
5422xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005423 name = "f16_prelu_test",
5424 srcs = [
5425 "test/f16-prelu.cc",
5426 "test/prelu-microkernel-tester.h",
5427 "src/xnnpack/AlignedAllocator.h",
5428 ] + MICROKERNEL_TEST_HDRS,
5429 deps = MICROKERNEL_TEST_DEPS,
5430)
5431
5432xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005433 name = "f32_prelu_test",
5434 srcs = [
5435 "test/f32-prelu.cc",
5436 "test/prelu-microkernel-tester.h",
5437 "src/xnnpack/AlignedAllocator.h",
5438 ] + MICROKERNEL_TEST_HDRS,
5439 deps = MICROKERNEL_TEST_DEPS,
5440)
5441
5442xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005443 name = "f32_raddexpminusmax_test",
5444 srcs = [
5445 "test/f32-raddexpminusmax.cc",
5446 "test/raddexpminusmax-microkernel-tester.h",
5447 ] + MICROKERNEL_TEST_HDRS,
5448 deps = MICROKERNEL_TEST_DEPS,
5449)
5450
5451xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005452 name = "f32_raddextexp_test",
5453 srcs = [
5454 "test/f32-raddextexp.cc",
5455 "test/raddextexp-microkernel-tester.h",
5456 ] + MICROKERNEL_TEST_HDRS,
5457 deps = MICROKERNEL_TEST_DEPS,
5458)
5459
5460xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005461 name = "f32_raddstoreexpminusmax_test",
5462 srcs = [
5463 "test/f32-raddstoreexpminusmax.cc",
5464 "test/raddstoreexpminusmax-microkernel-tester.h",
5465 ] + MICROKERNEL_TEST_HDRS,
5466 deps = MICROKERNEL_TEST_DEPS,
5467)
5468
5469xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005470 name = "f32_relu_test",
5471 srcs = [
5472 "test/f32-relu.cc",
5473 "test/vunary-microkernel-tester.h",
5474 ] + MICROKERNEL_TEST_HDRS,
5475 deps = MICROKERNEL_TEST_DEPS,
5476)
5477
5478xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005479 name = "f32_rmax_test",
5480 srcs = [
5481 "test/f32-rmax.cc",
5482 "test/rmax-microkernel-tester.h",
5483 ] + MICROKERNEL_TEST_HDRS,
5484 deps = MICROKERNEL_TEST_DEPS,
5485)
5486
5487xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005488 name = "f32_sigmoid_test",
5489 srcs = [
5490 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005491 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005492 ] + MICROKERNEL_TEST_HDRS,
5493 deps = MICROKERNEL_TEST_DEPS,
5494)
5495
5496xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005497 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005498 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005499 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005500 "test/spmm-microkernel-tester.h",
5501 "src/xnnpack/AlignedAllocator.h",
5502 ] + MICROKERNEL_TEST_HDRS,
5503 deps = MICROKERNEL_TEST_DEPS,
5504)
5505
5506xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005507 name = "f16_vadd_minmax_test",
5508 srcs = [
5509 "test/f16-vadd-minmax.cc",
5510 "test/vbinary-microkernel-tester.h",
5511 ] + MICROKERNEL_TEST_HDRS,
5512 deps = MICROKERNEL_TEST_DEPS,
5513)
5514
5515xnnpack_unit_test(
5516 name = "f16_vaddc_minmax_test",
5517 srcs = [
5518 "test/f16-vaddc-minmax.cc",
5519 "test/vbinaryc-microkernel-tester.h",
5520 ] + MICROKERNEL_TEST_HDRS,
5521 deps = MICROKERNEL_TEST_DEPS,
5522)
5523
5524xnnpack_unit_test(
5525 name = "f16_vdiv_minmax_test",
5526 srcs = [
5527 "test/f16-vdiv-minmax.cc",
5528 "test/vbinary-microkernel-tester.h",
5529 ] + MICROKERNEL_TEST_HDRS,
5530 deps = MICROKERNEL_TEST_DEPS,
5531)
5532
5533xnnpack_unit_test(
5534 name = "f16_vdivc_minmax_test",
5535 srcs = [
5536 "test/f16-vdivc-minmax.cc",
5537 "test/vbinaryc-microkernel-tester.h",
5538 ] + MICROKERNEL_TEST_HDRS,
5539 deps = MICROKERNEL_TEST_DEPS,
5540)
5541
5542xnnpack_unit_test(
5543 name = "f16_vrdivc_minmax_test",
5544 srcs = [
5545 "test/f16-vrdivc-minmax.cc",
5546 "test/vbinaryc-microkernel-tester.h",
5547 ] + MICROKERNEL_TEST_HDRS,
5548 deps = MICROKERNEL_TEST_DEPS,
5549)
5550
5551xnnpack_unit_test(
5552 name = "f16_vmax_test",
5553 srcs = [
5554 "test/f16-vmax.cc",
5555 "test/vbinary-microkernel-tester.h",
5556 ] + MICROKERNEL_TEST_HDRS,
5557 deps = MICROKERNEL_TEST_DEPS,
5558)
5559
5560xnnpack_unit_test(
5561 name = "f16_vmaxc_test",
5562 srcs = [
5563 "test/f16-vmaxc.cc",
5564 "test/vbinaryc-microkernel-tester.h",
5565 ] + MICROKERNEL_TEST_HDRS,
5566 deps = MICROKERNEL_TEST_DEPS,
5567)
5568
5569xnnpack_unit_test(
5570 name = "f16_vmin_test",
5571 srcs = [
5572 "test/f16-vmin.cc",
5573 "test/vbinary-microkernel-tester.h",
5574 ] + MICROKERNEL_TEST_HDRS,
5575 deps = MICROKERNEL_TEST_DEPS,
5576)
5577
5578xnnpack_unit_test(
5579 name = "f16_vminc_test",
5580 srcs = [
5581 "test/f16-vminc.cc",
5582 "test/vbinaryc-microkernel-tester.h",
5583 ] + MICROKERNEL_TEST_HDRS,
5584 deps = MICROKERNEL_TEST_DEPS,
5585)
5586
5587xnnpack_unit_test(
5588 name = "f16_vmul_minmax_test",
5589 srcs = [
5590 "test/f16-vmul-minmax.cc",
5591 "test/vbinary-microkernel-tester.h",
5592 ] + MICROKERNEL_TEST_HDRS,
5593 deps = MICROKERNEL_TEST_DEPS,
5594)
5595
5596xnnpack_unit_test(
5597 name = "f16_vmulc_minmax_test",
5598 srcs = [
5599 "test/f16-vmulc-minmax.cc",
5600 "test/vbinaryc-microkernel-tester.h",
5601 ] + MICROKERNEL_TEST_HDRS,
5602 deps = MICROKERNEL_TEST_DEPS,
5603)
5604
5605xnnpack_unit_test(
5606 name = "f16_vsub_minmax_test",
5607 srcs = [
5608 "test/f16-vsub-minmax.cc",
5609 "test/vbinary-microkernel-tester.h",
5610 ] + MICROKERNEL_TEST_HDRS,
5611 deps = MICROKERNEL_TEST_DEPS,
5612)
5613
5614xnnpack_unit_test(
5615 name = "f16_vsubc_minmax_test",
5616 srcs = [
5617 "test/f16-vsubc-minmax.cc",
5618 "test/vbinaryc-microkernel-tester.h",
5619 ] + MICROKERNEL_TEST_HDRS,
5620 deps = MICROKERNEL_TEST_DEPS,
5621)
5622
5623xnnpack_unit_test(
5624 name = "f16_vrsubc_minmax_test",
5625 srcs = [
5626 "test/f16-vrsubc-minmax.cc",
5627 "test/vbinaryc-microkernel-tester.h",
5628 ] + MICROKERNEL_TEST_HDRS,
5629 deps = MICROKERNEL_TEST_DEPS,
5630)
5631
5632xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005633 name = "f32_vabs_test",
