blob: a958d136af6fd318de723db89673ce38d1ddd132 [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",
Frank Barchardbeca6522020-10-30 22:34:35 -0700910 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c",
911 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700912 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700913 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c",
914 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700915 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700916 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c",
917 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700918 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700919 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c",
920 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700921 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700922 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c",
923 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700924 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700925 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c",
926 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700927 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700928 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c",
929 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -0700930 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700931 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c",
932 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -0700933 "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 Dukhan30d4b252020-10-29 16:33:22 -07001244 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1245 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1246 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1247 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1248 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1249 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1250 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1251 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001252 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001253 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001254 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001255 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1256 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001257 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001258 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1259 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001260 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001261 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1262 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1263 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1264 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1265 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001266 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1267 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001268 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1269 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001270 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1271 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001272 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1273 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1274 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1275 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1276 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1277 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1278 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1279 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1280 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1281 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1282 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1283 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1284 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1285 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1286 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1287 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001288 "src/f32-hswish/gen/hswish-neon-x4.c",
1289 "src/f32-hswish/gen/hswish-neon-x8.c",
1290 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001291 "src/f32-ibilinear/gen/neon-c4.c",
1292 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001293 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001294 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001295 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001296 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1297 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001298 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001299 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1300 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1301 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1302 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001303 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1304 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001305 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1306 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001307 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1308 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001309 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1310 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1311 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001312 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1313 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001314 "src/f32-prelu/gen/neon-1x4.c",
1315 "src/f32-prelu/gen/neon-1x8.c",
1316 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001317 "src/f32-prelu/gen/neon-2x4.c",
1318 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001319 "src/f32-prelu/gen/neon-2x16.c",
1320 "src/f32-prelu/gen/neon-4x4.c",
1321 "src/f32-prelu/gen/neon-4x8.c",
1322 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001323 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001324 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001325 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001326 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1327 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001328 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001329 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1330 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001331 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001332 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1333 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001334 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1335 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1336 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1337 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1338 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1339 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1340 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1341 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1342 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1343 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1344 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1345 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1346 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001347 "src/f32-relu/gen/neon-x4.c",
1348 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001349 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001350 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001351 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1352 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1353 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1354 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1355 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1356 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1357 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1358 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1359 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1360 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1361 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1362 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001363 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1364 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1365 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1366 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1367 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1368 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001369 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1370 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1371 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1372 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001373 "src/f32-vbinary/gen/vmax-neon-x4.c",
1374 "src/f32-vbinary/gen/vmax-neon-x8.c",
1375 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1376 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1377 "src/f32-vbinary/gen/vmin-neon-x4.c",
1378 "src/f32-vbinary/gen/vmin-neon-x8.c",
1379 "src/f32-vbinary/gen/vminc-neon-x4.c",
1380 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001381 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1382 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1383 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1384 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1385 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1386 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001387 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1388 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1389 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1390 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001391 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1392 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1393 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1394 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001395 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1396 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001397 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1398 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001399 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1400 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001401 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1402 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1403 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1404 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1405 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1406 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001407 "src/f32-vunary/gen/vabs-neon-x4.c",
1408 "src/f32-vunary/gen/vabs-neon-x8.c",
1409 "src/f32-vunary/gen/vneg-neon-x4.c",
1410 "src/f32-vunary/gen/vneg-neon-x8.c",
1411 "src/f32-vunary/gen/vsqr-neon-x4.c",
1412 "src/f32-vunary/gen/vsqr-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001413 "src/math/roundd-neon-addsub.c",
1414 "src/math/roundd-neon-cvt.c",
1415 "src/math/roundne-neon-addsub.c",
1416 "src/math/roundu-neon-addsub.c",
1417 "src/math/roundu-neon-cvt.c",
1418 "src/math/roundz-neon-addsub.c",
1419 "src/math/roundz-neon-cvt.c",
1420 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
1421 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1422 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1423 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1424 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1425 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1426 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1427 "src/math/sqrt-neon-nr1rsqrts.c",
1428 "src/math/sqrt-neon-nr2rsqrts.c",
1429 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001430 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1431 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1432 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1433 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001434 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1435 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1436 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1437 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001438 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1439 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1440 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1441 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001442 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001443 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001444 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001445 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1446 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001447 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001448 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001449 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001450 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001451 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001452 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001453 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1454 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1455 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1456 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1457 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1458 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1459 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1460 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001461 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1462 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1463 "src/qu8-dwconv/up8x9-minmax-neon.c",
1464 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1465 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1466 "src/qu8-gemm/4x8-minmax-neon.c",
1467 "src/qu8-gemm/8x8-minmax-neon.c",
1468 "src/qu8-igemm/4x8-minmax-neon.c",
1469 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001470 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001471 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001472 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001473 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001474 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001475 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001476 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001477 "src/x8-zip/x2-neon.c",
1478 "src/x8-zip/x3-neon.c",
1479 "src/x8-zip/x4-neon.c",
1480 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001481 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001482 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001483 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001484 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001485 "src/x32-zip/x2-neon.c",
1486 "src/x32-zip/x3-neon.c",
1487 "src/x32-zip/x4-neon.c",
1488 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001489]
1490
1491NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001492 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1493 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1494 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1495 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1496 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1497 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1498 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1499 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1500 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1501 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1502 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1503 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1504 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1505 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1506 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1507 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1508 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1509 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1510 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1511 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1512 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1513 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1514 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1515 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1516 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1517 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1518 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1519 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1520 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1521 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001522 "src/f32-ibilinear/gen/neonfma-c4.c",
1523 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001524 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001525 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001526 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001527 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1528 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001529 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1530 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001531 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1532 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001533 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1534 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001535 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001536 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001537 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001538 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1539 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001540 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001541 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1542 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001543 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001544 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1545 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001546 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1547 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1548 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1549 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1550 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1551 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1552 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1553 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1554 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1555 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1556 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1557 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1558 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001559 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1560 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1561 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1562 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1563 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1564 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001565 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1566 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1567 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1568 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1569 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1570 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001571 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1572 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1573 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1574 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1575 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1576 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001577 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1578 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1579 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1580 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1581 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1582 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001583 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1584 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1585 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1586 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1587 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1588 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001589 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1590 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1591 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1592 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1593 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1594 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001595 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1596 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1597 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1598 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1599 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1600 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1601 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1602 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1603 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1604 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1605 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1606 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1607 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1608 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1609 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1610 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1611 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1612 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001613 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1614 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001615 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1616 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1617 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1618 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1619 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1620 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1621 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1622 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1623 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1624 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1625 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1626 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1627 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1628 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1629 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1630 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1631 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1632 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1633 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1634 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001635 "src/math/exp-neonfma-lut64-p2.c",
1636 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001637 "src/math/expminus-neonfma-lut64-p2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001638 "src/math/expminus-neonfma-lut2048-p1.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001639 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001640 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1641 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1642 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001643 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1644 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1645 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001646 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1647 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1648 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001649 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1650 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1651 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001652 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1653 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1654 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001655 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1656 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1657 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001658 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001659 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001660 "src/math/sqrt-neonfma-nr2fma.c",
1661 "src/math/sqrt-neonfma-nr2fma1adj.c",
1662 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001663]
1664
1665AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001666 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001667 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001668 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001669 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001670 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001671 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001672 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001673 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001674 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001675 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
1676 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
1677 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001678 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001679 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001680 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
1681 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
1682 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
1683 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
1684 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001685 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
1686 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
1687 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001688 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001689 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001690 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
1691 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
1692 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001693 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
1694 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
1695 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
1696 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001697 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001698 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
1699 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001700 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001701 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001702 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001703 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001704 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
1705 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001706 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
1707 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
1708 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
1709 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
1710 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
1711 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
1712 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
1713 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001714 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001715 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001716 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1717 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1718 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1719 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1720 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1721 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1722 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1723 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1724 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1725 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1726 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1727 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1728 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1729 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1730 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1731 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1732 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1733 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1734 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1735 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001736 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1737 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1738 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1739 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1740 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1741 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1742 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1743 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1744 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1745 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1746 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1747 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001748 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1749 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1750 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1751 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1752 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1753 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001754 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001755 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001756 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1757 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1758 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1759 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001760 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001761 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1762 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1763 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001764 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1765 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1766 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1767 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001768 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001769 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1770 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1771 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001772 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001773 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001774 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
1775 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
1776 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001777 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1778 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1779 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1780 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1781 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1782 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001783 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1784 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001785 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001786 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001787 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001788 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001789 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001790 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001791]
1792
Marat Dukhan8853b822020-05-07 12:19:01 -07001793NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001794 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1795 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001796 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1797 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1798 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1799 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1800 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1801 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001802 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001803 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001804 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001805 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001806]
1807
Marat Dukhan08c4a432019-10-03 09:29:21 -07001808AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001809 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001810 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001811 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1812 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1813 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1814 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001815 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1816 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1817 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1818 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1819 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1820 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1821 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1822 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001823 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1824 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001825 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1826 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1827 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1828 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1829 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1830 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1831 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
1832 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1833 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1834 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1835 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1836 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1837 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1838 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1839 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1840 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001841 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001842 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1843 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1844 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1845 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1846 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1847 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1848 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1849 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1850 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001851 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001852 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001853 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001854 "src/f16-relu/gen/neonfp16arith-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001855 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001856 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001857 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001858 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001859 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001860 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001861 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001862 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1863 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1864 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1865 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1866 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1867 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1868 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1869 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1870 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1871 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1872 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1873 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1874 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1875 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1876 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1877 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1878 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1879 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1880 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1881 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1882 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1883 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1884 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1885 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1886 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1887 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
1888 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1889 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1890 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1891 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
1892 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001893]
1894
Benoit Jacoba9644732020-08-13 12:48:55 -07001895NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07001896 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001897 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001898 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001899 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001900 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001901 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001902 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001903 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001904 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001905 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001906 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001907 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001908 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001909 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001910 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001911 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001912 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001913 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001914]
1915
Marat Dukhan08c4a432019-10-03 09:29:21 -07001916SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001917 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1918 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001919 "src/f32-clamp/gen/sse-x4.c",
1920 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001921 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1922 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001923 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1924 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1925 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1926 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001927 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1928 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001929 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1930 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1931 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1932 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001933 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1934 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001935 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
1936 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
1937 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07001938 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001939 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07001940 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
1941 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
1942 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
1943 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
1944 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07001945 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
1946 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
1947 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001948 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07001949 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001950 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
1951 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
1952 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07001953 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
1954 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
1955 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
1956 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
1957 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
1958 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
1959 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
1960 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
1961 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
1962 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
1963 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
1964 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
1965 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07001966 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
1967 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
1968 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
1969 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
1970 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
1971 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
1972 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
1973 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
1974 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
1975 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001976 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001977 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1978 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001979 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1980 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1981 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1982 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1983 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1984 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001985 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1986 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1987 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1988 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1989 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1990 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1991 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001992 "src/f32-hswish/gen/hswish-sse-x4.c",
1993 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001994 "src/f32-ibilinear/gen/sse-c4.c",
1995 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001996 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1997 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1998 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1999 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2000 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2001 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2002 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002003 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2004 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2005 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002006 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002007 "src/f32-prelu/gen/sse-2x4.c",
2008 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002009 "src/f32-relu/gen/sse-x4.c",
2010 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002011 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002012 "src/f32-spmm/gen/4x1-minmax-sse.c",
2013 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002014 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002015 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002016 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2017 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2018 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2019 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2020 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2021 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2022 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2023 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002024 "src/f32-vbinary/gen/vmax-sse-x4.c",
2025 "src/f32-vbinary/gen/vmax-sse-x8.c",
2026 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2027 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2028 "src/f32-vbinary/gen/vmin-sse-x4.c",
2029 "src/f32-vbinary/gen/vmin-sse-x8.c",
2030 "src/f32-vbinary/gen/vminc-sse-x4.c",
2031 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002032 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2033 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2034 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2035 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2036 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2037 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2038 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2039 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002040 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2041 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2042 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2043 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002044 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2045 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2046 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2047 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002048 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2049 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002050 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2051 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002052 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2053 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002054 "src/f32-vunary/gen/vabs-sse-x4.c",
2055 "src/f32-vunary/gen/vabs-sse-x8.c",
2056 "src/f32-vunary/gen/vneg-sse-x4.c",
2057 "src/f32-vunary/gen/vneg-sse-x8.c",
2058 "src/f32-vunary/gen/vsqr-sse-x4.c",
2059 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002060 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002061 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002062 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002063 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002064 "src/math/sqrt-sse-hh1mac.c",
2065 "src/math/sqrt-sse-nr1mac.c",
2066 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002067 "src/x32-fill/sse.c",
2068 "src/x32-packx/x4-sse.c",
2069 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002070]
2071
2072SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002073 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002074 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002075 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002076 "src/f32-prelu/gen/sse2-2x4.c",
2077 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002078 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002079 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002080 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002081 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2082 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002083 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002084 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2085 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002086 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002087 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2088 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002089 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002090 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
2091 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
2092 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
2093 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
2094 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
2095 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002096 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002097 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002098 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002099 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002100 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
2101 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002102 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2103 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002104 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2105 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002106 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2107 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2108 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2109 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2110 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2111 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
2112 "src/math/exp-sse2-lut64-p2.c",
2113 "src/math/exp-sse2-p5.c",
2114 "src/math/expminus-sse2-p5.c",
2115 "src/math/roundd-sse2-cvt.c",
2116 "src/math/roundne-sse2-cvt.c",
2117 "src/math/roundu-sse2-cvt.c",
2118 "src/math/roundz-sse2-cvt.c",
2119 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2120 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2121 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2122 "src/math/sigmoid-sse2-rr2-p5-div.c",
2123 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2124 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002125 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2126 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2127 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002128 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2129 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2130 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002131 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2132 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2133 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002134 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002135 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002136 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002137 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002138 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002139 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002140 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2141 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002142 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002143 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2144 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002145 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002146 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2147 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2148 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002149 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002150 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002151 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002152 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002153 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002154 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002155 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002156 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002157 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2158 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002159 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002160 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002161 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002162 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2163 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2164 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2165 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002166 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2167 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2168 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2169 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002170 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2171 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002172 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2173 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2174 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2175 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2176 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002177 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002178 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002179 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002180 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002181 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002182 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002183 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002184 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002185 "src/x8-zip/x2-sse2.c",
2186 "src/x8-zip/x3-sse2.c",
2187 "src/x8-zip/x4-sse2.c",
2188 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002189 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002190 "src/x32-zip/x2-sse2.c",
2191 "src/x32-zip/x3-sse2.c",
2192 "src/x32-zip/x4-sse2.c",
2193 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002194]
2195
2196SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002197 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2198 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2199 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002200 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002201 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002202 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2203 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2204 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2205 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2206 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002207 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2208 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2209 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002210 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2211 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2212 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002213 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2214 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2215 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002216 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002217 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002218 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002219 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002220 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002221 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002222 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2223 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002224 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002225 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2226 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002227 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002228 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2229 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2230 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002231 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002232 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002233 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002234 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002235 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002236 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002237 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002238 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002239 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2240 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002241 "src/qs8-requantization/precise-ssse3.c",
2242 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002243 "src/qu8-requantization/precise-ssse3.c",
2244 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002245]
2246
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002247SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002248 "src/f32-prelu/gen/sse41-2x4.c",
2249 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002250 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2251 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2252 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2253 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2254 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2255 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002256 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002257 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002258 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002259 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002260 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2261 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002262 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2263 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002264 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2265 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002266 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2267 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2268 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2269 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2270 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2271 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2272 "src/math/roundd-sse41.c",
2273 "src/math/roundne-sse41.c",
2274 "src/math/roundu-sse41.c",
2275 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002276 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2277 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2278 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002279 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2280 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2281 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002282 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2283 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2284 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002285 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002286 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002287 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002288 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002289 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002290 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002291 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2292 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002293 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002294 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2295 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002296 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002297 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2298 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2299 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002300 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002301 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002302 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002303 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002304 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002305 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002306 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002307 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002308 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2309 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002310 "src/qs8-requantization/fp32-sse4.c",
2311 "src/qs8-requantization/precise-sse4.c",
2312 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002313 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2314 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2315 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2316 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002317 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2318 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2319 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2320 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002321 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2322 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2323 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2324 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002325 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2326 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2327 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2328 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002329 "src/qu8-requantization/precise-sse4.c",
2330 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002331]
2332
Marat Dukhan08c4a432019-10-03 09:29:21 -07002333AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002334 "src/f32-clamp/gen/avx-x8.c",
2335 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002336 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2337 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002338 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2339 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002340 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2341 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002342 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2343 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2344 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2345 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2346 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2347 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002348 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002349 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2350 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002351 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002352 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002353 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002354 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002355 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2356 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2357 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2358 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2359 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2360 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2361 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2362 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2363 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2364 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2365 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002366 "src/f32-hswish/gen/hswish-avx-x8.c",
2367 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002368 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002369 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2370 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002371 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002372 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002373 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002374 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002375 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2376 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002377 "src/f32-prelu/gen/avx-2x8.c",
2378 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002379 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002380 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002381 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002382 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2383 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2384 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2385 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2386 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2387 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2388 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2389 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2390 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2391 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2392 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2393 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2394 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2395 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2396 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2397 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2398 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2399 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2400 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2401 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002402 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2403 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2404 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2405 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2406 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2407 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2408 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2409 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002410 "src/f32-vbinary/gen/vmax-avx-x8.c",
2411 "src/f32-vbinary/gen/vmax-avx-x16.c",
2412 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2413 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2414 "src/f32-vbinary/gen/vmin-avx-x8.c",
2415 "src/f32-vbinary/gen/vmin-avx-x16.c",
2416 "src/f32-vbinary/gen/vminc-avx-x8.c",
2417 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002418 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2419 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2420 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2421 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2422 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2423 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2424 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2425 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002426 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2427 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2428 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2429 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002430 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2431 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2432 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2433 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002434 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2435 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002436 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2437 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002438 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2439 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2440 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2441 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2442 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2443 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002444 "src/f32-vscale/avx-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002445 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2446 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002447 "src/f32-vunary/gen/vabs-avx-x8.c",
2448 "src/f32-vunary/gen/vabs-avx-x16.c",
2449 "src/f32-vunary/gen/vneg-avx-x8.c",
2450 "src/f32-vunary/gen/vneg-avx-x16.c",
2451 "src/f32-vunary/gen/vsqr-avx-x8.c",
2452 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhand08abcb2020-09-21 00:26:25 -07002453 "src/math/exp-avx-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002454 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002455 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002456 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2457 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002458]
2459
Marat Dukhan1566fee2020-08-02 21:55:41 -07002460XOP_UKERNELS = [
2461 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002462 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002463 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002464 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002465 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002466 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002467 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2468 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002469 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002470 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2471 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002472 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002473 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2474 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2475 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002476 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002477 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002478 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002479 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002480 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002481 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002482 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002483 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002484 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2485 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002486 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2487 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2488 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2489 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2490 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2491 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2492 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2493 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002494]
2495
Marat Dukhanfda12b82019-11-21 12:27:59 -08002496FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002497 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2498 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002499 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2500 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002501 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2502 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002503 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2504 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2505 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2506 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2507 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2508 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002509 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002510 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2511 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2512 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2513 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002514 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002515 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2516 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002517 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002518 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2519 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002520 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2521 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2522 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002523 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2524 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2525 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2526 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2527 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2528 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2529 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2530 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2531 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2532 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2533 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2534 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2535 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2536 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002537 "src/f32-hswish/gen/hswish-fma3-x8.c",
2538 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002539 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002540 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2541 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2542 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2543 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002544 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002545 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2546 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002547 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002548 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2549 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002550 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2551 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2552 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002553 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2554 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2555 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2556 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2557 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2558 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2559 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2560 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002561 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002562 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002563 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002564]
2565
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002566AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002567 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2568 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002569 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002570 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002571 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002572 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2573 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002574 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002575 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2576 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2577 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002578 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002579 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2580 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002581 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002582 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002583 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002584 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2585 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002586 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002587 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2588 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2589 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002590 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002591 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2592 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002593 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002594 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002595 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002596 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2597 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002598 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002599 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2600 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2601 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002602 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002603 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2604 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2605 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2606 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2607 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2608 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2609 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2610 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2611 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2612 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2613 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2614 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2615 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2616 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2617 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2618 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2619 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2620 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2621 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2622 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2623 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2624 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2625 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2626 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2627 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2628 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2629 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2630 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2631 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2632 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002633 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2634 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2635 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2636 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2637 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2638 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2639 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2640 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2641 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2642 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2643 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2644 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2645 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2646 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2647 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2648 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2649 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2650 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2651 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2652 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2653 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2654 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2655 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2656 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002657 "src/math/exp-avx2-lut8-p3-perm.c",
2658 "src/math/exp-avx2-lut8-p4-perm.c",
2659 "src/math/exp-avx2-p5.c",
2660 "src/math/expminus-avx2-p5.c",
2661 "src/math/extexp-avx2-p5.c",
2662 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2663 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2664 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2665 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2666 "src/math/sigmoid-avx2-rr1-p5-div.c",
2667 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
2668 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2669 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2670 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2671 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2672 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2673 "src/math/sigmoid-avx2-rr2-p5-div.c",
2674 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2675 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002676 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002677 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002678 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2679 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002680 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002681 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002682 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002683 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002684 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002685 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002686 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002687 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002688 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2689 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2690 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002691 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2692 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2693 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2694 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2695 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2696 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2697 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2698 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002699]
2700
Marat Dukhan08c4a432019-10-03 09:29:21 -07002701AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002702 "src/f32-clamp/gen/avx512f-x16.c",
2703 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002704 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2705 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002706 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2707 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002708 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2709 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002710 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2711 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2712 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2713 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2714 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2715 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002716 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2717 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2718 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2719 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2720 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2721 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002722 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2723 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2724 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2725 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2726 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2727 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002728 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2729 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002730 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2731 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2732 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2733 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2734 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2735 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002736 "src/f32-prelu/gen/avx512f-2x16.c",
2737 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002738 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2739 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002740 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002741 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002742 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002743 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2744 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002745 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002746 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2747 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2748 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002749 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002750 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2751 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002752 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002753 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002754 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002755 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2756 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002757 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002758 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2759 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2760 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002761 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002762 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2763 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002764 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002765 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002766 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002767 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2768 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002769 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002770 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2771 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2772 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002773 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002774 "src/f32-relu/gen/avx512f-x16.c",