5634 srcs = [
5635 "test/f32-vabs.cc",
5636 "test/vunary-microkernel-tester.h",
5637 ] + MICROKERNEL_TEST_HDRS,
5638 deps = MICROKERNEL_TEST_DEPS,
5639)
5640
5641xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005642 name = "f32_vadd_test",
5643 srcs = [
5644 "test/f32-vadd.cc",
5645 "test/vbinary-microkernel-tester.h",
5646 ] + MICROKERNEL_TEST_HDRS,
5647 deps = MICROKERNEL_TEST_DEPS,
5648)
5649
5650xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005651 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005652 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005653 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005654 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005655 ] + MICROKERNEL_TEST_HDRS,
5656 deps = MICROKERNEL_TEST_DEPS,
5657)
5658
5659xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005660 name = "f32_vadd_relu_test",
5661 srcs = [
5662 "test/f32-vadd-relu.cc",
5663 "test/vbinary-microkernel-tester.h",
5664 ] + MICROKERNEL_TEST_HDRS,
5665 deps = MICROKERNEL_TEST_DEPS,
5666)
5667
5668xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005669 name = "f32_vaddc_test",
5670 srcs = [
5671 "test/f32-vaddc.cc",
5672 "test/vbinaryc-microkernel-tester.h",
5673 ] + MICROKERNEL_TEST_HDRS,
5674 deps = MICROKERNEL_TEST_DEPS,
5675)
5676
5677xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005678 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005679 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005680 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005681 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005682 ] + MICROKERNEL_TEST_HDRS,
5683 deps = MICROKERNEL_TEST_DEPS,
5684)
5685
5686xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005687 name = "f32_vaddc_relu_test",
5688 srcs = [
5689 "test/f32-vaddc-relu.cc",
5690 "test/vbinaryc-microkernel-tester.h",
5691 ] + MICROKERNEL_TEST_HDRS,
5692 deps = MICROKERNEL_TEST_DEPS,
5693)
5694
5695xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005696 name = "f32_vdiv_test",
5697 srcs = [
5698 "test/f32-vdiv.cc",
5699 "test/vbinary-microkernel-tester.h",
5700 ] + MICROKERNEL_TEST_HDRS,
5701 deps = MICROKERNEL_TEST_DEPS,
5702)
5703
5704xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005705 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005706 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005707 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005708 "test/vbinary-microkernel-tester.h",
5709 ] + MICROKERNEL_TEST_HDRS,
5710 deps = MICROKERNEL_TEST_DEPS,
5711)
5712
5713xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005714 name = "f32_vdiv_relu_test",
5715 srcs = [
5716 "test/f32-vdiv-relu.cc",
5717 "test/vbinary-microkernel-tester.h",
5718 ] + MICROKERNEL_TEST_HDRS,
5719 deps = MICROKERNEL_TEST_DEPS,
5720)
5721
5722xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005723 name = "f32_vdivc_test",
5724 srcs = [
5725 "test/f32-vdivc.cc",
5726 "test/vbinaryc-microkernel-tester.h",
5727 ] + MICROKERNEL_TEST_HDRS,
5728 deps = MICROKERNEL_TEST_DEPS,
5729)
5730
5731xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005732 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005733 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005734 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005735 "test/vbinaryc-microkernel-tester.h",
5736 ] + MICROKERNEL_TEST_HDRS,
5737 deps = MICROKERNEL_TEST_DEPS,
5738)
5739
5740xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005741 name = "f32_vdivc_relu_test",
5742 srcs = [
5743 "test/f32-vdivc-relu.cc",
5744 "test/vbinaryc-microkernel-tester.h",
5745 ] + MICROKERNEL_TEST_HDRS,
5746 deps = MICROKERNEL_TEST_DEPS,
5747)
5748
5749xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005750 name = "f32_vrdivc_test",
5751 srcs = [
5752 "test/f32-vrdivc.cc",
5753 "test/vbinaryc-microkernel-tester.h",
5754 ] + MICROKERNEL_TEST_HDRS,
5755 deps = MICROKERNEL_TEST_DEPS,
5756)
5757
5758xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005759 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005760 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005761 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005762 "test/vbinaryc-microkernel-tester.h",
5763 ] + MICROKERNEL_TEST_HDRS,
5764 deps = MICROKERNEL_TEST_DEPS,
5765)
5766
5767xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005768 name = "f32_vrdivc_relu_test",
5769 srcs = [
5770 "test/f32-vrdivc-relu.cc",
5771 "test/vbinaryc-microkernel-tester.h",
5772 ] + MICROKERNEL_TEST_HDRS,
5773 deps = MICROKERNEL_TEST_DEPS,
5774)
5775
5776xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005777 name = "f32_vmax_test",
5778 srcs = [
5779 "test/f32-vmax.cc",
5780 "test/vbinary-microkernel-tester.h",
5781 ] + MICROKERNEL_TEST_HDRS,
5782 deps = MICROKERNEL_TEST_DEPS,
5783)
5784
5785xnnpack_unit_test(
5786 name = "f32_vmaxc_test",
5787 srcs = [
5788 "test/f32-vmaxc.cc",
5789 "test/vbinaryc-microkernel-tester.h",
5790 ] + MICROKERNEL_TEST_HDRS,
5791 deps = MICROKERNEL_TEST_DEPS,
5792)
5793
5794xnnpack_unit_test(
5795 name = "f32_vmin_test",
5796 srcs = [
5797 "test/f32-vmin.cc",
5798 "test/vbinary-microkernel-tester.h",
5799 ] + MICROKERNEL_TEST_HDRS,
5800 deps = MICROKERNEL_TEST_DEPS,
5801)
5802
5803xnnpack_unit_test(
5804 name = "f32_vminc_test",
5805 srcs = [
5806 "test/f32-vminc.cc",
5807 "test/vbinaryc-microkernel-tester.h",
5808 ] + MICROKERNEL_TEST_HDRS,
5809 deps = MICROKERNEL_TEST_DEPS,
5810)
5811
5812xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005813 name = "f32_vmul_test",
5814 srcs = [
5815 "test/f32-vmul.cc",
5816 "test/vbinary-microkernel-tester.h",
5817 ] + MICROKERNEL_TEST_HDRS,
5818 deps = MICROKERNEL_TEST_DEPS,
5819)
5820