2775 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002776 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002777 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2778 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2779 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2780 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2781 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2782 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2783 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2784 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2785 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2786 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2787 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2788 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2789 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2790 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2791 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2792 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002793 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2794 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2795 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2796 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2797 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2798 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2799 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2800 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2801 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2802 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2803 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2804 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2805 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2806 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2807 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2808 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002809 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2810 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2811 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2812 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2813 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2814 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2815 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2816 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2817 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2818 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2819 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2820 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2821 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2822 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2823 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2824 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002825 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2826 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2827 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2828 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2829 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2830 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2831 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2832 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002833 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2834 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2835 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2836 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2837 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2838 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2839 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2840 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002841 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2842 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2843 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2844 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2845 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2846 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2847 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2848 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002849 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2850 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2851 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2852 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002853 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2854 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2855 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2856 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002857 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2858 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002859 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2860 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
2861 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2862 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2863 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2864 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2865 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2866 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002867 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002868 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2869 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2870 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2871 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2872 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2873 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2874 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2875 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2876 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2877 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2878 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2879 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2880 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2881 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2882 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2883 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2884 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2885 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2886 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2887 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2888 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2889 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2890 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2891 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002892 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2893 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2894 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2895 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2896 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2897 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2898 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2899 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002900 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2901 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2902 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2903 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2904 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2905 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002906 "src/math/exp-avx512f-lut16-p3-perm-scalef.c",
2907 "src/math/exp-avx512f-lut16-p3-perm.c",
2908 "src/math/exp-avx512f-lut32-p2-perm2-scalef.c",
2909 "src/math/exp-avx512f-lut32-p2-perm2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002910 "src/math/exp-avx512f-p5-scalef.c",
2911 "src/math/exp-avx512f-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002912 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002913 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002914 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002915 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002916 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002917 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002918 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002919 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002920 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002921 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
2922 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
2923 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
2924 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
2925 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2926 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2927 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2928 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2929 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
2930 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002931 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002932 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002933 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
2934 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
2935 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
2936 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002937 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002938 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002939 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002940]
2941
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002942AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002943 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2944 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002945 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2946 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2947 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2948 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2949 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2950 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2951 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2952 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2953]
2954
Frank Barchardbcedc082020-08-17 18:00:51 -07002955WASM32_ASM_UKERNELS = [
2956 "src/f32-relu/wasm_shr_x1.S",
2957 "src/f32-relu/wasm_shr_x2.S",
2958 "src/f32-relu/wasm_shr_x4.S",
2959]
2960
Marat Dukhan08c4a432019-10-03 09:29:21 -07002961AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002962 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002963 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002964 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2965 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002966 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002967 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002968 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002969 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002970 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2971 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002972 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2973 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2974 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2975 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002976]
2977
2978AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002979 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002980 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002981 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002982 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002983 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002984 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002985 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002986 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2987 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
2988 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2989 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2990 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
2991 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
2992 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002993 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2994 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002995 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2996 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2997 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002998 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
2999 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003000 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3001 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3002 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3003 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003004 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003005 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3006 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003007 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3008 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3009 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3010 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003011 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003012 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003013 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003014 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003015 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3016 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3017 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3018 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3019 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3020 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3021 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3022 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3023 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3024 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3025 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3026 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3027 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3028 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3029 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3030 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3031 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3032 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3033 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3034 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3035 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3036 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003037 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003038 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003039 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3040 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003041 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3042 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3043 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3044 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3045 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3046 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003047 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3048 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3049 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3050 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003051 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3052 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003053 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07003054 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07003055 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003056]
3057
Marat Dukhan1b354632020-03-23 12:50:22 -07003058INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003059 "src/xnnpack/argmaxpool.h",
3060 "src/xnnpack/avgpool.h",
3061 "src/xnnpack/clamp.h",
3062 "src/xnnpack/common.h",
3063 "src/xnnpack/conv.h",
3064 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003065 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003066 "src/xnnpack/gavgpool.h",
3067 "src/xnnpack/gemm.h",
3068 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003069 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003070 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003071 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003072 "src/xnnpack/lut.h",
3073 "src/xnnpack/math.h",
3074 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003075 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003076 "src/xnnpack/packx.h",
3077 "src/xnnpack/pad.h",
3078 "src/xnnpack/params.h",
3079 "src/xnnpack/pavgpool.h",
3080 "src/xnnpack/ppmm.h",
3081 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003082 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003083 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003084 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003085 "src/xnnpack/rmax.h",
3086 "src/xnnpack/scalar-utils.h",
3087 "src/xnnpack/spmm.h",
3088 "src/xnnpack/unpool.h",
3089 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003090 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003091 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003092 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003093 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003094 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003095 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003096 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003097]
3098
3099INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003100 "include/xnnpack.h",
3101 "src/xnnpack/allocator.h",
3102 "src/xnnpack/compute.h",
3103 "src/xnnpack/im2col.h",
3104 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003105 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003106 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003107 "src/xnnpack/operator.h",
3108 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003109 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003110 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003111 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003112 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003113]
3114
Marat Dukhan1b354632020-03-23 12:50:22 -07003115ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003116 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003117]
3118
Marat Dukhan1b354632020-03-23 12:50:22 -07003119MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003120 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003121 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003122]
3123
Marat Dukhan1b354632020-03-23 12:50:22 -07003124MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003125 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003126 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003127 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003128 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003129]
3130
3131OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003132 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003133 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003134]
3135
3136WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003137 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003138 "src/xnnpack/operator.h",
3139 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003140]
3141
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003142LOGGING_COPTS = select({
3143 # No logging in optimized mode
3144 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3145 # Full logging in debug mode
3146 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3147 # Error-only logging in default (fastbuild) mode
3148 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3149})
3150
Marat Dukhan3b59de22020-06-03 20:15:19 -07003151LOGGING_SRCS = select({
3152 # No logging in optimized mode
3153 ":optimized_build": [],
3154 "//conditions:default": [
3155 "src/operator-strings.c",
3156 "src/subgraph-strings.c",
3157 ],
3158})
3159
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003160LOGGING_HDRS = [
3161 "src/xnnpack/log.h",
3162]
3163
Marat Dukhan08c4a432019-10-03 09:29:21 -07003164xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003165 name = "tables",
3166 srcs = TABLE_SRCS,
3167 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003168 gcc_copts = xnnpack_gcc_std_copts(),
3169 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003170)
3171
3172xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003173 name = "scalar_ukernels",
3174 srcs = SCALAR_UKERNELS,
3175 hdrs = INTERNAL_HDRS,
3176 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003177 gcc_copts = xnnpack_gcc_std_copts(),
3178 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003179 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003180 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003181 "@FP16",
3182 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003183 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003184 ],
3185)
3186
3187xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003188 name = "scalar_ukernels_test_mode",
3189 srcs = SCALAR_UKERNELS,
3190 hdrs = INTERNAL_HDRS,
3191 aarch32_copts = ["-marm"],
3192 copts = [
3193 "-UNDEBUG",
3194 "-DXNN_TEST_MODE=1",
3195 ],
3196 gcc_copts = xnnpack_gcc_std_copts(),
3197 msvc_copts = xnnpack_msvc_std_copts(),
3198 deps = [
3199 ":tables",
3200 "@FP16",
3201 "@FXdiv",
3202 "@pthreadpool",
3203 ],
3204)
3205
3206xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003207 name = "wasm_ukernels",
3208 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003209 gcc_copts = xnnpack_gcc_std_copts(),
3210 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003211 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003212 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003213 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003214 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003215 "@FP16",
3216 "@FXdiv",
3217 "@pthreadpool",
3218 ],
3219)
3220
3221xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003222 name = "wasm_ukernels_test_mode",
3223 hdrs = INTERNAL_HDRS,
3224 copts = [
3225 "-UNDEBUG",
3226 "-DXNN_TEST_MODE=1",
3227 ],
3228 gcc_copts = xnnpack_gcc_std_copts(),
3229 msvc_copts = xnnpack_msvc_std_copts(),
3230 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003231 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003232 deps = [
3233 ":tables",
3234 "@FP16",
3235 "@FXdiv",
3236 "@pthreadpool",
3237 ],
3238)
3239
3240xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003241 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003242 hdrs = INTERNAL_HDRS,
3243 aarch32_copts = [
3244 "-marm",
3245 "-mfpu=neon",
3246 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003247 gcc_copts = xnnpack_gcc_std_copts(),
3248 gcc_x86_copts = ["-msse2"],
3249 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003250 optimized_copts = [
3251 "-O3",
3252 "-ffast-math",
3253 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003254 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003255 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003256 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003257 "@FP16",
3258 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003259 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003260 ],
3261)
3262
3263xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003264 name = "psimd_fastmath_ukernels_test_mode",
3265 hdrs = INTERNAL_HDRS,
3266 aarch32_copts = [
3267 "-marm",
3268 "-mfpu=neon",
3269 ],
3270 copts = [
3271 "-UNDEBUG",
3272 "-DXNN_TEST_MODE=1",
3273 ],
3274 gcc_copts = xnnpack_gcc_std_copts(),
3275 gcc_x86_copts = ["-msse2"],
3276 msvc_copts = xnnpack_msvc_std_copts(),
3277 optimized_copts = [
3278 "-O3",
3279 "-ffast-math",
3280 ],
3281 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
3282 deps = [
3283 ":tables",
3284 "@FP16",
3285 "@psimd",
3286 "@pthreadpool",
3287 ],
3288)
3289
3290xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003291 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003292 hdrs = INTERNAL_HDRS,
3293 aarch32_copts = [
3294 "-marm",
3295 "-mfpu=neon",
3296 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003297 gcc_copts = xnnpack_gcc_std_copts(),
3298 gcc_x86_copts = ["-msse2"],
3299 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003300 optimized_copts = [
3301 "-O3",
3302 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003303 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003304 deps = [
3305 ":tables",
3306 "@FP16",
3307 "@psimd",
3308 "@pthreadpool",
3309 ],
3310)
3311
3312xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003313 name = "psimd_accmath_ukernels_test_mode",
3314 hdrs = INTERNAL_HDRS,
3315 aarch32_copts = [
3316 "-marm",
3317 "-mfpu=neon",
3318 ],
3319 copts = [
3320 "-UNDEBUG",
3321 "-DXNN_TEST_MODE=1",
3322 ],
3323 gcc_copts = xnnpack_gcc_std_copts(),
3324 gcc_x86_copts = ["-msse2"],
3325 msvc_copts = xnnpack_msvc_std_copts(),
3326 optimized_copts = [
3327 "-O3",
3328 ],
3329 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3330 deps = [
3331 ":tables",
3332 "@FP16",
3333 "@psimd",
3334 "@pthreadpool",
3335 ],
3336)
3337
3338xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003339 name = "neon_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",
3345 ],
3346 aarch32_srcs = NEON_UKERNELS,
3347 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003348 gcc_copts = xnnpack_gcc_std_copts(),
3349 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003350 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003351 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003352 "@FP16",
3353 "@pthreadpool",
3354 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003355)
3356
3357xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003358 name = "neon_ukernels_test_mode",
3359 hdrs = INTERNAL_HDRS,
3360 aarch32_copts = [
3361 "-marm",
3362 "-march=armv7-a",
3363 "-mfpu=neon",
3364 ],
3365 aarch32_srcs = NEON_UKERNELS,
3366 aarch64_srcs = NEON_UKERNELS,
3367 copts = [
3368 "-UNDEBUG",
3369 "-DXNN_TEST_MODE=1",
3370 ],
3371 gcc_copts = xnnpack_gcc_std_copts(),
3372 msvc_copts = xnnpack_msvc_std_copts(),
3373 deps = [
3374 ":tables",
3375 "@FP16",
3376 "@pthreadpool",
3377 ],
3378)
3379
3380xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003381 name = "neonfma_ukernels",
3382 hdrs = INTERNAL_HDRS,
3383 aarch32_copts = [
3384 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003385 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003386 "-mfpu=neon-vfpv4",
3387 ],
3388 aarch32_srcs = NEONFMA_UKERNELS,
3389 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003390 apple_aarch32_copts = [
3391 "-mcpu=swift",
3392 "-mtune=generic",
3393 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003394 gcc_copts = xnnpack_gcc_std_copts(),
3395 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003396 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003397 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003398 "@FP16",
3399 "@pthreadpool",
3400 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003401)
3402
3403xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003404 name = "neonfma_ukernels_test_mode",
3405 hdrs = INTERNAL_HDRS,
3406 aarch32_copts = [
3407 "-marm",
3408 "-march=armv7-a",
3409 "-mfpu=neon-vfpv4",
3410 ],
3411 aarch32_srcs = NEONFMA_UKERNELS,
3412 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003413 apple_aarch32_copts = [
3414 "-mcpu=swift",
3415 "-mtune=generic",
3416 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003417 copts = [
3418 "-UNDEBUG",
3419 "-DXNN_TEST_MODE=1",
3420 ],
3421 gcc_copts = xnnpack_gcc_std_copts(),
3422 msvc_copts = xnnpack_msvc_std_copts(),
3423 deps = [
3424 ":tables",
3425 "@FP16",
3426 "@pthreadpool",
3427 ],
3428)
3429
3430xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003431 name = "neonv8_ukernels",
3432 hdrs = INTERNAL_HDRS,
3433 aarch32_copts = [
3434 "-marm",
3435 "-march=armv8-a",
3436 "-mfpu=neon-fp-armv8",
3437 ],
3438 aarch32_srcs = NEONV8_UKERNELS,
3439 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003440 apple_aarch32_copts = [
3441 "-mcpu=cyclone",
3442 "-mtune=generic",
3443 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003444 gcc_copts = xnnpack_gcc_std_copts(),
3445 msvc_copts = xnnpack_msvc_std_copts(),
3446 deps = [
3447 ":tables",
3448 "@FP16",
3449 "@pthreadpool",
3450 ],
3451)
3452
3453xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003454 name = "neonv8_ukernels_test_mode",
3455 hdrs = INTERNAL_HDRS,
3456 aarch32_copts = [
3457 "-marm",
3458 "-march=armv8-a",
3459 "-mfpu=neon-fp-armv8",
3460 ],
3461 aarch32_srcs = NEONV8_UKERNELS,
3462 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003463 apple_aarch32_copts = [
3464 "-mcpu=cyclone",
3465 "-mtune=generic",
3466 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003467 copts = [
3468 "-UNDEBUG",
3469 "-DXNN_TEST_MODE=1",
3470 ],
3471 gcc_copts = xnnpack_gcc_std_copts(),
3472 msvc_copts = xnnpack_msvc_std_copts(),
3473 deps = [
3474 ":tables",
3475 "@FP16",
3476 "@pthreadpool",
3477 ],
3478)
3479
3480xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003481 name = "neonfp16arith_ukernels",
3482 hdrs = INTERNAL_HDRS,
3483 aarch64_copts = ["-march=armv8.2-a+fp16"],
3484 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003485 gcc_copts = xnnpack_gcc_std_copts(),
3486 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003487 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003488 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003489 "@FP16",
3490 "@pthreadpool",
3491 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003492)
3493
3494xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003495 name = "neonfp16arith_ukernels_test_mode",
3496 hdrs = INTERNAL_HDRS,
3497 aarch64_copts = ["-march=armv8.2-a+fp16"],
3498 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3499 copts = [
3500 "-UNDEBUG",
3501 "-DXNN_TEST_MODE=1",
3502 ],
3503 gcc_copts = xnnpack_gcc_std_copts(),
3504 msvc_copts = xnnpack_msvc_std_copts(),
3505 deps = [
3506 ":tables",
3507 "@FP16",
3508 "@pthreadpool",
3509 ],
3510)
3511
3512xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003513 name = "neondot_ukernels",
3514 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003515 aarch32_copts = [
3516 "-marm",
3517 "-march=armv8.2-a+dotprod",
3518 "-mfpu=neon-fp-armv8",
3519 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003520 aarch32_srcs = NEONDOT_UKERNELS,
3521 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3522 aarch64_srcs = NEONDOT_UKERNELS,
3523 gcc_copts = xnnpack_gcc_std_copts(),
3524 msvc_copts = xnnpack_msvc_std_copts(),
3525 deps = [
3526 ":tables",
3527 "@FP16",
3528 "@pthreadpool",
3529 ],
3530)
3531
3532xnnpack_cc_library(
3533 name = "neondot_ukernels_test_mode",
3534 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003535 aarch32_copts = [
3536 "-marm",
3537 "-march=armv8.2-a+dotprod",
3538 "-mfpu=neon-fp-armv8",
3539 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003540 aarch32_srcs = NEONDOT_UKERNELS,
3541 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3542 aarch64_srcs = NEONDOT_UKERNELS,
3543 copts = [
3544 "-UNDEBUG",
3545 "-DXNN_TEST_MODE=1",
3546 ],
3547 gcc_copts = xnnpack_gcc_std_copts(),
3548 msvc_copts = xnnpack_msvc_std_copts(),
3549 deps = [
3550 ":tables",
3551 "@FP16",
3552 "@pthreadpool",
3553 ],
3554)
3555
3556xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003557 name = "sse2_ukernels",
3558 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003559 gcc_copts = xnnpack_gcc_std_copts(),
3560 gcc_x86_copts = ["-msse2"],
3561 msvc_copts = xnnpack_msvc_std_copts(),
3562 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003563 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003564 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003565 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003566 "@FP16",
3567 "@pthreadpool",
3568 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003569)
3570
3571xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003572 name = "sse2_ukernels_test_mode",
3573 hdrs = INTERNAL_HDRS,
3574 copts = [
3575 "-UNDEBUG",
3576 "-DXNN_TEST_MODE=1",
3577 ],
3578 gcc_copts = xnnpack_gcc_std_copts(),
3579 gcc_x86_copts = ["-msse2"],
3580 msvc_copts = xnnpack_msvc_std_copts(),
3581 msvc_x86_32_copts = ["/arch:SSE2"],
3582 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3583 deps = [
3584 ":tables",
3585 "@FP16",
3586 "@pthreadpool",
3587 ],
3588)
3589
3590xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003591 name = "ssse3_ukernels",
3592 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003593 gcc_copts = xnnpack_gcc_std_copts(),
3594 gcc_x86_copts = ["-mssse3"],
3595 msvc_copts = xnnpack_msvc_std_copts(),
3596 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003597 x86_srcs = SSSE3_UKERNELS,
3598 deps = [
3599 ":tables",
3600 "@FP16",
3601 "@pthreadpool",
3602 ],
3603)
3604
3605xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003606 name = "ssse3_ukernels_test_mode",
3607 hdrs = INTERNAL_HDRS,
3608 copts = [
3609 "-UNDEBUG",
3610 "-DXNN_TEST_MODE=1",
3611 ],
3612 gcc_copts = xnnpack_gcc_std_copts(),
3613 gcc_x86_copts = ["-mssse3"],
3614 msvc_copts = xnnpack_msvc_std_copts(),
3615 msvc_x86_32_copts = ["/arch:SSE2"],
3616 x86_srcs = SSSE3_UKERNELS,
3617 deps = [
3618 ":tables",
3619 "@FP16",
3620 "@pthreadpool",
3621 ],
3622)
3623
3624xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003625 name = "sse41_ukernels",
3626 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003627 gcc_copts = xnnpack_gcc_std_copts(),
3628 gcc_x86_copts = ["-msse4.1"],
3629 msvc_copts = xnnpack_msvc_std_copts(),
3630 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003631 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003632 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003633 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003634 "@FP16",
3635 "@pthreadpool",
3636 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003637)
3638
3639xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003640 name = "sse41_ukernels_test_mode",
3641 hdrs = INTERNAL_HDRS,
3642 copts = [
3643 "-UNDEBUG",
3644 "-DXNN_TEST_MODE=1",
3645 ],
3646 gcc_copts = xnnpack_gcc_std_copts(),
3647 gcc_x86_copts = ["-msse4.1"],
3648 msvc_copts = xnnpack_msvc_std_copts(),
3649 msvc_x86_32_copts = ["/arch:SSE2"],
3650 x86_srcs = SSE41_UKERNELS,
3651 deps = [
3652 ":tables",
3653 "@FP16",
3654 "@pthreadpool",
3655 ],
3656)
3657
3658xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003659 name = "avx_ukernels",
3660 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003661 gcc_copts = xnnpack_gcc_std_copts(),
3662 gcc_x86_copts = ["-mavx"],
3663 msvc_copts = xnnpack_msvc_std_copts(),
3664 msvc_x86_32_copts = ["/arch:AVX"],
3665 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003666 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003667 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003668 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003669 "@FP16",
3670 "@pthreadpool",
3671 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003672)
3673
3674xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003675 name = "avx_ukernels_test_mode",
3676 hdrs = INTERNAL_HDRS,
3677 copts = [
3678 "-UNDEBUG",
3679 "-DXNN_TEST_MODE=1",
3680 ],
3681 gcc_copts = xnnpack_gcc_std_copts(),
3682 gcc_x86_copts = ["-mavx"],
3683 msvc_copts = xnnpack_msvc_std_copts(),
3684 msvc_x86_32_copts = ["/arch:AVX"],
3685 msvc_x86_64_copts = ["/arch:AVX"],
3686 x86_srcs = AVX_UKERNELS,
3687 deps = [
3688 ":tables",
3689 "@FP16",
3690 "@pthreadpool",
3691 ],
3692)
3693
3694xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003695 name = "xop_ukernels",
3696 hdrs = INTERNAL_HDRS,
3697 gcc_copts = xnnpack_gcc_std_copts(),
3698 gcc_x86_copts = ["-mxop"],
3699 msvc_copts = xnnpack_msvc_std_copts(),
3700 msvc_x86_32_copts = ["/arch:AVX"],
3701 msvc_x86_64_copts = ["/arch:AVX"],
3702 x86_srcs = XOP_UKERNELS,
3703 deps = [
3704 ":tables",
3705 "@FP16",
3706 "@pthreadpool",
3707 ],
3708)
3709
3710xnnpack_cc_library(
3711 name = "xop_ukernels_test_mode",
3712 hdrs = INTERNAL_HDRS,
3713 copts = [
3714 "-UNDEBUG",
3715 "-DXNN_TEST_MODE=1",
3716 ],
3717 gcc_copts = xnnpack_gcc_std_copts(),
3718 gcc_x86_copts = ["-mxop"],
3719 msvc_copts = xnnpack_msvc_std_copts(),
3720 msvc_x86_32_copts = ["/arch:AVX"],
3721 msvc_x86_64_copts = ["/arch:AVX"],
3722 x86_srcs = XOP_UKERNELS,
3723 deps = [
3724 ":tables",
3725 "@FP16",
3726 "@pthreadpool",
3727 ],
3728)
3729
3730xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003731 name = "fma3_ukernels",
3732 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003733 gcc_copts = xnnpack_gcc_std_copts(),
3734 gcc_x86_copts = ["-mfma"],
3735 msvc_copts = xnnpack_msvc_std_copts(),
3736 msvc_x86_32_copts = ["/arch:AVX"],
3737 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003738 x86_srcs = FMA3_UKERNELS,
3739 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003740 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003741 "@FP16",
3742 "@pthreadpool",
3743 ],
3744)
3745
3746xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003747 name = "fma3_ukernels_test_mode",
3748 hdrs = INTERNAL_HDRS,
3749 copts = [
3750 "-UNDEBUG",
3751 "-DXNN_TEST_MODE=1",
3752 ],
3753 gcc_copts = xnnpack_gcc_std_copts(),
3754 gcc_x86_copts = ["-mfma"],
3755 msvc_copts = xnnpack_msvc_std_copts(),
3756 msvc_x86_32_copts = ["/arch:AVX"],
3757 msvc_x86_64_copts = ["/arch:AVX"],
3758 x86_srcs = FMA3_UKERNELS,
3759 deps = [
3760 ":tables",
3761 "@FP16",
3762 "@pthreadpool",
3763 ],
3764)
3765
3766xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003767 name = "avx2_ukernels",
3768 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003769 gcc_copts = xnnpack_gcc_std_copts(),
3770 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003771 "-mfma",
3772 "-mavx2",
3773 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003774 msvc_copts = xnnpack_msvc_std_copts(),
3775 msvc_x86_32_copts = ["/arch:AVX2"],
3776 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003777 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003778 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003779 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003780 "@FP16",
3781 "@pthreadpool",
3782 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003783)
3784
3785xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003786 name = "avx2_ukernels_test_mode",
3787 hdrs = INTERNAL_HDRS,
3788 copts = [
3789 "-UNDEBUG",
3790 "-DXNN_TEST_MODE=1",
3791 ],
3792 gcc_copts = xnnpack_gcc_std_copts(),
3793 gcc_x86_copts = [
3794 "-mfma",
3795 "-mavx2",
3796 ],
3797 msvc_copts = xnnpack_msvc_std_copts(),
3798 msvc_x86_32_copts = ["/arch:AVX2"],
3799 msvc_x86_64_copts = ["/arch:AVX2"],
3800 x86_srcs = AVX2_UKERNELS,
3801 deps = [
3802 ":tables",
3803 "@FP16",
3804 "@pthreadpool",
3805 ],
3806)
3807
3808xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003809 name = "avx512f_ukernels",
3810 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003811 gcc_copts = xnnpack_gcc_std_copts(),
3812 gcc_x86_copts = ["-mavx512f"],
3813 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3814 msvc_copts = xnnpack_msvc_std_copts(),
3815 msvc_x86_32_copts = ["/arch:AVX512"],
3816 msvc_x86_64_copts = ["/arch:AVX512"],
3817 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003818 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003819 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003820 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003821 "@FP16",
3822 "@pthreadpool",
3823 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003824)
3825
3826xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003827 name = "avx512f_ukernels_test_mode",
3828 hdrs = INTERNAL_HDRS,
3829 copts = [
3830 "-UNDEBUG",
3831 "-DXNN_TEST_MODE=1",
3832 ],
3833 gcc_copts = xnnpack_gcc_std_copts(),
3834 gcc_x86_copts = ["-mavx512f"],
3835 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3836 msvc_copts = xnnpack_msvc_std_copts(),
3837 msvc_x86_32_copts = ["/arch:AVX512"],
3838 msvc_x86_64_copts = ["/arch:AVX512"],
3839 msys_copts = ["-fno-asynchronous-unwind-tables"],
3840 x86_srcs = AVX512F_UKERNELS,
3841 deps = [
3842 ":tables",
3843 "@FP16",
3844 "@pthreadpool",
3845 ],
3846)
3847
3848xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003849 name = "avx512skx_ukernels",
3850 hdrs = INTERNAL_HDRS,
3851 gcc_copts = xnnpack_gcc_std_copts(),
3852 gcc_x86_copts = [
3853 "-mavx512f",
3854 "-mavx512cd",
3855 "-mavx512bw",
3856 "-mavx512dq",
3857 "-mavx512vl",
3858 ],
3859 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3860 msvc_copts = xnnpack_msvc_std_copts(),
3861 msvc_x86_32_copts = ["/arch:AVX512"],
3862 msvc_x86_64_copts = ["/arch:AVX512"],
3863 msys_copts = ["-fno-asynchronous-unwind-tables"],
3864 x86_srcs = AVX512SKX_UKERNELS,
3865 deps = [
3866 ":tables",
3867 "@FP16",
3868 "@pthreadpool",
3869 ],
3870)
3871
3872xnnpack_cc_library(
3873 name = "avx512skx_ukernels_test_mode",
3874 hdrs = INTERNAL_HDRS,
3875 copts = [
3876 "-UNDEBUG",
3877 "-DXNN_TEST_MODE=1",
3878 ],
3879 gcc_copts = xnnpack_gcc_std_copts(),
3880 gcc_x86_copts = [
3881 "-mavx512f",
3882 "-mavx512cd",
3883 "-mavx512bw",
3884 "-mavx512dq",
3885 "-mavx512vl",
3886 ],
3887 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3888 msvc_copts = xnnpack_msvc_std_copts(),
3889 msvc_x86_32_copts = ["/arch:AVX512"],
3890 msvc_x86_64_copts = ["/arch:AVX512"],
3891 msys_copts = ["-fno-asynchronous-unwind-tables"],
3892 x86_srcs = AVX512SKX_UKERNELS,
3893 deps = [
3894 ":tables",
3895 "@FP16",
3896 "@pthreadpool",
3897 ],
3898)
3899
3900xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003901 name = "asm_ukernels",
3902 hdrs = ["src/xnnpack/assembly.h"],
3903 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07003904 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003905 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003906 wasm_srcs = WASM32_ASM_UKERNELS,
3907 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003908)
3909
Marat Dukhan3b59de22020-06-03 20:15:19 -07003910xnnpack_cc_library(
3911 name = "logging_utils",
3912 srcs = LOGGING_SRCS,
3913 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3914 copts = LOGGING_COPTS + [
3915 "-Isrc",
3916 "-Iinclude",
3917 ] + select({
3918 ":debug_build": [],
3919 "//conditions:default": xnnpack_min_size_copts(),
3920 }),
3921 gcc_copts = xnnpack_gcc_std_copts(),
3922 msvc_copts = xnnpack_msvc_std_copts(),
3923 visibility = xnnpack_visibility(),
3924 deps = [
3925 "@FP16",
3926 "@clog",
3927 "@pthreadpool",
3928 ],
3929)
3930
Marat Dukhan08c4a432019-10-03 09:29:21 -07003931xnnpack_aggregate_library(
3932 name = "ukernels",
3933 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003934 ":neon_ukernels",
3935 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003936 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003937 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003938 ":asm_ukernels",
3939 ],
3940 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003941 ":neon_ukernels",
3942 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003943 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003944 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003945 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003946 ":asm_ukernels",
3947 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003948 generic_deps = [
3949 ":scalar_ukernels",
3950 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003951 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003952 ":psimd_fastmath_ukernels",
3953 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003954 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003955 wasm_deps = [
3956 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003957 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003958 ],
3959 wasmsimd_deps = [
3960 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003961 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003962 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003963 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003964 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003965 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003966 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003967 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003968 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003969 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003970 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003971 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003972 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003973 ],
3974)
3975
Marat Dukhan33fcf782020-05-24 14:27:15 -07003976xnnpack_aggregate_library(
3977 name = "ukernels_test_mode",
3978 aarch32_deps = [
3979 ":neon_ukernels_test_mode",
3980 ":neonfma_ukernels_test_mode",
3981 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003982 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003983 ":asm_ukernels",
3984 ],
3985 aarch64_deps = [
3986 ":neon_ukernels_test_mode",
3987 ":neonfma_ukernels_test_mode",
3988 ":neonv8_ukernels_test_mode",
3989 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003990 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003991 ":asm_ukernels",
3992 ],
3993 generic_deps = [
3994 ":scalar_ukernels_test_mode",
3995 ],
3996 psimd_deps = [
3997 ":psimd_fastmath_ukernels_test_mode",
3998 ":psimd_accmath_ukernels_test_mode",
3999 ],
4000 wasm_deps = [
4001 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004002 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004003 ],
4004 wasmsimd_deps = [
4005 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004006 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004007 ],
4008 x86_deps = [
4009 ":sse2_ukernels_test_mode",
4010 ":ssse3_ukernels_test_mode",
4011 ":sse41_ukernels_test_mode",
4012 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004013 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004014 ":fma3_ukernels_test_mode",
4015 ":avx2_ukernels_test_mode",
4016 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004017 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004018 ],
4019)
4020
Marat Dukhan86df3842020-06-17 14:03:31 -07004021# This target must be compatible with Android NDK r17 and compile with Clang 6.