5821xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005822 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005823 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005824 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005825 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005826 ] + MICROKERNEL_TEST_HDRS,
5827 deps = MICROKERNEL_TEST_DEPS,
5828)
5829
5830xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005831 name = "f32_vmul_relu_test",
5832 srcs = [
5833 "test/f32-vmul-relu.cc",
5834 "test/vbinary-microkernel-tester.h",
5835 ] + MICROKERNEL_TEST_HDRS,
5836 deps = MICROKERNEL_TEST_DEPS,
5837)
5838
5839xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005840 name = "f32_vmulc_test",
5841 srcs = [
5842 "test/f32-vmulc.cc",
5843 "test/vbinaryc-microkernel-tester.h",
5844 ] + MICROKERNEL_TEST_HDRS,
5845 deps = MICROKERNEL_TEST_DEPS,
5846)
5847
5848xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005849 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005850 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005851 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005852 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005853 ] + MICROKERNEL_TEST_HDRS,
5854 deps = MICROKERNEL_TEST_DEPS,
5855)
5856
5857xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005858 name = "f32_vmulc_relu_test",
5859 srcs = [
5860 "test/f32-vmulc-relu.cc",
5861 "test/vbinaryc-microkernel-tester.h",
5862 ] + MICROKERNEL_TEST_HDRS,
5863 deps = MICROKERNEL_TEST_DEPS,
5864)
5865
5866xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005867 name = "f16_vmulcaddc_minmax_test",
5868 srcs = [
5869 "test/f16-vmulcaddc-minmax.cc",
5870 "test/vmulcaddc-microkernel-tester.h",
5871 "src/xnnpack/AlignedAllocator.h",
5872 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005873 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005874)
5875
5876xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005877 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005878 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005879 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005880 "test/vmulcaddc-microkernel-tester.h",
5881 "src/xnnpack/AlignedAllocator.h",
5882 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005883 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005884)
5885
5886xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005887 name = "f32_vlrelu_test",
5888 srcs = [
5889 "test/f32-vlrelu.cc",
5890 "test/vunary-microkernel-tester.h",
5891 ] + MICROKERNEL_TEST_HDRS,
5892 deps = MICROKERNEL_TEST_DEPS,
5893)
5894
5895xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005896 name = "f32_vneg_test",
5897 srcs = [
5898 "test/f32-vneg.cc",
5899 "test/vunary-microkernel-tester.h",
5900 ] + MICROKERNEL_TEST_HDRS,
5901 deps = MICROKERNEL_TEST_DEPS,
5902)
5903
5904xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005905 name = "f32_vrndne_test",
5906 srcs = [
5907 "test/f32-vrndne.cc",
5908 "test/vunary-microkernel-tester.h",
5909 ] + MICROKERNEL_TEST_HDRS,
5910 deps = MICROKERNEL_TEST_DEPS,
5911)
5912
5913xnnpack_unit_test(
5914 name = "f32_vrndz_test",
5915 srcs = [
5916 "test/f32-vrndz.cc",
5917 "test/vunary-microkernel-tester.h",
5918 ] + MICROKERNEL_TEST_HDRS,
5919 deps = MICROKERNEL_TEST_DEPS,
5920)
5921
5922xnnpack_unit_test(
5923 name = "f32_vrndu_test",
5924 srcs = [
5925 "test/f32-vrndu.cc",
5926 "test/vunary-microkernel-tester.h",
5927 ] + MICROKERNEL_TEST_HDRS,
5928 deps = MICROKERNEL_TEST_DEPS,
5929)
5930
5931xnnpack_unit_test(
5932 name = "f32_vrndd_test",
5933 srcs = [
5934 "test/f32-vrndd.cc",
5935 "test/vunary-microkernel-tester.h",
5936 ] + MICROKERNEL_TEST_HDRS,
5937 deps = MICROKERNEL_TEST_DEPS,
5938)
5939
5940xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005941 name = "f32_vscale_test",
5942 srcs = [
5943 "test/f32-vscale.cc",
5944 "test/vscale-microkernel-tester.h",
5945 ] + MICROKERNEL_TEST_HDRS,
5946 deps = MICROKERNEL_TEST_DEPS,
5947)
5948
5949xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005950 name = "f32_vscaleexpminusmax_test",
5951 srcs = [
5952 "test/f32-vscaleexpminusmax.cc",
5953 "test/vscaleexpminusmax-microkernel-tester.h",
5954 ] + MICROKERNEL_TEST_HDRS,
5955 deps = MICROKERNEL_TEST_DEPS,
5956)
5957
5958xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005959 name = "f32_vscaleextexp_test",
5960 srcs = [
5961 "test/f32-vscaleextexp.cc",
5962 "test/vscaleextexp-microkernel-tester.h",
5963 ] + MICROKERNEL_TEST_HDRS,
5964 deps = MICROKERNEL_TEST_DEPS,
5965)
5966
5967xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005968 name = "f32_vsqr_test",
5969 srcs = [
5970 "test/f32-vsqr.cc",
5971 "test/vunary-microkernel-tester.h",
5972 ] + MICROKERNEL_TEST_HDRS,
5973 deps = MICROKERNEL_TEST_DEPS,
5974)
5975
5976xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005977 name = "f32_vsqrdiff_test",
5978 srcs = [
5979 "test/f32-vsqrdiff.cc",
5980 "test/vbinary-microkernel-tester.h",
5981 ] + MICROKERNEL_TEST_HDRS,
5982 deps = MICROKERNEL_TEST_DEPS,
5983)
5984
5985xnnpack_unit_test(
5986 name = "f32_vsqrdiffc_test",
5987 srcs = [
5988 "test/f32-vsqrdiffc.cc",
5989 "test/vbinaryc-microkernel-tester.h",
5990 ] + MICROKERNEL_TEST_HDRS,
5991 deps = MICROKERNEL_TEST_DEPS,
5992)
5993
5994xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005995 name = "f32_vsqrt_test",
5996 srcs = [
5997 "test/f32-vsqrt.cc",
5998 "test/vunary-microkernel-tester.h",
5999 ] + MICROKERNEL_TEST_HDRS,
6000 deps = MICROKERNEL_TEST_DEPS,
6001)
6002
6003xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006004 name = "f32_vsub_test",
6005 srcs = [
6006 "test/f32-vsub.cc",
6007 "test/vbinary-microkernel-tester.h",
6008 ] + MICROKERNEL_TEST_HDRS,
6009 deps = MICROKERNEL_TEST_DEPS,
6010)
6011
6012xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006013 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006014 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006015 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006016 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006017 ] + MICROKERNEL_TEST_HDRS,