4022# It is similar to :ukernels target, but excludes microkernels which require
4023# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07004024xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07004025 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004026 aarch32_deps = [
4027 ":neon_ukernels",
4028 ":neonfma_ukernels",
4029 ":neonv8_ukernels",
4030 ":asm_ukernels",
4031 ],
4032 aarch64_deps = [
4033 ":neon_ukernels",
4034 ":neonfma_ukernels",
4035 ":neonv8_ukernels",
4036 ":asm_ukernels",
4037 ],
4038 generic_deps = [
4039 ":scalar_ukernels",
4040 ],
4041 psimd_deps = [
4042 ":psimd_fastmath_ukernels",
4043 ":psimd_accmath_ukernels",
4044 ],
4045 wasm_deps = [
4046 ":wasm_ukernels",
4047 ],
4048 wasmsimd_deps = [
4049 ":wasm_ukernels",
4050 ],
4051 x86_deps = [
4052 ":sse2_ukernels",
4053 ":ssse3_ukernels",
4054 ":sse41_ukernels",
4055 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004056 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004057 ":fma3_ukernels",
4058 ":avx2_ukernels",
4059 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004060 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004061 ],
4062)
4063
Marat Dukhan08c4a432019-10-03 09:29:21 -07004064xnnpack_cc_library(
4065 name = "im2col",
4066 srcs = ["src/im2col.c"],
4067 hdrs = [
4068 "src/xnnpack/common.h",
4069 "src/xnnpack/im2col.h",
4070 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004071 gcc_copts = xnnpack_gcc_std_copts(),
4072 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004073)
4074
4075xnnpack_cc_library(
4076 name = "indirection",
4077 srcs = ["src/indirection.c"],
4078 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004079 gcc_copts = xnnpack_gcc_std_copts(),
4080 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004081 deps = [
4082 "@FP16",
4083 "@FXdiv",
4084 "@pthreadpool",
4085 ],
4086)
4087
4088xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004089 name = "indirection_test_mode",
4090 srcs = ["src/indirection.c"],
4091 hdrs = INTERNAL_HDRS,
4092 copts = [
4093 "-UNDEBUG",
4094 "-DXNN_TEST_MODE=1",
4095 ],
4096 gcc_copts = xnnpack_gcc_std_copts(),
4097 msvc_copts = xnnpack_msvc_std_copts(),
4098 deps = [
4099 "@FP16",
4100 "@FXdiv",
4101 "@pthreadpool",
4102 ],
4103)
4104
4105xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004106 name = "packing",
4107 srcs = ["src/packing.c"],
4108 hdrs = INTERNAL_HDRS,
4109 gcc_copts = xnnpack_gcc_std_copts(),
4110 msvc_copts = xnnpack_msvc_std_copts(),
4111 deps = [
4112 "@FP16",
4113 "@FXdiv",
4114 "@pthreadpool",
4115 ],
4116)
4117
4118xnnpack_cc_library(
4119 name = "packing_test_mode",
4120 srcs = ["src/packing.c"],
4121 hdrs = INTERNAL_HDRS,
4122 copts = [
4123 "-UNDEBUG",
4124 "-DXNN_TEST_MODE=1",
4125 ],
4126 gcc_copts = xnnpack_gcc_std_copts(),
4127 msvc_copts = xnnpack_msvc_std_copts(),
4128 deps = [
4129 "@FP16",
4130 "@FXdiv",
4131 "@pthreadpool",
4132 ],
4133)
4134
4135xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004136 name = "operator_run",
4137 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004138 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004139 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004140 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4141 "//conditions:default": [],
4142 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004143 gcc_copts = xnnpack_gcc_std_copts(),
4144 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004145 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004146 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004147 "@FP16",
4148 "@FXdiv",
4149 "@clog",
4150 "@pthreadpool",
4151 ],
4152)
4153
Chao Mei6ddfc602020-05-13 22:29:36 -07004154xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004155 name = "operator_run_test_mode",
4156 srcs = ["src/operator-run.c"],
4157 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4158 copts = LOGGING_COPTS + [
4159 "-UNDEBUG",
4160 "-DXNN_TEST_MODE=1",
4161 ] + select({
4162 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4163 "//conditions:default": [],
4164 }),
4165 gcc_copts = xnnpack_gcc_std_copts(),
4166 msvc_copts = xnnpack_msvc_std_copts(),
4167 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004168 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004169 "@FP16",
4170 "@FXdiv",
4171 "@clog",
4172 "@pthreadpool",
4173 ],
4174)
4175
4176xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004177 name = "memory_planner",
4178 srcs = ["src/memory-planner.c"],
4179 hdrs = INTERNAL_HDRS,
4180 defines = select({
4181 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4182 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4183 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4184 }),
4185 gcc_copts = xnnpack_gcc_std_copts(),
4186 msvc_copts = xnnpack_msvc_std_copts(),
4187 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004188 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004189 "@pthreadpool",
4190 ],
4191)
4192
Marat Dukhan33fcf782020-05-24 14:27:15 -07004193xnnpack_cc_library(
4194 name = "memory_planner_test_mode",
4195 srcs = ["src/memory-planner.c"],
4196 hdrs = INTERNAL_HDRS,
4197 copts = [
4198 "-UNDEBUG",
4199 "-DXNN_TEST_MODE=1",
4200 ],
4201 defines = select({
4202 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4203 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4204 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4205 }),
4206 gcc_copts = xnnpack_gcc_std_copts(),
4207 msvc_copts = xnnpack_msvc_std_copts(),
4208 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004209 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004210 "@pthreadpool",
4211 ],
4212)
4213
Marat Dukhan08c4a432019-10-03 09:29:21 -07004214cc_library(
4215 name = "enable_assembly",
4216 defines = select({
4217 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4218 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004219 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004220 }),
4221)
4222
Marat Dukhan9de90e02020-06-18 16:04:12 -07004223cc_library(
4224 name = "enable_sparse",
4225 defines = select({
4226 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4227 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07004228 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004229 }),
4230)
4231
Marat Dukhancf056b22019-10-07 10:26:29 -07004232xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004233 name = "operators",
4234 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004235 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004236 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004237 ],
4238 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004239 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004240 "-Isrc",
4241 "-Iinclude",
4242 ] + select({
4243 ":debug_build": [],
4244 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004245 }) + select({
4246 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4247 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004248 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004249 gcc_copts = xnnpack_gcc_std_copts(),
4250 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004251 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004252 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004253 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004254 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004255 "@FP16",
4256 "@FXdiv",
4257 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004258 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004259 ],
4260)
4261
Marat Dukhan10a38082020-04-17 03:58:35 -07004262xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004263 name = "operators_test_mode",
4264 srcs = OPERATOR_SRCS + [
4265 "src/memory.c",
4266 "src/operator-delete.c",
4267 ],
4268 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4269 copts = LOGGING_COPTS + [
4270 "-Isrc",
4271 "-Iinclude",
4272 "-UNDEBUG",
4273 "-DXNN_TEST_MODE=1",
4274 ] + select({
4275 ":debug_build": [],
4276 "//conditions:default": xnnpack_min_size_copts(),
4277 }) + select({
4278 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4279 "//conditions:default": [],
4280 }),
4281 gcc_copts = xnnpack_gcc_std_copts(),
4282 msvc_copts = xnnpack_msvc_std_copts(),
4283 deps = [
4284 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004285 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004286 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004287 "@FP16",
4288 "@FXdiv",
4289 "@clog",
4290 "@pthreadpool",
4291 ],
4292)
4293
4294xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004295 name = "XNNPACK",
4296 srcs = [
4297 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004298 "src/runtime.c",
4299 "src/subgraph.c",
4300 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004301 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004302 hdrs = ["include/xnnpack.h"],
4303 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004304 "-Isrc",
4305 "-Iinclude",
4306 ] + select({
4307 ":debug_build": [],
4308 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004309 }) + select({
4310 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4311 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004312 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004313 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004314 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004315 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004316 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",
Chao Mei6ddfc602020-05-13 22:29:36 -07004321 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004322 ":operator_run",
4323 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004324 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004325 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004326 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004327 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004328 ] + select({
4329 ":emscripten": [],
4330 "//conditions:default": ["@cpuinfo"],
4331 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004332)
4333
Marat Dukhan10a38082020-04-17 03:58:35 -07004334xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004335 name = "XNNPACK_test_mode",
4336 srcs = [
4337 "src/init.c",
4338 "src/runtime.c",
4339 "src/subgraph.c",
4340 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004341 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004342 hdrs = ["include/xnnpack.h"],
4343 copts = LOGGING_COPTS + [
4344 "-Isrc",
4345 "-Iinclude",
4346 "-UNDEBUG",
4347 "-DXNN_TEST_MODE=1",
4348 ] + select({
4349 ":debug_build": [],
4350 "//conditions:default": xnnpack_min_size_copts(),
4351 }) + select({
4352 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4353 "//conditions:default": [],
4354 }),
4355 gcc_copts = xnnpack_gcc_std_copts(),
4356 includes = ["include"],
4357 msvc_copts = xnnpack_msvc_std_copts(),
4358 visibility = xnnpack_visibility(),
4359 deps = [
4360 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004361 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004362 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004363 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004364 ":operator_run_test_mode",
4365 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004366 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004367 "@clog",
4368 "@FP16",
4369 "@pthreadpool",
4370 ] + select({
4371 ":emscripten": [],
4372 "//conditions:default": ["@cpuinfo"],
4373 }),
4374)
4375
4376xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004377 name = "xnnpack_f32",
4378 srcs = [
4379 "src/init.c",
4380 "src/runtime.c",
4381 "src/subgraph.c",
4382 "src/tensor.c",
4383 ] + SUBGRAPH_SRCS,
4384 hdrs = ["include/xnnpack.h"],
4385 copts = LOGGING_COPTS + [
4386 "-Isrc",
4387 "-Iinclude",
4388 ] + select({
4389 ":debug_build": [],
4390 "//conditions:default": xnnpack_min_size_copts(),
4391 }) + select({
4392 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4393 "//conditions:default": [],
4394 }),
4395 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004396 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004397 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004398 "XNN_NO_U8_OPERATORS",
4399 "XNN_NO_X8_OPERATORS",
4400 "XNN_NO_F16_OPERATORS",
4401 "XNN_NO_X16_OPERATORS",
4402 ],
4403 gcc_copts = xnnpack_gcc_std_copts(),
4404 includes = ["include"],
4405 msvc_copts = xnnpack_msvc_std_copts(),
4406 visibility = xnnpack_visibility(),
4407 deps = [
4408 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004409 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004410 ":logging_utils",
4411 ":memory_planner",
4412 ":operator_run",
4413 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004414 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004415 "@clog",
4416 "@FP16",
4417 "@pthreadpool",
4418 ] + select({
4419 ":emscripten": [],
4420 "//conditions:default": ["@cpuinfo"],
4421 }),
4422)
4423
4424xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004425 name = "xnnpack_operators_nhwc_f32",
4426 srcs = [
4427 "src/init.c",
4428 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004429 hdrs = ["include/xnnpack.h"],
4430 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004431 "-Isrc",
4432 "-Iinclude",
4433 ] + select({
4434 ":debug_build": [],
4435 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004436 }) + select({
4437 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4438 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004439 }),
4440 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004441 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004442 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004443 "XNN_NO_U8_OPERATORS",
4444 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004445 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004446 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004447 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004448 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004449 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004450 visibility = xnnpack_visibility(),
4451 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004452 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004453 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004454 ":operator_run",
4455 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004456 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004457 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004458 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004459 ] + select({
4460 ":emscripten": [],
4461 "//conditions:default": ["@cpuinfo"],
4462 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004463)
4464
Marat Dukhancf056b22019-10-07 10:26:29 -07004465xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004466 name = "bench_utils",
4467 srcs = ["bench/utils.cc"],
4468 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004469 deps = [
4470 "@com_google_benchmark//:benchmark",
4471 "@cpuinfo",
4472 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004473)
4474
Frank Barchard7e955972019-10-11 10:34:25 -07004475######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004476
4477xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004478 name = "qs8_gemm_bench",
4479 srcs = [
4480 "bench/gemm.h",
4481 "bench/qs8-gemm.cc",
4482 "src/xnnpack/AlignedAllocator.h",
4483 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004484 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4485 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004486)
4487
4488xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004489 name = "qs8_requantization_bench",
4490 srcs = [
4491 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004492 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004493 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004494 ] + MICROKERNEL_BENCHMARK_HDRS,
4495 deps = MICROKERNEL_BENCHMARK_DEPS,
4496)
4497
4498xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004499 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004500 srcs = [
4501 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004502 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004503 "src/xnnpack/AlignedAllocator.h",
4504 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004505 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004506 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004507)
4508
4509xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004510 name = "qu8_requantization_bench",
4511 srcs = [
4512 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004513 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004514 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004515 ] + MICROKERNEL_BENCHMARK_HDRS,
4516 deps = MICROKERNEL_BENCHMARK_DEPS,
4517)
4518
4519xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004520 name = "f16_igemm_bench",
4521 srcs = [
4522 "bench/f16-igemm.cc",
4523 "bench/conv.h",
4524 "bench/google/conv.h",
4525 "src/xnnpack/AlignedAllocator.h",
4526 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004527 deps = MICROKERNEL_BENCHMARK_DEPS + [
4528 ":indirection",
4529 ":packing",
4530 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004531)
4532
4533xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004534 name = "f16_gemm_bench",
4535 srcs = [
4536 "bench/f16-gemm.cc",
4537 "bench/gemm.h",
4538 "src/xnnpack/AlignedAllocator.h",
4539 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004540 deps = MICROKERNEL_BENCHMARK_DEPS + [
4541 ":packing",
4542 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004543)
4544
4545xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004546 name = "f16_spmm_bench",
4547 srcs = [
4548 "bench/f16-spmm.cc",
4549 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004550 "src/xnnpack/AlignedAllocator.h",
4551 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004552 deps = MICROKERNEL_BENCHMARK_DEPS,
4553)
4554
4555xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004556 name = "f32_igemm_bench",
4557 srcs = [
4558 "bench/f32-igemm.cc",
4559 "bench/conv.h",
4560 "src/xnnpack/AlignedAllocator.h",
4561 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004562 deps = MICROKERNEL_BENCHMARK_DEPS + [
4563 ":indirection",
4564 ":packing",
4565 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004566)
4567
4568xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004569 name = "f16_relu_bench",
4570 srcs = [
4571 "bench/f16-relu.cc",
4572 "src/xnnpack/AlignedAllocator.h",
4573 ] + MICROKERNEL_BENCHMARK_HDRS,
4574 deps = MICROKERNEL_BENCHMARK_DEPS,
4575)
4576
4577xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004578 name = "f32_conv_hwc_bench",
4579 srcs = [
4580 "bench/f32-conv-hwc.cc",
4581 "bench/dconv.h",
4582 "src/xnnpack/AlignedAllocator.h",
4583 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004584 deps = MICROKERNEL_BENCHMARK_DEPS + [
4585 ":packing",
4586 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004587)
4588
4589xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004590 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004591 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004592 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004593 "bench/dconv.h",
4594 "src/xnnpack/AlignedAllocator.h",
4595 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004596 deps = MICROKERNEL_BENCHMARK_DEPS + [
4597 ":packing",
4598 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004599)
4600
4601xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004602 name = "f16_dwconv_bench",
4603 srcs = [
4604 "bench/f16-dwconv.cc",
4605 "bench/dwconv.h",
4606 "bench/google/dwconv.h",
4607 "src/xnnpack/AlignedAllocator.h",
4608 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004609 deps = MICROKERNEL_BENCHMARK_DEPS + [
4610 ":indirection",
4611 ":packing",
4612 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004613)
4614
4615xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004616 name = "f32_dwconv_bench",
4617 srcs = [
4618 "bench/f32-dwconv.cc",
4619 "bench/dwconv.h",
4620 "src/xnnpack/AlignedAllocator.h",
4621 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004622 deps = MICROKERNEL_BENCHMARK_DEPS + [
4623 ":indirection",
4624 ":packing",
4625 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004626)
4627
4628xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07004629 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004630 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07004631 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004632 "bench/dwconv.h",
4633 "src/xnnpack/AlignedAllocator.h",
4634 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004635 deps = MICROKERNEL_BENCHMARK_DEPS + [
4636 ":indirection",
4637 ":packing",
4638 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004639)
4640
4641xnnpack_benchmark(
4642 name = "f32_gemm_bench",
4643 srcs = [
4644 "bench/f32-gemm.cc",
4645 "bench/gemm.h",
4646 "src/xnnpack/AlignedAllocator.h",
4647 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004648 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004649 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004650)
4651
4652xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004653 name = "f32_hswish_bench",
4654 srcs = [
4655 "bench/f32-hswish.cc",
4656 "src/xnnpack/AlignedAllocator.h",
4657 ] + MICROKERNEL_BENCHMARK_HDRS,
4658 deps = MICROKERNEL_BENCHMARK_DEPS,
4659)
4660
4661xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004662 name = "f32_raddexpminusmax_bench",
4663 srcs = [
4664 "bench/f32-raddexpminusmax.cc",
4665 "src/xnnpack/AlignedAllocator.h",
4666 ] + MICROKERNEL_BENCHMARK_HDRS,
4667 deps = MICROKERNEL_BENCHMARK_DEPS,
4668)
4669
4670xnnpack_benchmark(
4671 name = "f32_raddextexp_bench",
4672 srcs = [
4673 "bench/f32-raddextexp.cc",
4674 "src/xnnpack/AlignedAllocator.h",
4675 ] + MICROKERNEL_BENCHMARK_HDRS,
4676 deps = MICROKERNEL_BENCHMARK_DEPS,
4677)
4678
4679xnnpack_benchmark(
4680 name = "f32_raddstoreexpminusmax_bench",
4681 srcs = [
4682 "bench/f32-raddstoreexpminusmax.cc",
4683 "src/xnnpack/AlignedAllocator.h",
4684 ] + MICROKERNEL_BENCHMARK_HDRS,
4685 deps = MICROKERNEL_BENCHMARK_DEPS,
4686)
4687
4688xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004689 name = "f32_relu_bench",
4690 srcs = [
4691 "bench/f32-relu.cc",
4692 "src/xnnpack/AlignedAllocator.h",
4693 ] + MICROKERNEL_BENCHMARK_HDRS,
4694 deps = MICROKERNEL_BENCHMARK_DEPS,
4695)
4696
4697xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004698 name = "f32_rmax_bench",
4699 srcs = [
4700 "bench/f32-rmax.cc",
4701 "src/xnnpack/AlignedAllocator.h",
4702 ] + MICROKERNEL_BENCHMARK_HDRS,
4703 deps = MICROKERNEL_BENCHMARK_DEPS,
4704)
4705
4706xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004707 name = "f32_sigmoid_bench",
4708 srcs = [
4709 "bench/f32-sigmoid.cc",
4710 "src/xnnpack/AlignedAllocator.h",
4711 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004712 deps = MICROKERNEL_BENCHMARK_DEPS,
4713)
4714
4715xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004716 name = "f32_spmm_bench",
4717 srcs = [
4718 "bench/f32-spmm.cc",
4719 "bench/gemm.h",
4720 "src/xnnpack/AlignedAllocator.h",
4721 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004722 deps = MICROKERNEL_BENCHMARK_DEPS,
4723)
4724
4725xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004726 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004727 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004728 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004729 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004730 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004731 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004732)
4733
4734xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004735 name = "f32_vscaleexpminusmax_bench",
4736 srcs = [
4737 "bench/f32-vscaleexpminusmax.cc",
4738 "src/xnnpack/AlignedAllocator.h",
4739 ] + MICROKERNEL_BENCHMARK_HDRS,
4740 deps = MICROKERNEL_BENCHMARK_DEPS,
4741)
4742
4743xnnpack_benchmark(
4744 name = "f32_vscaleextexp_bench",
4745 srcs = [
4746 "bench/f32-vscaleextexp.cc",
4747 "src/xnnpack/AlignedAllocator.h",
4748 ] + MICROKERNEL_BENCHMARK_HDRS,
4749 deps = MICROKERNEL_BENCHMARK_DEPS,
4750)
4751
4752xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004753 name = "f32_vsqrt_bench",
4754 srcs = [
4755 "bench/f32-vsqrt.cc",
4756 "src/xnnpack/AlignedAllocator.h",
4757 ] + MICROKERNEL_BENCHMARK_HDRS,
4758 deps = MICROKERNEL_BENCHMARK_DEPS,
4759)
4760
4761xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004762 name = "f32_im2col_gemm_bench",
4763 srcs = [
4764 "bench/f32-im2col-gemm.cc",
4765 "bench/conv.h",
4766 "src/xnnpack/AlignedAllocator.h",
4767 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004768 deps = MICROKERNEL_BENCHMARK_DEPS + [
4769 ":im2col",
4770 ":packing",
4771 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004772)
4773
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004774xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004775 name = "rounding_bench",
4776 srcs = [
4777 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004778 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004779 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004780 ] + MICROKERNEL_BENCHMARK_HDRS,
4781 deps = MICROKERNEL_BENCHMARK_DEPS,
4782)
4783
Marat Dukhan08c4a432019-10-03 09:29:21 -07004784########################### Benchmarks for operators ###########################
4785
4786xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004787 name = "average_pooling_bench",
4788 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004789 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004790 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004791 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004792)
4793
4794xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004795 name = "bankers_rounding_bench",
4796 srcs = ["bench/bankers-rounding.cc"],
4797 copts = xnnpack_optional_tflite_copts(),
4798 tags = ["nowin32"],
4799 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4800)
4801
4802xnnpack_benchmark(
4803 name = "ceiling_bench",
4804 srcs = ["bench/ceiling.cc"],
4805 copts = xnnpack_optional_tflite_copts(),
4806 tags = ["nowin32"],
4807 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4808)
4809
4810xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004811 name = "channel_shuffle_bench",
4812 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004813 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004814)
4815
4816xnnpack_benchmark(
4817 name = "convolution_bench",
4818 srcs = ["bench/convolution.cc"],
4819 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004820 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004821 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004822)
4823
4824xnnpack_benchmark(
4825 name = "deconvolution_bench",
4826 srcs = ["bench/deconvolution.cc"],
4827 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004828 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004829 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004830)
4831
4832xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004833 name = "floor_bench",
4834 srcs = ["bench/floor.cc"],
4835 copts = xnnpack_optional_tflite_copts(),
4836 tags = ["nowin32"],
4837 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4838)
4839
4840xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004841 name = "global_average_pooling_bench",
4842 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004843 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004844)
4845
4846xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004847 name = "hardswish_bench",
4848 srcs = ["bench/hardswish.cc"],
4849 copts = xnnpack_optional_tflite_copts(),
4850 tags = ["nowin32"],
4851 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4852)
4853
4854xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004855 name = "max_pooling_bench",
4856 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004857 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004858)
4859
4860xnnpack_benchmark(
4861 name = "sigmoid_bench",
4862 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004863 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004864 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004865 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004866)
4867
4868xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004869 name = "prelu_bench",
4870 srcs = ["bench/prelu.cc"],
4871 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004872 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004873 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004874)
4875
4876xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004877 name = "softmax_bench",
4878 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004879 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004880 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004881 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004882)
4883
Marat Dukhan87727142020-06-24 15:24:10 -07004884xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004885 name = "square_root_bench",
4886 srcs = ["bench/square-root.cc"],
4887 copts = xnnpack_optional_tflite_copts(),
4888 tags = ["nowin32"],
4889 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4890)
4891
4892xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004893 name = "truncation_bench",
4894 srcs = ["bench/truncation.cc"],
4895 deps = OPERATOR_BENCHMARK_DEPS,
4896)
4897
Marat Dukhanc068bb62019-10-04 13:24:39 -07004898############################# End-to-end benchmarks ############################
4899
4900cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004901 name = "fp32_mobilenet_v1",
4902 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004903 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004904 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004905 linkstatic = True,
4906 deps = [
4907 ":XNNPACK",
4908 "@pthreadpool",
4909 ],
4910)
4911
4912cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004913 name = "fp16_mobilenet_v1",
4914 srcs = ["models/fp16-mobilenet-v1.cc"],
4915 hdrs = ["models/models.h"],
4916 copts = xnnpack_std_cxxopts(),
4917 linkstatic = True,
4918 deps = [
4919 ":XNNPACK",
4920 "@FP16",
4921 "@pthreadpool",
4922 ],
4923)
4924
4925cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004926 name = "qs8_mobilenet_v1",
4927 srcs = ["models/qs8-mobilenet-v1.cc"],
4928 hdrs = ["models/models.h"],
4929 copts = xnnpack_std_cxxopts(),
4930 linkstatic = True,
4931 deps = [
4932 ":XNNPACK",
4933 "@pthreadpool",
4934 ],
4935)
4936
4937cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004938 name = "qs8_mobilenet_v2",
4939 srcs = ["models/qs8-mobilenet-v2.cc"],
4940 hdrs = ["models/models.h"],
4941 copts = xnnpack_std_cxxopts(),
4942 linkstatic = True,
4943 deps = [
4944 ":XNNPACK",
4945 "@pthreadpool",
4946 ],
4947)
4948
4949cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004950 name = "fp32_mobilenet_v2",
4951 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004952 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004953 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004954 linkstatic = True,
4955 deps = [
4956 ":XNNPACK",
4957 "@pthreadpool",
4958 ],
4959)
4960
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004961cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004962 name = "fp16_mobilenet_v2",
4963 srcs = ["models/fp16-mobilenet-v2.cc"],
4964 hdrs = ["models/models.h"],
4965 copts = xnnpack_std_cxxopts(),
4966 linkstatic = True,
4967 deps = [
4968 ":XNNPACK",
4969 "@FP16",
4970 "@pthreadpool",
4971 ],
4972)
4973
4974cc_library(
4975 name = "fp32_mobilenet_v3_large",
4976 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004977 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004978 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004979 linkstatic = True,
4980 deps = [
4981 ":XNNPACK",
4982 "@pthreadpool",
4983 ],
4984)
4985
4986cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004987 name = "fp16_mobilenet_v3_large",
4988 srcs = ["models/fp16-mobilenet-v3-large.cc"],
4989 hdrs = ["models/models.h"],
4990 copts = xnnpack_std_cxxopts(),
4991 linkstatic = True,
4992 deps = [
4993 ":XNNPACK",
4994 "@FP16",
4995 "@pthreadpool",
4996 ],
4997)
4998
4999cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005000 name = "fp32_mobilenet_v3_small",
5001 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005002 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005003 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005004 linkstatic = True,
5005 deps = [
5006 ":XNNPACK",
5007 "@pthreadpool",
5008 ],
5009)
5010
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005011cc_library(
5012 name = "fp16_mobilenet_v3_small",
5013 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5014 hdrs = ["models/models.h"],
5015 copts = xnnpack_std_cxxopts(),
5016 linkstatic = True,
5017 deps = [
5018 ":XNNPACK",
5019 "@FP16",
5020 "@pthreadpool",
5021 ],
5022)
5023
Marat Dukhanc068bb62019-10-04 13:24:39 -07005024xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005025 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005026 srcs = [
5027 "bench/f32-dwconv-e2e.cc",
5028 "bench/end2end.h",
5029 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005030 deps = MICROKERNEL_BENCHMARK_DEPS + [
5031 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005032 ":fp32_mobilenet_v1",
5033 ":fp32_mobilenet_v2",
5034 ":fp32_mobilenet_v3_large",
5035 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005036 ],
5037)
5038
5039xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005040 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005041 srcs = [
5042 "bench/f32-gemm-e2e.cc",
5043 "bench/end2end.h",
5044 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005045 deps = MICROKERNEL_BENCHMARK_DEPS + [
5046 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005047 ":fp32_mobilenet_v1",
5048 ":fp32_mobilenet_v2",
5049 ":fp32_mobilenet_v3_large",
5050 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005051 ],
5052)
5053
5054xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005055 name = "end2end_bench",
5056 srcs = ["bench/end2end.cc"],
5057 deps = [
5058 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005059 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005060 ":fp16_mobilenet_v1",
5061 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005062 ":fp16_mobilenet_v3_large",
5063 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005064 ":fp32_mobilenet_v1",
5065 ":fp32_mobilenet_v2",
5066 ":fp32_mobilenet_v3_large",
5067 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005068 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005069 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005070 "@pthreadpool",
5071 ],
5072)
5073
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005074#################### Accuracy evaluation for math functions ####################
5075
5076xnnpack_benchmark(
5077 name = "f32_exp_eval",
5078 srcs = [
5079 "eval/f32-exp.cc",
5080 "src/xnnpack/AlignedAllocator.h",
5081 ] + ACCURACY_EVAL_HDRS,
5082 deps = ACCURACY_EVAL_DEPS,
5083)
5084
Marat Dukhan515c9772019-10-17 18:07:57 -07005085xnnpack_benchmark(
5086 name = "f32_expminus_eval",
5087 srcs = [
5088 "eval/f32-expminus.cc",
5089 "src/xnnpack/AlignedAllocator.h",
5090 ] + ACCURACY_EVAL_HDRS,
5091 deps = ACCURACY_EVAL_DEPS,
5092)
5093
Marat Dukhan98ba4412019-10-23 02:14:28 -07005094xnnpack_benchmark(
5095 name = "f32_extexp_eval",
5096 srcs = [
5097 "eval/f32-extexp.cc",
5098 "src/xnnpack/AlignedAllocator.h",