6018 deps = MICROKERNEL_TEST_DEPS,
6019)
6020
6021xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006022 name = "f32_vsub_relu_test",
6023 srcs = [
6024 "test/f32-vsub-relu.cc",
6025 "test/vbinary-microkernel-tester.h",
6026 ] + MICROKERNEL_TEST_HDRS,
6027 deps = MICROKERNEL_TEST_DEPS,
6028)
6029
6030xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006031 name = "f32_vsubc_test",
6032 srcs = [
6033 "test/f32-vsubc.cc",
6034 "test/vbinaryc-microkernel-tester.h",
6035 ] + MICROKERNEL_TEST_HDRS,
6036 deps = MICROKERNEL_TEST_DEPS,
6037)
6038
6039xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006040 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006041 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006042 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006043 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006044 ] + MICROKERNEL_TEST_HDRS,
6045 deps = MICROKERNEL_TEST_DEPS,
6046)
6047
6048xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006049 name = "f32_vsubc_relu_test",
6050 srcs = [
6051 "test/f32-vsubc-relu.cc",
6052 "test/vbinaryc-microkernel-tester.h",
6053 ] + MICROKERNEL_TEST_HDRS,
6054 deps = MICROKERNEL_TEST_DEPS,
6055)
6056
6057xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006058 name = "f32_vrsubc_test",
6059 srcs = [
6060 "test/f32-vrsubc.cc",
6061 "test/vbinaryc-microkernel-tester.h",
6062 ] + MICROKERNEL_TEST_HDRS,
6063 deps = MICROKERNEL_TEST_DEPS,
6064)
6065
6066xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006067 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006068 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006069 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006070 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006071 ] + MICROKERNEL_TEST_HDRS,
6072 deps = MICROKERNEL_TEST_DEPS,
6073)
6074
6075xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006076 name = "f32_vrsubc_relu_test",
6077 srcs = [
6078 "test/f32-vrsubc-relu.cc",
6079 "test/vbinaryc-microkernel-tester.h",
6080 ] + MICROKERNEL_TEST_HDRS,
6081 deps = MICROKERNEL_TEST_DEPS,
6082)
6083
6084xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006085 name = "qs8_dwconv_minmax_test",
6086 srcs = [
6087 "test/qs8-dwconv-minmax.cc",
6088 "test/dwconv-microkernel-tester.h",
6089 "src/xnnpack/AlignedAllocator.h",
6090 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6091 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6092)
6093
6094xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006095 name = "qs8_gavgpool_minmax_test",
6096 srcs = [
6097 "test/qs8-gavgpool-minmax.cc",
6098 "test/gavgpool-microkernel-tester.h",
6099 "src/xnnpack/AlignedAllocator.h",
6100 ] + MICROKERNEL_TEST_HDRS,
6101 deps = MICROKERNEL_TEST_DEPS,
6102)
6103
6104xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006105 name = "qs8_gemm_minmax_test",
6106 srcs = [
6107 "test/qs8-gemm-minmax.cc",
6108 "test/gemm-microkernel-tester.h",
6109 "src/xnnpack/AlignedAllocator.h",
6110 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6111 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6112)
6113
6114xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006115 name = "qs8_igemm_minmax_test",
6116 srcs = [
6117 "test/qs8-igemm-minmax.cc",
6118 "test/gemm-microkernel-tester.h",
6119 "src/xnnpack/AlignedAllocator.h",
6120 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6121 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6122)
6123
6124xnnpack_unit_test(
6125 name = "qs8_requantization_test",
6126 srcs = [
6127 "src/xnnpack/requantization-stubs.h",
6128 "test/qs8-requantization.cc",
6129 "test/requantization-tester.h",
6130 ] + MICROKERNEL_TEST_HDRS,
6131 deps = MICROKERNEL_TEST_DEPS,
6132)
6133
6134xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006135 name = "qs8_vadd_minmax_test",
6136 srcs = [
6137 "test/qs8-vadd-minmax.cc",
6138 "test/vadd-microkernel-tester.h",
6139 ] + MICROKERNEL_TEST_HDRS,
6140 deps = MICROKERNEL_TEST_DEPS,
6141)
6142
6143xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006144 name = "qs8_vaddc_minmax_test",
6145 srcs = [
6146 "test/qs8-vaddc-minmax.cc",
6147 "test/vaddc-microkernel-tester.h",
6148 ] + MICROKERNEL_TEST_HDRS,
6149 deps = MICROKERNEL_TEST_DEPS,
6150)
6151
6152xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006153 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006154 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006155 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006156 "test/avgpool-microkernel-tester.h",
6157 "src/xnnpack/AlignedAllocator.h",
6158 ] + MICROKERNEL_TEST_HDRS,
6159 deps = MICROKERNEL_TEST_DEPS,
6160)
6161
6162xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006163 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006164 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006165 "test/qu8-dwconv-minmax.cc",
6166 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006167 "src/xnnpack/AlignedAllocator.h",
6168 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006169 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006170)
6171
6172xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006173 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006174 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006175 "test/qu8-igemm-minmax.cc",
6176 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006177 "src/xnnpack/AlignedAllocator.h",
6178 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006179 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006180)
6181
6182xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006183 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006184 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006185 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006186 "test/gavgpool-microkernel-tester.h",