5099 ] + ACCURACY_EVAL_HDRS,
5100 deps = ACCURACY_EVAL_DEPS,
5101)
5102
Marat Dukhan8853b822020-05-07 12:19:01 -07005103xnnpack_unit_test(
5104 name = "f32_roundne_eval",
5105 srcs = [
5106 "eval/f32-roundne.cc",
5107 "src/xnnpack/AlignedAllocator.h",
5108 "src/xnnpack/math-stubs.h",
5109 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005110 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005111 deps = MICROKERNEL_TEST_DEPS,
5112)
5113
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005114xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005115 name = "f32_roundd_eval",
5116 srcs = [
5117 "eval/f32-roundd.cc",
5118 "src/xnnpack/AlignedAllocator.h",
5119 "src/xnnpack/math-stubs.h",
5120 ] + MICROKERNEL_TEST_HDRS,
5121 automatic = False,
5122 deps = MICROKERNEL_TEST_DEPS,
5123)
5124
5125xnnpack_unit_test(
5126 name = "f32_roundu_eval",
5127 srcs = [
5128 "eval/f32-roundu.cc",
5129 "src/xnnpack/AlignedAllocator.h",
5130 "src/xnnpack/math-stubs.h",
5131 ] + MICROKERNEL_TEST_HDRS,
5132 automatic = False,
5133 deps = MICROKERNEL_TEST_DEPS,
5134)
5135
5136xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005137 name = "f32_roundz_eval",
5138 srcs = [
5139 "eval/f32-roundz.cc",
5140 "src/xnnpack/AlignedAllocator.h",
5141 "src/xnnpack/math-stubs.h",
5142 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005143 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005144 deps = MICROKERNEL_TEST_DEPS,
5145)
5146
Marat Dukhan346a9e52019-11-15 09:06:30 -08005147xnnpack_benchmark(
5148 name = "f32_sigmoid_eval",
5149 srcs = [
5150 "eval/f32-sigmoid.cc",
5151 "src/xnnpack/AlignedAllocator.h",
5152 ] + ACCURACY_EVAL_HDRS,
5153 deps = ACCURACY_EVAL_DEPS,
5154)
5155
Marat Dukhan84000762020-06-29 18:38:43 -07005156xnnpack_benchmark(
5157 name = "f32_sqrt_eval",
5158 srcs = [
5159 "eval/f32-sqrt.cc",
5160 "src/xnnpack/AlignedAllocator.h",
5161 ] + ACCURACY_EVAL_HDRS,
5162 deps = ACCURACY_EVAL_DEPS + [
5163 ":bench_utils",
5164 ],
5165)
5166
Marat Dukhan08c4a432019-10-03 09:29:21 -07005167######################### Unit tests for micro-kernels #########################
5168
5169xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005170 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005171 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005172 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005173 "test/gemm-microkernel-tester.h",
5174 "src/xnnpack/AlignedAllocator.h",
5175 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005176 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005177)
5178
5179xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005180 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005181 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005182 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005183 "test/spmm-microkernel-tester.h",
5184 "src/xnnpack/AlignedAllocator.h",
5185 ] + MICROKERNEL_TEST_HDRS,
5186 deps = MICROKERNEL_TEST_DEPS,
5187)
5188
5189xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005190 name = "f32_argmaxpool_test",
5191 srcs = [
5192 "test/f32-argmaxpool.cc",
5193 "test/argmaxpool-microkernel-tester.h",
5194 "src/xnnpack/AlignedAllocator.h",
5195 ] + MICROKERNEL_TEST_HDRS,
5196 deps = MICROKERNEL_TEST_DEPS,
5197)
5198
5199xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005200 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005201 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005202 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005203 "test/avgpool-microkernel-tester.h",
5204 "src/xnnpack/AlignedAllocator.h",
5205 ] + MICROKERNEL_TEST_HDRS,
5206 deps = MICROKERNEL_TEST_DEPS,
5207)
5208
5209xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005210 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005211 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005212 "test/f32-ibilinear.cc",
5213 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005214 "src/xnnpack/AlignedAllocator.h",
5215 ] + MICROKERNEL_TEST_HDRS,
5216 deps = MICROKERNEL_TEST_DEPS,
5217)
5218
5219xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005220 name = "f32_ibilinear_chw_test",
5221 srcs = [
5222 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005223 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005224 "src/xnnpack/AlignedAllocator.h",
5225 ] + MICROKERNEL_TEST_HDRS,
5226 deps = MICROKERNEL_TEST_DEPS,
5227)
5228
5229xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005230 name = "f16_clamp_test",
5231 srcs = [
5232 "test/f16-clamp.cc",
5233 "test/clamp-microkernel-tester.h",
5234 ] + MICROKERNEL_TEST_HDRS,
5235 deps = MICROKERNEL_TEST_DEPS,
5236)
5237
5238xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005239 name = "f32_clamp_test",
5240 srcs = [
5241 "test/f32-clamp.cc",
5242 "test/clamp-microkernel-tester.h",
5243 ] + MICROKERNEL_TEST_HDRS,
5244 deps = MICROKERNEL_TEST_DEPS,
5245)
5246
5247xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005248 name = "f32_igemm_test",
5249 srcs = [
5250 "test/f32-igemm.cc",
5251 "test/gemm-microkernel-tester.h",
5252 "src/xnnpack/AlignedAllocator.h",
5253 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005254 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005255)
5256
5257xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005258 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005259 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005260 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005261 "test/gemm-microkernel-tester.h",
5262 "src/xnnpack/AlignedAllocator.h",
5263 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005264 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005265)
5266
5267xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005268 name = "f32_igemm_minmax_test",
5269 srcs = [
5270 "test/f32-igemm-minmax.cc",
5271 "test/gemm-microkernel-tester.h",
5272 "src/xnnpack/AlignedAllocator.h",
5273 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005274 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005275)
5276
5277xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005278 name = "f16_igemm_minmax_test",
5279 srcs = [
5280 "test/f16-igemm-minmax.cc",
5281 "test/gemm-microkernel-tester.h",
5282 "src/xnnpack/AlignedAllocator.h",
5283 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005284 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005285)
5286
5287xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005288 name = "f32_conv_hwc_test",
5289 srcs = [
5290 "test/f32-conv-hwc.cc",
5291 "test/conv-hwc-microkernel-tester.h",
5292 "src/xnnpack/AlignedAllocator.h",
5293 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005294 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005295)
5296
5297xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005298 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005299 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005300 "test/f32-conv-hwc2chw.cc",
5301 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005302 "src/xnnpack/AlignedAllocator.h",
5303 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005304 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005305)
5306
5307xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005308 name = "f32_dwconv_test",
5309 srcs = [
5310 "test/f32-dwconv.cc",
5311 "test/dwconv-microkernel-tester.h",
5312 "src/xnnpack/AlignedAllocator.h",
5313 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005314 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005315)
5316
5317xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005318 name = "f16_dwconv_minmax_test",
5319 srcs = [
5320 "test/f16-dwconv-minmax.cc",
5321 "test/dwconv-microkernel-tester.h",
5322 "src/xnnpack/AlignedAllocator.h",
5323 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005324 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005325)
5326
5327xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005328 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005329 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005330 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005331 "test/dwconv-microkernel-tester.h",
5332 "src/xnnpack/AlignedAllocator.h",
5333 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005334 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005335)
5336
5337xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005338 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005339 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005340 "test/f32-dwconv2d-chw.cc",
5341 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005342 "src/xnnpack/AlignedAllocator.h",
5343 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005344 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005345)
5346
5347xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005348 name = "f16_gavgpool_minmax_test",
5349 srcs = [
5350 "test/f16-gavgpool-minmax.cc",
5351 "test/gavgpool-microkernel-tester.h",
5352 "src/xnnpack/AlignedAllocator.h",
5353 ] + MICROKERNEL_TEST_HDRS,
5354 deps = MICROKERNEL_TEST_DEPS,
5355)
5356
5357xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005358 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005359 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005360 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005361 "test/gavgpool-microkernel-tester.h",
5362 "src/xnnpack/AlignedAllocator.h",
5363 ] + MICROKERNEL_TEST_HDRS,
5364 deps = MICROKERNEL_TEST_DEPS,
5365)
5366
5367xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005368 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005369 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005370 "test/f32-gavgpool-cw.cc",
5371 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005372 "src/xnnpack/AlignedAllocator.h",
5373 ] + MICROKERNEL_TEST_HDRS,
5374 deps = MICROKERNEL_TEST_DEPS,
5375)
5376
5377xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005378 name = "f32_gemm_test",
5379 srcs = [
5380 "test/f32-gemm.cc",
5381 "test/gemm-microkernel-tester.h",
5382 "src/xnnpack/AlignedAllocator.h",
5383 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005384 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005385)
5386
5387xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005388 name = "f32_gemm_relu_test",
5389 srcs = [
5390 "test/f32-gemm-relu.cc",
5391 "test/gemm-microkernel-tester.h",
5392 "src/xnnpack/AlignedAllocator.h",
5393 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005394 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005395)
5396
5397xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005398 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005399 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005400 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005401 "test/gemm-microkernel-tester.h",
5402 "src/xnnpack/AlignedAllocator.h",
5403 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005404 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005405)
5406
5407xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005408 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005409 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005410 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005411 "test/gemm-microkernel-tester.h",
5412 "src/xnnpack/AlignedAllocator.h",
5413 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005414 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005415)
5416
5417xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005418 name = "f16_hswish_test",
5419 srcs = [
5420 "test/f16-hswish.cc",
5421 "test/hswish-microkernel-tester.h",
5422 ] + MICROKERNEL_TEST_HDRS,
5423 deps = MICROKERNEL_TEST_DEPS,
5424)
5425
5426xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005427 name = "f32_hswish_test",
5428 srcs = [
5429 "test/f32-hswish.cc",
5430 "test/hswish-microkernel-tester.h",
5431 ] + MICROKERNEL_TEST_HDRS,
5432 deps = MICROKERNEL_TEST_DEPS,
5433)
5434
5435xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005436 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005437 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005438 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005439 "test/maxpool-microkernel-tester.h",
5440 ] + MICROKERNEL_TEST_HDRS,
5441 deps = MICROKERNEL_TEST_DEPS,
5442)
5443
5444xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005445 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005446 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005447 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005448 "test/avgpool-microkernel-tester.h",
5449 "src/xnnpack/AlignedAllocator.h",
5450 ] + MICROKERNEL_TEST_HDRS,
5451 deps = MICROKERNEL_TEST_DEPS,
5452)
5453
5454xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005455 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005456 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005457 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005458 "test/gemm-microkernel-tester.h",
5459 "src/xnnpack/AlignedAllocator.h",
5460 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005461 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005462)
5463
5464xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005465 name = "f16_prelu_test",
5466 srcs = [
5467 "test/f16-prelu.cc",
5468 "test/prelu-microkernel-tester.h",
5469 "src/xnnpack/AlignedAllocator.h",
5470 ] + MICROKERNEL_TEST_HDRS,
5471 deps = MICROKERNEL_TEST_DEPS,
5472)
5473
5474xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005475 name = "f32_prelu_test",
5476 srcs = [
5477 "test/f32-prelu.cc",
5478 "test/prelu-microkernel-tester.h",
5479 "src/xnnpack/AlignedAllocator.h",
5480 ] + MICROKERNEL_TEST_HDRS,
5481 deps = MICROKERNEL_TEST_DEPS,
5482)
5483
5484xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005485 name = "f32_raddexpminusmax_test",
5486 srcs = [
5487 "test/f32-raddexpminusmax.cc",
5488 "test/raddexpminusmax-microkernel-tester.h",
5489 ] + MICROKERNEL_TEST_HDRS,
5490 deps = MICROKERNEL_TEST_DEPS,
5491)
5492
5493xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005494 name = "f32_raddextexp_test",
5495 srcs = [
5496 "test/f32-raddextexp.cc",
5497 "test/raddextexp-microkernel-tester.h",
5498 ] + MICROKERNEL_TEST_HDRS,
5499 deps = MICROKERNEL_TEST_DEPS,
5500)
5501
5502xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005503 name = "f32_raddstoreexpminusmax_test",
5504 srcs = [
5505 "test/f32-raddstoreexpminusmax.cc",
5506 "test/raddstoreexpminusmax-microkernel-tester.h",
5507 ] + MICROKERNEL_TEST_HDRS,
5508 deps = MICROKERNEL_TEST_DEPS,
5509)
5510
5511xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005512 name = "f32_relu_test",
5513 srcs = [
5514 "test/f32-relu.cc",
5515 "test/vunary-microkernel-tester.h",
5516 ] + MICROKERNEL_TEST_HDRS,
5517 deps = MICROKERNEL_TEST_DEPS,
5518)
5519
5520xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005521 name = "f32_rmax_test",
5522 srcs = [
5523 "test/f32-rmax.cc",
5524 "test/rmax-microkernel-tester.h",
5525 ] + MICROKERNEL_TEST_HDRS,
5526 deps = MICROKERNEL_TEST_DEPS,
5527)
5528
5529xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005530 name = "f32_sigmoid_test",
5531 srcs = [
5532 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005533 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005534 ] + MICROKERNEL_TEST_HDRS,
5535 deps = MICROKERNEL_TEST_DEPS,
5536)
5537
5538xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005539 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005540 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005541 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005542 "test/spmm-microkernel-tester.h",
5543 "src/xnnpack/AlignedAllocator.h",
5544 ] + MICROKERNEL_TEST_HDRS,
5545 deps = MICROKERNEL_TEST_DEPS,
5546)
5547
5548xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005549 name = "f16_vadd_minmax_test",
5550 srcs = [
5551 "test/f16-vadd-minmax.cc",
5552 "test/vbinary-microkernel-tester.h",
5553 ] + MICROKERNEL_TEST_HDRS,
5554 deps = MICROKERNEL_TEST_DEPS,
5555)
5556
5557xnnpack_unit_test(
5558 name = "f16_vaddc_minmax_test",
5559 srcs = [
5560 "test/f16-vaddc-minmax.cc",
5561 "test/vbinaryc-microkernel-tester.h",
5562 ] + MICROKERNEL_TEST_HDRS,
5563 deps = MICROKERNEL_TEST_DEPS,
5564)
5565
5566xnnpack_unit_test(
5567 name = "f16_vdiv_minmax_test",
5568 srcs = [
5569 "test/f16-vdiv-minmax.cc",
5570 "test/vbinary-microkernel-tester.h",
5571 ] + MICROKERNEL_TEST_HDRS,
5572 deps = MICROKERNEL_TEST_DEPS,
5573)
5574
5575xnnpack_unit_test(
5576 name = "f16_vdivc_minmax_test",
5577 srcs = [
5578 "test/f16-vdivc-minmax.cc",
5579 "test/vbinaryc-microkernel-tester.h",
5580 ] + MICROKERNEL_TEST_HDRS,
5581 deps = MICROKERNEL_TEST_DEPS,
5582)
5583
5584xnnpack_unit_test(
5585 name = "f16_vrdivc_minmax_test",
5586 srcs = [
5587 "test/f16-vrdivc-minmax.cc",
5588 "test/vbinaryc-microkernel-tester.h",
5589 ] + MICROKERNEL_TEST_HDRS,
5590 deps = MICROKERNEL_TEST_DEPS,
5591)
5592
5593xnnpack_unit_test(
5594 name = "f16_vmax_test",
5595 srcs = [
5596 "test/f16-vmax.cc",
5597 "test/vbinary-microkernel-tester.h",
5598 ] + MICROKERNEL_TEST_HDRS,
5599 deps = MICROKERNEL_TEST_DEPS,
5600)
5601
5602xnnpack_unit_test(
5603 name = "f16_vmaxc_test",
5604 srcs = [
5605 "test/f16-vmaxc.cc",
5606 "test/vbinaryc-microkernel-tester.h",
5607 ] + MICROKERNEL_TEST_HDRS,
5608 deps = MICROKERNEL_TEST_DEPS,
5609)
5610
5611xnnpack_unit_test(
5612 name = "f16_vmin_test",
5613 srcs = [
5614 "test/f16-vmin.cc",
5615 "test/vbinary-microkernel-tester.h",
5616 ] + MICROKERNEL_TEST_HDRS,
5617 deps = MICROKERNEL_TEST_DEPS,
5618)
5619
5620xnnpack_unit_test(
5621 name = "f16_vminc_test",
5622 srcs = [
5623 "test/f16-vminc.cc",
5624 "test/vbinaryc-microkernel-tester.h",
5625 ] + MICROKERNEL_TEST_HDRS,
5626 deps = MICROKERNEL_TEST_DEPS,
5627)
5628
5629xnnpack_unit_test(
5630 name = "f16_vmul_minmax_test",
5631 srcs = [
5632 "test/f16-vmul-minmax.cc",
5633 "test/vbinary-microkernel-tester.h",
5634 ] + MICROKERNEL_TEST_HDRS,
5635 deps = MICROKERNEL_TEST_DEPS,
5636)
5637
5638xnnpack_unit_test(
5639 name = "f16_vmulc_minmax_test",
5640 srcs = [
5641 "test/f16-vmulc-minmax.cc",
5642 "test/vbinaryc-microkernel-tester.h",
5643 ] + MICROKERNEL_TEST_HDRS,
5644 deps = MICROKERNEL_TEST_DEPS,
5645)
5646
5647xnnpack_unit_test(
5648 name = "f16_vsub_minmax_test",
5649 srcs = [
5650 "test/f16-vsub-minmax.cc",
5651 "test/vbinary-microkernel-tester.h",
5652 ] + MICROKERNEL_TEST_HDRS,
5653 deps = MICROKERNEL_TEST_DEPS,
5654)
5655
5656xnnpack_unit_test(
5657 name = "f16_vsubc_minmax_test",
5658 srcs = [
5659 "test/f16-vsubc-minmax.cc",
5660 "test/vbinaryc-microkernel-tester.h",
5661 ] + MICROKERNEL_TEST_HDRS,
5662 deps = MICROKERNEL_TEST_DEPS,
5663)
5664
5665xnnpack_unit_test(
5666 name = "f16_vrsubc_minmax_test",
5667 srcs = [
5668 "test/f16-vrsubc-minmax.cc",
5669 "test/vbinaryc-microkernel-tester.h",
5670 ] + MICROKERNEL_TEST_HDRS,
5671 deps = MICROKERNEL_TEST_DEPS,
5672)
5673
5674xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005675 name = "f32_vabs_test",
5676 srcs = [
5677 "test/f32-vabs.cc",
5678 "test/vunary-microkernel-tester.h",
5679 ] + MICROKERNEL_TEST_HDRS,
5680 deps = MICROKERNEL_TEST_DEPS,
5681)
5682
5683xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005684 name = "f32_vadd_test",
5685 srcs = [
5686 "test/f32-vadd.cc",
5687 "test/vbinary-microkernel-tester.h",
5688 ] + MICROKERNEL_TEST_HDRS,
5689 deps = MICROKERNEL_TEST_DEPS,
5690)
5691
5692xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005693 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005694 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005695 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005696 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005697 ] + MICROKERNEL_TEST_HDRS,
5698 deps = MICROKERNEL_TEST_DEPS,
5699)
5700
5701xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005702 name = "f32_vadd_relu_test",
5703 srcs = [
5704 "test/f32-vadd-relu.cc",
5705 "test/vbinary-microkernel-tester.h",
5706 ] + MICROKERNEL_TEST_HDRS,
5707 deps = MICROKERNEL_TEST_DEPS,
5708)
5709
5710xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005711 name = "f32_vaddc_test",
5712 srcs = [
5713 "test/f32-vaddc.cc",
5714 "test/vbinaryc-microkernel-tester.h",
5715 ] + MICROKERNEL_TEST_HDRS,
5716 deps = MICROKERNEL_TEST_DEPS,
5717)
5718
5719xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005720 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005721 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005722 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005723 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005724 ] + MICROKERNEL_TEST_HDRS,
5725 deps = MICROKERNEL_TEST_DEPS,
5726)
5727
5728xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005729 name = "f32_vaddc_relu_test",
5730 srcs = [
5731 "test/f32-vaddc-relu.cc",
5732 "test/vbinaryc-microkernel-tester.h",
5733 ] + MICROKERNEL_TEST_HDRS,
5734 deps = MICROKERNEL_TEST_DEPS,
5735)
5736
5737xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005738 name = "f32_vdiv_test",
5739 srcs = [
5740 "test/f32-vdiv.cc",
5741 "test/vbinary-microkernel-tester.h",
5742 ] + MICROKERNEL_TEST_HDRS,
5743 deps = MICROKERNEL_TEST_DEPS,
5744)
5745
5746xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005747 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005748 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005749 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005750 "test/vbinary-microkernel-tester.h",
5751 ] + MICROKERNEL_TEST_HDRS,
5752 deps = MICROKERNEL_TEST_DEPS,
5753)
5754
5755xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005756 name = "f32_vdiv_relu_test",
5757 srcs = [
5758 "test/f32-vdiv-relu.cc",
5759 "test/vbinary-microkernel-tester.h",
5760 ] + MICROKERNEL_TEST_HDRS,
5761 deps = MICROKERNEL_TEST_DEPS,
5762)
5763
5764xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005765 name = "f32_vdivc_test",
5766 srcs = [
5767 "test/f32-vdivc.cc",
5768 "test/vbinaryc-microkernel-tester.h",
5769 ] + MICROKERNEL_TEST_HDRS,
5770 deps = MICROKERNEL_TEST_DEPS,
5771)
5772
5773xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005774 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005775 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005776 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005777 "test/vbinaryc-microkernel-tester.h",
5778 ] + MICROKERNEL_TEST_HDRS,
5779 deps = MICROKERNEL_TEST_DEPS,
5780)
5781
5782xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005783 name = "f32_vdivc_relu_test",
5784 srcs = [
5785 "test/f32-vdivc-relu.cc",
5786 "test/vbinaryc-microkernel-tester.h",
5787 ] + MICROKERNEL_TEST_HDRS,
5788 deps = MICROKERNEL_TEST_DEPS,
5789)
5790
5791xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005792 name = "f32_vrdivc_test",
5793 srcs = [
5794 "test/f32-vrdivc.cc",
5795 "test/vbinaryc-microkernel-tester.h",
5796 ] + MICROKERNEL_TEST_HDRS,
5797 deps = MICROKERNEL_TEST_DEPS,
5798)
5799
5800xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005801 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005802 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005803 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005804 "test/vbinaryc-microkernel-tester.h",
5805 ] + MICROKERNEL_TEST_HDRS,
5806 deps = MICROKERNEL_TEST_DEPS,
5807)
5808
5809xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005810 name = "f32_vrdivc_relu_test",
5811 srcs = [
5812 "test/f32-vrdivc-relu.cc",
5813 "test/vbinaryc-microkernel-tester.h",
5814 ] + MICROKERNEL_TEST_HDRS,
5815 deps = MICROKERNEL_TEST_DEPS,
5816)
5817
5818xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005819 name = "f32_vmax_test",
5820 srcs = [
5821 "test/f32-vmax.cc",
5822 "test/vbinary-microkernel-tester.h",
5823 ] + MICROKERNEL_TEST_HDRS,
5824 deps = MICROKERNEL_TEST_DEPS,
5825)
5826
5827xnnpack_unit_test(
5828 name = "f32_vmaxc_test",
5829 srcs = [
5830 "test/f32-vmaxc.cc",
5831 "test/vbinaryc-microkernel-tester.h",
5832 ] + MICROKERNEL_TEST_HDRS,
5833 deps = MICROKERNEL_TEST_DEPS,
5834)
5835
5836xnnpack_unit_test(
5837 name = "f32_vmin_test",
5838 srcs = [
5839 "test/f32-vmin.cc",
5840 "test/vbinary-microkernel-tester.h",
5841 ] + MICROKERNEL_TEST_HDRS,
5842 deps = MICROKERNEL_TEST_DEPS,
5843)
5844
5845xnnpack_unit_test(
5846 name = "f32_vminc_test",
5847 srcs = [
5848 "test/f32-vminc.cc",
5849 "test/vbinaryc-microkernel-tester.h",
5850 ] + MICROKERNEL_TEST_HDRS,
5851 deps = MICROKERNEL_TEST_DEPS,
5852)
5853
5854xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005855 name = "f32_vmul_test",
5856 srcs = [
5857 "test/f32-vmul.cc",
5858 "test/vbinary-microkernel-tester.h",
5859 ] + MICROKERNEL_TEST_HDRS,
5860 deps = MICROKERNEL_TEST_DEPS,
5861)
5862
5863xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005864 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005865 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005866 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005867 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005868 ] + MICROKERNEL_TEST_HDRS,
5869 deps = MICROKERNEL_TEST_DEPS,
5870)
5871
5872xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005873 name = "f32_vmul_relu_test",
5874 srcs = [
5875 "test/f32-vmul-relu.cc",
5876 "test/vbinary-microkernel-tester.h",
5877 ] + MICROKERNEL_TEST_HDRS,
5878 deps = MICROKERNEL_TEST_DEPS,
5879)
5880
5881xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005882 name = "f32_vmulc_test",
5883 srcs = [
5884 "test/f32-vmulc.cc",
5885 "test/vbinaryc-microkernel-tester.h",
5886 ] + MICROKERNEL_TEST_HDRS,
5887 deps = MICROKERNEL_TEST_DEPS,
5888)
5889
5890xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005891 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005892 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005893 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005894 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005895 ] + MICROKERNEL_TEST_HDRS,
5896 deps = MICROKERNEL_TEST_DEPS,
5897)
5898
5899xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005900 name = "f32_vmulc_relu_test",
5901 srcs = [
5902 "test/f32-vmulc-relu.cc",
5903 "test/vbinaryc-microkernel-tester.h",
5904 ] + MICROKERNEL_TEST_HDRS,
5905 deps = MICROKERNEL_TEST_DEPS,
5906)
5907
5908xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005909 name = "f16_vmulcaddc_minmax_test",
5910 srcs = [
5911 "test/f16-vmulcaddc-minmax.cc",
5912 "test/vmulcaddc-microkernel-tester.h",
5913 "src/xnnpack/AlignedAllocator.h",
5914 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005915 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005916)
5917
5918xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005919 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005920 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005921 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005922 "test/vmulcaddc-microkernel-tester.h",
5923 "src/xnnpack/AlignedAllocator.h",
5924 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005925 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005926)
5927
5928xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005929 name = "f32_vlrelu_test",
5930 srcs = [
5931 "test/f32-vlrelu.cc",
5932 "test/vunary-microkernel-tester.h",
5933 ] + MICROKERNEL_TEST_HDRS,
5934 deps = MICROKERNEL_TEST_DEPS,
5935)
5936
5937xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005938 name = "f32_vneg_test",
5939 srcs = [
5940 "test/f32-vneg.cc",
5941 "test/vunary-microkernel-tester.h",
5942 ] + MICROKERNEL_TEST_HDRS,
5943 deps = MICROKERNEL_TEST_DEPS,
5944)
5945
5946xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005947 name = "f32_vrndne_test",
5948 srcs = [
5949 "test/f32-vrndne.cc",
5950 "test/vunary-microkernel-tester.h",
5951 ] + MICROKERNEL_TEST_HDRS,
5952 deps = MICROKERNEL_TEST_DEPS,
5953)
5954
5955xnnpack_unit_test(
5956 name = "f32_vrndz_test",
5957 srcs = [
5958 "test/f32-vrndz.cc",
5959 "test/vunary-microkernel-tester.h",
5960 ] + MICROKERNEL_TEST_HDRS,
5961 deps = MICROKERNEL_TEST_DEPS,
5962)
5963
5964xnnpack_unit_test(
5965 name = "f32_vrndu_test",
5966 srcs = [
5967 "test/f32-vrndu.cc",
5968 "test/vunary-microkernel-tester.h",
5969 ] + MICROKERNEL_TEST_HDRS,
5970 deps = MICROKERNEL_TEST_DEPS,
5971)
5972
5973xnnpack_unit_test(
5974 name = "f32_vrndd_test",
5975 srcs = [
5976 "test/f32-vrndd.cc",
5977 "test/vunary-microkernel-tester.h",
5978 ] + MICROKERNEL_TEST_HDRS,
5979 deps = MICROKERNEL_TEST_DEPS,
5980)
5981
5982xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005983 name = "f32_vscale_test",
5984 srcs = [
5985 "test/f32-vscale.cc",
5986 "test/vscale-microkernel-tester.h",
5987 ] + MICROKERNEL_TEST_HDRS,
5988 deps = MICROKERNEL_TEST_DEPS,
5989)
5990
5991xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005992 name = "f32_vscaleexpminusmax_test",
5993 srcs = [
5994 "test/f32-vscaleexpminusmax.cc",
5995 "test/vscaleexpminusmax-microkernel-tester.h",
5996 ] + MICROKERNEL_TEST_HDRS,
5997 deps = MICROKERNEL_TEST_DEPS,
5998)
5999
6000xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006001 name = "f32_vscaleextexp_test",
6002 srcs = [
6003 "test/f32-vscaleextexp.cc",
6004 "test/vscaleextexp-microkernel-tester.h",
6005 ] + MICROKERNEL_TEST_HDRS,
6006 deps = MICROKERNEL_TEST_DEPS,
6007)
6008
6009xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006010 name = "f32_vsqr_test",
6011 srcs = [
6012 "test/f32-vsqr.cc",
6013 "test/vunary-microkernel-tester.h",
6014 ] + MICROKERNEL_TEST_HDRS,
6015 deps = MICROKERNEL_TEST_DEPS,
6016)
6017
6018xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006019 name = "f32_vsqrdiff_test",
6020 srcs = [
6021 "test/f32-vsqrdiff.cc",
6022 "test/vbinary-microkernel-tester.h",
6023 ] + MICROKERNEL_TEST_HDRS,
6024 deps = MICROKERNEL_TEST_DEPS,
6025)
6026
6027xnnpack_unit_test(
6028 name = "f32_vsqrdiffc_test",
6029 srcs = [
6030 "test/f32-vsqrdiffc.cc",
6031 "test/vbinaryc-microkernel-tester.h",
6032 ] + MICROKERNEL_TEST_HDRS,
6033 deps = MICROKERNEL_TEST_DEPS,
6034)
6035
6036xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006037 name = "f32_vsqrt_test",
6038 srcs = [
6039 "test/f32-vsqrt.cc",
6040 "test/vunary-microkernel-tester.h",
6041 ] + MICROKERNEL_TEST_HDRS,
6042 deps = MICROKERNEL_TEST_DEPS,
6043)
6044
6045xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006046 name = "f32_vsub_test",
6047 srcs = [
6048 "test/f32-vsub.cc",
6049 "test/vbinary-microkernel-tester.h",
6050 ] + MICROKERNEL_TEST_HDRS,
6051 deps = MICROKERNEL_TEST_DEPS,
6052)
6053
6054xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006055 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006056 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006057 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006058 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006059 ] + MICROKERNEL_TEST_HDRS,
6060 deps = MICROKERNEL_TEST_DEPS,
6061)
6062
6063xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006064 name = "f32_vsub_relu_test",
6065 srcs = [
6066 "test/f32-vsub-relu.cc",
6067 "test/vbinary-microkernel-tester.h",
6068 ] + MICROKERNEL_TEST_HDRS,
6069 deps = MICROKERNEL_TEST_DEPS,
6070)
6071
6072xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006073 name = "f32_vsubc_test",
6074 srcs = [
6075 "test/f32-vsubc.cc",
6076 "test/vbinaryc-microkernel-tester.h",
6077 ] + MICROKERNEL_TEST_HDRS,
6078 deps = MICROKERNEL_TEST_DEPS,
6079)
6080
6081xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006082 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006083 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006084 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006085 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006086 ] + MICROKERNEL_TEST_HDRS,
6087 deps = MICROKERNEL_TEST_DEPS,
6088)
6089
6090xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006091 name = "f32_vsubc_relu_test",
6092 srcs = [
6093 "test/f32-vsubc-relu.cc",
6094 "test/vbinaryc-microkernel-tester.h",
6095 ] + MICROKERNEL_TEST_HDRS,
6096 deps = MICROKERNEL_TEST_DEPS,
6097)
6098
6099xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006100 name = "f32_vrsubc_test",
6101 srcs = [
6102 "test/f32-vrsubc.cc",
6103 "test/vbinaryc-microkernel-tester.h",
6104 ] + MICROKERNEL_TEST_HDRS,
6105 deps = MICROKERNEL_TEST_DEPS,
6106)
6107
6108xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006109 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006110 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006111 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006112 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006113 ] + MICROKERNEL_TEST_HDRS,
6114 deps = MICROKERNEL_TEST_DEPS,
6115)
6116
6117xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006118 name = "f32_vrsubc_relu_test",
6119 srcs = [
6120 "test/f32-vrsubc-relu.cc",
6121 "test/vbinaryc-microkernel-tester.h",
6122 ] + MICROKERNEL_TEST_HDRS,
6123 deps = MICROKERNEL_TEST_DEPS,
6124)
6125
6126xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006127 name = "qs8_dwconv_minmax_test",
6128 srcs = [
6129 "test/qs8-dwconv-minmax.cc",
6130 "test/dwconv-microkernel-tester.h",
6131 "src/xnnpack/AlignedAllocator.h",
6132 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6133 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6134)
6135
6136xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006137 name = "qs8_gavgpool_minmax_test",
6138 srcs = [
6139 "test/qs8-gavgpool-minmax.cc",
6140 "test/gavgpool-microkernel-tester.h",
6141 "src/xnnpack/AlignedAllocator.h",
6142 ] + MICROKERNEL_TEST_HDRS,
6143 deps = MICROKERNEL_TEST_DEPS,
6144)
6145
6146xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006147 name = "qs8_gemm_minmax_test",
6148 srcs = [
6149 "test/qs8-gemm-minmax.cc",
6150 "test/gemm-microkernel-tester.h",
6151 "src/xnnpack/AlignedAllocator.h",
6152 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6153 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6154)
6155
6156xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006157 name = "qs8_igemm_minmax_test",
6158 srcs = [
6159 "test/qs8-igemm-minmax.cc",
6160 "test/gemm-microkernel-tester.h",
6161 "src/xnnpack/AlignedAllocator.h",
6162 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6163 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6164)
6165
6166xnnpack_unit_test(
6167 name = "qs8_requantization_test",
6168 srcs = [
6169 "src/xnnpack/requantization-stubs.h",
6170 "test/qs8-requantization.cc",
6171 "test/requantization-tester.h",
6172 ] + MICROKERNEL_TEST_HDRS,
6173 deps = MICROKERNEL_TEST_DEPS,
6174)
6175
6176xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006177 name = "qs8_vadd_minmax_test",
6178 srcs = [
6179 "test/qs8-vadd-minmax.cc",
6180 "test/vadd-microkernel-tester.h",
6181 ] + MICROKERNEL_TEST_HDRS,
6182 deps = MICROKERNEL_TEST_DEPS,
6183)
6184
6185xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006186 name = "qs8_vaddc_minmax_test",
6187 srcs = [
6188 "test/qs8-vaddc-minmax.cc",
6189 "test/vaddc-microkernel-tester.h",
6190 ] + MICROKERNEL_TEST_HDRS,
6191 deps = MICROKERNEL_TEST_DEPS,
6192)
6193
6194xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006195 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006196 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006197 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006198 "test/avgpool-microkernel-tester.h",
6199 "src/xnnpack/AlignedAllocator.h",
6200 ] + MICROKERNEL_TEST_HDRS,
6201 deps = MICROKERNEL_TEST_DEPS,
6202)
6203
6204xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006205 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006206 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006207 "test/qu8-dwconv-minmax.cc",
6208 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006209 "src/xnnpack/AlignedAllocator.h",