6187 "src/xnnpack/AlignedAllocator.h",
6188 ] + MICROKERNEL_TEST_HDRS,
6189 deps = MICROKERNEL_TEST_DEPS,
6190)
6191
6192xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006193 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006194 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006195 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006196 "test/gemm-microkernel-tester.h",
6197 "src/xnnpack/AlignedAllocator.h",
6198 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006199 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006200)
6201
6202xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006203 name = "qu8_requantization_test",
6204 srcs = [
6205 "src/xnnpack/requantization-stubs.h",
6206 "test/qu8-requantization.cc",
6207 "test/requantization-tester.h",
6208 ] + MICROKERNEL_TEST_HDRS,
6209 deps = MICROKERNEL_TEST_DEPS,
6210)
6211
6212xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006213 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006214 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006215 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006216 "test/vadd-microkernel-tester.h",
6217 ] + MICROKERNEL_TEST_HDRS,
6218 deps = MICROKERNEL_TEST_DEPS,
6219)
6220
6221xnnpack_unit_test(
6222 name = "u8_clamp_test",
6223 srcs = [
6224 "test/u8-clamp.cc",
6225 "test/clamp-microkernel-tester.h",
6226 ] + MICROKERNEL_TEST_HDRS,
6227 deps = MICROKERNEL_TEST_DEPS,
6228)
6229
6230xnnpack_unit_test(
6231 name = "u8_lut32norm_test",
6232 srcs = [
6233 "test/u8-lut32norm.cc",
6234 "test/lut-norm-microkernel-tester.h",
6235 ] + MICROKERNEL_TEST_HDRS,
6236 deps = MICROKERNEL_TEST_DEPS,
6237)
6238
6239xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006240 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006241 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006242 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006243 "test/maxpool-microkernel-tester.h",
6244 ] + MICROKERNEL_TEST_HDRS,
6245 deps = MICROKERNEL_TEST_DEPS,
6246)
6247
6248xnnpack_unit_test(
6249 name = "u8_rmax_test",
6250 srcs = [
6251 "test/u8-rmax.cc",
6252 "test/rmax-microkernel-tester.h",
6253 ] + MICROKERNEL_TEST_HDRS,
6254 deps = MICROKERNEL_TEST_DEPS,
6255)
6256
6257xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006258 name = "x32_fill_test",
6259 srcs = [
6260 "test/x32-fill.cc",
6261 "test/fill-microkernel-tester.h",
6262 ] + MICROKERNEL_TEST_HDRS,
6263 deps = MICROKERNEL_TEST_DEPS,
6264)
6265
6266xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006267 name = "x32_packx_test",
6268 srcs = [
6269 "test/x32-packx.cc",
6270 "test/pack-microkernel-tester.h",
6271 "src/xnnpack/AlignedAllocator.h",
6272 ] + MICROKERNEL_TEST_HDRS,
6273 deps = MICROKERNEL_TEST_DEPS,
6274)
6275
6276xnnpack_unit_test(
6277 name = "x32_pad_test",
6278 srcs = [
6279 "test/x32-pad.cc",
6280 "test/pad-microkernel-tester.h",
6281 ] + MICROKERNEL_TEST_HDRS,
6282 deps = MICROKERNEL_TEST_DEPS,
6283)
6284
6285xnnpack_unit_test(
6286 name = "x32_unpool_test",
6287 srcs = [
6288 "test/x32-unpool.cc",
6289 "test/unpool-microkernel-tester.h",
6290 ] + MICROKERNEL_TEST_HDRS,
6291 deps = MICROKERNEL_TEST_DEPS,
6292)
6293
6294xnnpack_unit_test(
6295 name = "x32_zip_test",
6296 srcs = [
6297 "test/x32-zip.cc",
6298 "test/zip-microkernel-tester.h",
6299 ] + MICROKERNEL_TEST_HDRS,
6300 deps = MICROKERNEL_TEST_DEPS,
6301)
6302
6303xnnpack_unit_test(
6304 name = "x8_lut_test",
6305 srcs = [
6306 "test/x8-lut.cc",
6307 "test/lut-microkernel-tester.h",
6308 ] + MICROKERNEL_TEST_HDRS,
6309 deps = MICROKERNEL_TEST_DEPS,
6310)
6311
6312xnnpack_unit_test(
6313 name = "x8_zip_test",
6314 srcs = [
6315 "test/x8-zip.cc",
6316 "test/zip-microkernel-tester.h",
6317 ] + MICROKERNEL_TEST_HDRS,
6318 deps = MICROKERNEL_TEST_DEPS,
6319)
6320
Marat Dukhan20c3b922020-03-10 03:45:06 -07006321########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006322
6323xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006324 name = "operator_size_test",
6325 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006326 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006327)
6328
Marat Dukhan20c3b922020-03-10 03:45:06 -07006329xnnpack_binary(
6330 name = "subgraph_size_test",
6331 srcs = ["test/subgraph-size.c"],
6332 deps = [":XNNPACK"],
6333)
6334
6335########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006336
6337xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006338 name = "abs_nc_test",
6339 srcs = [
6340 "test/abs-nc.cc",
6341 "test/abs-operator-tester.h",
6342 ],
6343 deps = OPERATOR_TEST_DEPS,
6344)
6345
6346xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006347 name = "add_nd_test",
6348 srcs = [
6349 "test/add-nd.cc",
6350 "test/binary-elementwise-operator-tester.h",
6351 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006352 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006353)
6354
6355xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006356 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006357 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006358 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006359 "test/argmax-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 Dukhanefc47b82019-11-18 09:25:38 -08006365 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006366 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006367 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006368 "test/average-pooling-operator-tester.h",
6369 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006370 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006371)
6372
6373xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006374 name = "bankers_rounding_nc_test",
6375 srcs = [
6376 "test/bankers-rounding-nc.cc",
6377 "test/bankers-rounding-operator-tester.h",
6378 ],
6379 deps = OPERATOR_TEST_DEPS,
6380)
6381
6382xnnpack_unit_test(
6383 name = "ceiling_nc_test",
6384 srcs = [
6385 "test/ceiling-nc.cc",
6386 "test/ceiling-operator-tester.h",
6387 ],
6388 deps = OPERATOR_TEST_DEPS,
6389)
6390
6391xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006392 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006393 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006394 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006395 "test/channel-shuffle-operator-tester.h",
6396 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006397 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006398)
6399
6400xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006401 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006402 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006403 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006404 "test/clamp-operator-tester.h",
6405 ],
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 Dukhan065b11e2020-05-22 09:49:41 -07006410 name = "constant_pad_nd_test",
6411 srcs = [
6412 "test/constant-pad-nd.cc",
6413 "test/constant-pad-operator-tester.h",
6414 ],
6415 deps = OPERATOR_TEST_DEPS,
6416)
6417
6418xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006419 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006420 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006421 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006422 "test/convolution-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 Dukhanefc47b82019-11-18 09:25:38 -08006428 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006429 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006430 "test/convolution-nchw.cc",
6431 "test/convolution-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 Dukhan4e21b272020-06-04 18:45:01 -07006437 name = "copy_nc_test",
6438 srcs = [
6439 "test/copy-nc.cc",
6440 "test/copy-operator-tester.h",
6441 ],
6442 deps = OPERATOR_TEST_DEPS,
6443)
6444
6445xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006446 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006447 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006448 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006449 "test/deconvolution-operator-tester.h",
6450 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006451 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006452)
6453
6454xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006455 name = "divide_nd_test",
6456 srcs = [
6457 "test/binary-elementwise-operator-tester.h",
6458 "test/divide-nd.cc",
6459 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006460 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006461)
6462
6463xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006464 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006465 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006466 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006467 "test/fully-connected-operator-tester.h",
6468 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006469 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006470)
6471
6472xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006473 name = "floor_nc_test",
6474 srcs = [
6475 "test/floor-nc.cc",
6476 "test/floor-operator-tester.h",
6477 ],
6478 deps = OPERATOR_TEST_DEPS,
6479)
6480
6481xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006482 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006483 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006484 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006485 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006486 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006487 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006488)
6489
6490xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006491 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006492 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006493 "test/global-average-pooling-ncw.cc",
6494 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006495 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006496 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006497)
6498
6499xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006500 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006501 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006502 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006503 "test/hardswish-operator-tester.h",
6504 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006505 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006506)
6507
6508xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006509 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006510 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006511 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006512 "test/leaky-relu-operator-tester.h",
6513 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006514 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006515)
6516
6517xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006518 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006519 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006520 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006521 "test/max-pooling-operator-tester.h",
6522 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006523 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006524)
6525
6526xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006527 name = "maximum_nd_test",
6528 srcs = [
6529 "test/binary-elementwise-operator-tester.h",
6530 "test/maximum-nd.cc",
6531 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006532 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006533)
6534
6535xnnpack_unit_test(
6536 name = "minimum_nd_test",
6537 srcs = [
6538 "test/binary-elementwise-operator-tester.h",
6539 "test/minimum-nd.cc",
6540 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006541 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006542)
6543
6544xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006545 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006546 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006547 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006548 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006549 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006550 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006551)
6552
6553xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006554 name = "negate_nc_test",
6555 srcs = [
6556 "test/negate-nc.cc",
6557 "test/negate-operator-tester.h",
6558 ],
6559 deps = OPERATOR_TEST_DEPS,
6560)
6561
6562xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006563 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006564 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006565 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006566 "test/prelu-operator-tester.h",
6567 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006568 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006569)
6570
6571xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006572 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006573 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006574 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006575 "test/resize-bilinear-operator-tester.h",
6576 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006577 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006578)
6579
6580xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07006581 name = "resize_bilinear_nchw_test",
6582 srcs = [
6583 "test/resize-bilinear-nchw.cc",
6584 "test/resize-bilinear-operator-tester.h",
6585 ] + OPERATOR_TEST_PARAMS_HDRS,
6586 deps = OPERATOR_TEST_DEPS,
6587)
6588
6589xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006590 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006591 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006592 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006593 "test/sigmoid-operator-tester.h",
6594 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006595 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006596)
6597
6598xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006599 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006600 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006601 "test/softmax-nc.cc",
6602 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006603 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006604 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006605)
6606
6607xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006608 name = "square_nc_test",
6609 srcs = [
6610 "test/square-nc.cc",
6611 "test/square-operator-tester.h",
6612 ],
6613 deps = OPERATOR_TEST_DEPS,
6614)
6615
6616xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006617 name = "square_root_nc_test",
6618 srcs = [
6619 "test/square-root-nc.cc",
6620 "test/square-root-operator-tester.h",
6621 ],
6622 deps = OPERATOR_TEST_DEPS,
6623)
6624
6625xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006626 name = "squared_difference_nd_test",
6627 srcs = [
6628 "test/binary-elementwise-operator-tester.h",
6629 "test/squared-difference-nd.cc",
6630 ],
6631 deps = OPERATOR_TEST_DEPS,
6632)
6633
6634xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006635 name = "subtract_nd_test",
6636 srcs = [
6637 "test/binary-elementwise-operator-tester.h",
6638 "test/subtract-nd.cc",
6639 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006640 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006641)
6642
6643xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006644 name = "truncation_nc_test",
6645 srcs = [
6646 "test/truncation-nc.cc",
6647 "test/truncation-operator-tester.h",
6648 ],
6649 deps = OPERATOR_TEST_DEPS,
6650)
6651
6652xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006653 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006654 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006655 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006656 "test/unpooling-operator-tester.h",
6657 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006658 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006659)
6660
Chao Mei6ddfc602020-05-13 22:29:36 -07006661############################### Misc unit tests ###############################
6662
6663xnnpack_unit_test(
6664 name = "memory_planner_test",
6665 srcs = [
6666 "test/memory-planner-test.cc",
6667 ],
6668 deps = [
6669 ":XNNPACK",
6670 ":memory_planner",
6671 ],
6672)
6673
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07006674xnnpack_unit_test(
6675 name = "subgraph_nchw_test",
6676 srcs = [
6677 "src/xnnpack/subgraph.h",
6678 "test/subgraph-nchw.cc",
6679 "test/subgraph-tester.h",
6680 ],
6681 deps = [
6682 ":XNNPACK",
6683 ],
6684)
6685
Marat Dukhan08c4a432019-10-03 09:29:21 -07006686############################# Build configurations #############################
6687
Marat Dukhanb8642352019-10-30 15:43:02 -07006688# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006689config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006690 name = "xnn_enable_assembly_explicit_true",
6691 define_values = {"xnn_enable_assembly": "true"},
6692)
6693
6694# Disables usage of assembly kernels.
6695config_setting(
6696 name = "xnn_enable_assembly_explicit_false",
6697 define_values = {"xnn_enable_assembly": "false"},
6698)
6699
Marat Dukhan9de90e02020-06-18 16:04:12 -07006700# Enables usage of sparse inference.
6701config_setting(
6702 name = "xnn_enable_sparse_explicit_true",
6703 define_values = {"xnn_enable_sparse": "true"},
6704)
6705
6706# Disables usage of sparse inference.
6707config_setting(
6708 name = "xnn_enable_sparse_explicit_false",
6709 define_values = {"xnn_enable_sparse": "false"},
6710)
6711
Marat Dukhan05702cf2020-03-26 15:41:33 -07006712# Disables usage of HMP-aware optimizations.
6713config_setting(
6714 name = "xnn_enable_hmp_explicit_false",
6715 define_values = {"xnn_enable_hmp": "false"},
6716)
6717
Chao Mei6ddfc602020-05-13 22:29:36 -07006718# Enable usage of optimized memory allocation
6719config_setting(
6720 name = "xnn_enable_memopt_explicit_true",
6721 define_values = {"xnnpack_enable_memopt": "true"},
6722)
6723
6724# Disable usage of optimized memory allocation
6725config_setting(
6726 name = "xnn_enable_memopt_explicit_false",
6727 define_values = {"xnnpack_enable_memopt": "false"},
6728)
6729
Marat Dukhanb8642352019-10-30 15:43:02 -07006730# Builds with -c dbg
6731config_setting(
6732 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006733 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006734 "compilation_mode": "dbg",
6735 },
6736)
6737
6738# Builds with -c opt
6739config_setting(
6740 name = "optimized_build",
6741 values = {
6742 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006743 },
6744)
6745
6746config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006747 name = "linux_k8",
6748 values = {"cpu": "k8"},
6749)
6750
6751config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006752 name = "linux_arm",
6753 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006754)
6755
6756config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006757 name = "linux_armeabi",
6758 values = {"cpu": "armeabi"},
6759)
6760
6761config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006762 name = "linux_armhf",
6763 values = {"cpu": "armhf"},
6764)
6765
6766config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006767 name = "linux_armv7a",
6768 values = {"cpu": "armv7a"},
6769)
6770
6771config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006772 name = "linux_aarch64",
6773 values = {"cpu": "aarch64"},
6774)
6775
6776config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006777 name = "android",
6778 values = {"crosstool_top": "//external:android/crosstool"},
6779)
6780
6781config_setting(
6782 name = "android_armv7",
6783 values = {
6784 "crosstool_top": "//external:android/crosstool",
6785 "cpu": "armeabi-v7a",
6786 },
6787)
6788
6789config_setting(
6790 name = "android_arm64",
6791 values = {
6792 "crosstool_top": "//external:android/crosstool",
6793 "cpu": "arm64-v8a",
6794 },
6795)
6796
6797config_setting(
6798 name = "android_x86",
6799 values = {
6800 "crosstool_top": "//external:android/crosstool",
6801 "cpu": "x86",
6802 },
6803)
6804
6805config_setting(
6806 name = "android_x86_64",
6807 values = {
6808 "crosstool_top": "//external:android/crosstool",
6809 "cpu": "x86_64",
6810 },
6811)
6812
6813config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006814 name = "windows_x86_64",
6815 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006816)
6817
6818config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006819 name = "windows_x86_64_clang",
6820 values = {
6821 "compiler": "clang-cl",
6822 "cpu": "x64_windows",
6823 },
6824)
6825
6826config_setting(
6827 name = "windows_x86_64_mingw",
6828 values = {
6829 "compiler": "mingw-gcc",
6830 "cpu": "x64_windows",
6831 },
6832)
6833
6834config_setting(
6835 name = "windows_x86_64_msys",
6836 values = {
6837 "compiler": "msys-gcc",
6838 "cpu": "x64_windows",
6839 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006840)
6841
6842config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006843 name = "macos_x86_64",
6844 values = {
6845 "apple_platform_type": "macos",
6846 "cpu": "darwin",
6847 },
6848)
6849
6850config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006851 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006852 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006853)
6854
6855config_setting(
6856 name = "emscripten_wasm",
6857 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006858 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006859 "cpu": "wasm",
6860 },
6861)
6862
6863config_setting(
6864 name = "emscripten_wasmsimd",
6865 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006866 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006867 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006868 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006869 },
6870)
6871
6872config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006873 name = "ios_armv7",
6874 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006875 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006876 "cpu": "ios_armv7",
6877 },
6878)
6879
6880config_setting(
6881 name = "ios_arm64",
6882 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006883 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006884 "cpu": "ios_arm64",
6885 },
6886)
6887
6888config_setting(
6889 name = "ios_arm64e",
6890 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006891 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006892 "cpu": "ios_arm64e",
6893 },
6894)
6895
6896config_setting(
6897 name = "ios_x86",
6898 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006899 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006900 "cpu": "ios_i386",
6901 },
6902)
6903
6904config_setting(
6905 name = "ios_x86_64",
6906 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006907 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006908 "cpu": "ios_x86_64",
6909 },
6910)
6911
6912config_setting(
6913 name = "watchos_armv7k",
6914 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006915 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006916 "cpu": "watchos_armv7k",
6917 },
6918)
6919
6920config_setting(
6921 name = "watchos_arm64_32",
6922 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006923 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006924 "cpu": "watchos_arm64_32",
6925 },
6926)
6927
6928config_setting(
6929 name = "watchos_x86",
6930 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006931 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006932 "cpu": "watchos_i386",
6933 },
6934)
6935
6936config_setting(
6937 name = "watchos_x86_64",
6938 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006939 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006940 "cpu": "watchos_x86_64",
6941 },
6942)
6943
6944config_setting(
6945 name = "tvos_arm64",
6946 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006947 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006948 "cpu": "tvos_arm64",
6949 },
6950)
6951
6952config_setting(
6953 name = "tvos_x86_64",
6954 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006955 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006956 "cpu": "tvos_x86_64",
6957 },
6958)