6210 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006211 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006212)
6213
6214xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006215 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006216 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006217 "test/qu8-igemm-minmax.cc",
6218 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006219 "src/xnnpack/AlignedAllocator.h",
6220 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006221 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006222)
6223
6224xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006225 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006226 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006227 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006228 "test/gavgpool-microkernel-tester.h",
6229 "src/xnnpack/AlignedAllocator.h",
6230 ] + MICROKERNEL_TEST_HDRS,
6231 deps = MICROKERNEL_TEST_DEPS,
6232)
6233
6234xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006235 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006236 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006237 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006238 "test/gemm-microkernel-tester.h",
6239 "src/xnnpack/AlignedAllocator.h",
6240 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006241 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006242)
6243
6244xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006245 name = "qu8_requantization_test",
6246 srcs = [
6247 "src/xnnpack/requantization-stubs.h",
6248 "test/qu8-requantization.cc",
6249 "test/requantization-tester.h",
6250 ] + MICROKERNEL_TEST_HDRS,
6251 deps = MICROKERNEL_TEST_DEPS,
6252)
6253
6254xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006255 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006256 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006257 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006258 "test/vadd-microkernel-tester.h",
6259 ] + MICROKERNEL_TEST_HDRS,
6260 deps = MICROKERNEL_TEST_DEPS,
6261)
6262
6263xnnpack_unit_test(
6264 name = "u8_clamp_test",
6265 srcs = [
6266 "test/u8-clamp.cc",
6267 "test/clamp-microkernel-tester.h",
6268 ] + MICROKERNEL_TEST_HDRS,
6269 deps = MICROKERNEL_TEST_DEPS,
6270)
6271
6272xnnpack_unit_test(
6273 name = "u8_lut32norm_test",
6274 srcs = [
6275 "test/u8-lut32norm.cc",
6276 "test/lut-norm-microkernel-tester.h",
6277 ] + MICROKERNEL_TEST_HDRS,
6278 deps = MICROKERNEL_TEST_DEPS,
6279)
6280
6281xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006282 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006283 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006284 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006285 "test/maxpool-microkernel-tester.h",
6286 ] + MICROKERNEL_TEST_HDRS,
6287 deps = MICROKERNEL_TEST_DEPS,
6288)
6289
6290xnnpack_unit_test(
6291 name = "u8_rmax_test",
6292 srcs = [
6293 "test/u8-rmax.cc",
6294 "test/rmax-microkernel-tester.h",
6295 ] + MICROKERNEL_TEST_HDRS,
6296 deps = MICROKERNEL_TEST_DEPS,
6297)
6298
6299xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006300 name = "x32_fill_test",
6301 srcs = [
6302 "test/x32-fill.cc",
6303 "test/fill-microkernel-tester.h",
6304 ] + MICROKERNEL_TEST_HDRS,
6305 deps = MICROKERNEL_TEST_DEPS,
6306)
6307
6308xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006309 name = "x32_packx_test",
6310 srcs = [
6311 "test/x32-packx.cc",
6312 "test/pack-microkernel-tester.h",
6313 "src/xnnpack/AlignedAllocator.h",
6314 ] + MICROKERNEL_TEST_HDRS,
6315 deps = MICROKERNEL_TEST_DEPS,
6316)
6317
6318xnnpack_unit_test(
6319 name = "x32_pad_test",
6320 srcs = [
6321 "test/x32-pad.cc",
6322 "test/pad-microkernel-tester.h",
6323 ] + MICROKERNEL_TEST_HDRS,
6324 deps = MICROKERNEL_TEST_DEPS,
6325)
6326
6327xnnpack_unit_test(
6328 name = "x32_unpool_test",
6329 srcs = [
6330 "test/x32-unpool.cc",
6331 "test/unpool-microkernel-tester.h",
6332 ] + MICROKERNEL_TEST_HDRS,
6333 deps = MICROKERNEL_TEST_DEPS,
6334)
6335
6336xnnpack_unit_test(
6337 name = "x32_zip_test",
6338 srcs = [
6339 "test/x32-zip.cc",
6340 "test/zip-microkernel-tester.h",
6341 ] + MICROKERNEL_TEST_HDRS,
6342 deps = MICROKERNEL_TEST_DEPS,
6343)
6344
6345xnnpack_unit_test(
6346 name = "x8_lut_test",
6347 srcs = [
6348 "test/x8-lut.cc",
6349 "test/lut-microkernel-tester.h",
6350 ] + MICROKERNEL_TEST_HDRS,
6351 deps = MICROKERNEL_TEST_DEPS,
6352)
6353
6354xnnpack_unit_test(
6355 name = "x8_zip_test",
6356 srcs = [
6357 "test/x8-zip.cc",
6358 "test/zip-microkernel-tester.h",
6359 ] + MICROKERNEL_TEST_HDRS,
6360 deps = MICROKERNEL_TEST_DEPS,
6361)
6362
Marat Dukhan20c3b922020-03-10 03:45:06 -07006363########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006364
6365xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006366 name = "operator_size_test",
6367 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006368 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006369)
6370
Marat Dukhan20c3b922020-03-10 03:45:06 -07006371xnnpack_binary(
6372 name = "subgraph_size_test",
6373 srcs = ["test/subgraph-size.c"],
6374 deps = [":XNNPACK"],
6375)
6376
6377########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006378
6379xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006380 name = "abs_nc_test",
6381 srcs = [
6382 "test/abs-nc.cc",
6383 "test/abs-operator-tester.h",
6384 ],
6385 deps = OPERATOR_TEST_DEPS,
6386)
6387
6388xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006389 name = "add_nd_test",
6390 srcs = [
6391 "test/add-nd.cc",
6392 "test/binary-elementwise-operator-tester.h",
6393 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006394 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006395)
6396
6397xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006398 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006399 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006400 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006401 "test/argmax-pooling-operator-tester.h",
6402 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006403 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006404)
6405
6406xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006407 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006408 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006409 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006410 "test/average-pooling-operator-tester.h",
6411 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006412 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006413)
6414
6415xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006416 name = "bankers_rounding_nc_test",
6417 srcs = [
6418 "test/bankers-rounding-nc.cc",
6419 "test/bankers-rounding-operator-tester.h",
6420 ],
6421 deps = OPERATOR_TEST_DEPS,
6422)
6423
6424xnnpack_unit_test(
6425 name = "ceiling_nc_test",
6426 srcs = [
6427 "test/ceiling-nc.cc",
6428 "test/ceiling-operator-tester.h",
6429 ],
6430 deps = OPERATOR_TEST_DEPS,
6431)
6432
6433xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006434 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006435 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006436 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006437 "test/channel-shuffle-operator-tester.h",
6438 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006439 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006440)
6441
6442xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006443 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006444 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006445 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006446 "test/clamp-operator-tester.h",
6447 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006448 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006449)
6450
6451xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006452 name = "constant_pad_nd_test",
6453 srcs = [
6454 "test/constant-pad-nd.cc",
6455 "test/constant-pad-operator-tester.h",
6456 ],
6457 deps = OPERATOR_TEST_DEPS,
6458)
6459
6460xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006461 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006462 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006463 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006464 "test/convolution-operator-tester.h",
6465 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006466 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006467)
6468
6469xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006470 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006471 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006472 "test/convolution-nchw.cc",
6473 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006474 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006475 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006476)
6477
6478xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006479 name = "copy_nc_test",
6480 srcs = [
6481 "test/copy-nc.cc",
6482 "test/copy-operator-tester.h",
6483 ],
6484 deps = OPERATOR_TEST_DEPS,
6485)
6486
6487xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006488 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006489 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006490 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006491 "test/deconvolution-operator-tester.h",
6492 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006493 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006494)
6495
6496xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006497 name = "divide_nd_test",
6498 srcs = [
6499 "test/binary-elementwise-operator-tester.h",
6500 "test/divide-nd.cc",
6501 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006502 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006503)
6504
6505xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006506 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006507 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006508 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006509 "test/fully-connected-operator-tester.h",
6510 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006511 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006512)
6513
6514xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006515 name = "floor_nc_test",
6516 srcs = [
6517 "test/floor-nc.cc",
6518 "test/floor-operator-tester.h",
6519 ],
6520 deps = OPERATOR_TEST_DEPS,
6521)
6522
6523xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006524 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006525 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006526 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006527 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006528 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006529 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006530)
6531
6532xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006533 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006534 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006535 "test/global-average-pooling-ncw.cc",
6536 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006537 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006538 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006539)
6540
6541xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006542 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006543 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006544 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006545 "test/hardswish-operator-tester.h",
6546 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006547 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006548)
6549
6550xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006551 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006552 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006553 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006554 "test/leaky-relu-operator-tester.h",
6555 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006556 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006557)
6558
6559xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006560 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006561 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006562 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006563 "test/max-pooling-operator-tester.h",
6564 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006565 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006566)
6567
6568xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006569 name = "maximum_nd_test",
6570 srcs = [
6571 "test/binary-elementwise-operator-tester.h",
6572 "test/maximum-nd.cc",
6573 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006574 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006575)
6576
6577xnnpack_unit_test(
6578 name = "minimum_nd_test",
6579 srcs = [
6580 "test/binary-elementwise-operator-tester.h",
6581 "test/minimum-nd.cc",
6582 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006583 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006584)
6585
6586xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006587 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006588 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006589 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006590 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006591 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006592 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006593)
6594
6595xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006596 name = "negate_nc_test",
6597 srcs = [
6598 "test/negate-nc.cc",
6599 "test/negate-operator-tester.h",
6600 ],
6601 deps = OPERATOR_TEST_DEPS,
6602)
6603
6604xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006605 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006606 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006607 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006608 "test/prelu-operator-tester.h",
6609 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006610 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006611)
6612
6613xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006614 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006615 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006616 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006617 "test/resize-bilinear-operator-tester.h",
6618 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006619 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006620)
6621
6622xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07006623 name = "resize_bilinear_nchw_test",
6624 srcs = [
6625 "test/resize-bilinear-nchw.cc",
6626 "test/resize-bilinear-operator-tester.h",
6627 ] + OPERATOR_TEST_PARAMS_HDRS,
6628 deps = OPERATOR_TEST_DEPS,
6629)
6630
6631xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006632 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006633 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006634 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006635 "test/sigmoid-operator-tester.h",
6636 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006637 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006638)
6639
6640xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006641 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006642 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006643 "test/softmax-nc.cc",
6644 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006645 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006646 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006647)
6648
6649xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006650 name = "square_nc_test",
6651 srcs = [
6652 "test/square-nc.cc",
6653 "test/square-operator-tester.h",
6654 ],
6655 deps = OPERATOR_TEST_DEPS,
6656)
6657
6658xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006659 name = "square_root_nc_test",
6660 srcs = [
6661 "test/square-root-nc.cc",
6662 "test/square-root-operator-tester.h",
6663 ],
6664 deps = OPERATOR_TEST_DEPS,
6665)
6666
6667xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006668 name = "squared_difference_nd_test",
6669 srcs = [
6670 "test/binary-elementwise-operator-tester.h",
6671 "test/squared-difference-nd.cc",
6672 ],
6673 deps = OPERATOR_TEST_DEPS,
6674)
6675
6676xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006677 name = "subtract_nd_test",
6678 srcs = [
6679 "test/binary-elementwise-operator-tester.h",
6680 "test/subtract-nd.cc",
6681 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006682 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006683)
6684
6685xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006686 name = "truncation_nc_test",
6687 srcs = [
6688 "test/truncation-nc.cc",
6689 "test/truncation-operator-tester.h",
6690 ],
6691 deps = OPERATOR_TEST_DEPS,
6692)
6693
6694xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006695 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006696 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006697 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006698 "test/unpooling-operator-tester.h",
6699 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006700 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006701)
6702
Chao Mei6ddfc602020-05-13 22:29:36 -07006703############################### Misc unit tests ###############################
6704
6705xnnpack_unit_test(
6706 name = "memory_planner_test",
6707 srcs = [
6708 "test/memory-planner-test.cc",
6709 ],
6710 deps = [
6711 ":XNNPACK",
6712 ":memory_planner",
6713 ],
6714)
6715
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07006716xnnpack_unit_test(
6717 name = "subgraph_nchw_test",
6718 srcs = [
6719 "src/xnnpack/subgraph.h",
6720 "test/subgraph-nchw.cc",
6721 "test/subgraph-tester.h",
6722 ],
6723 deps = [
6724 ":XNNPACK",
6725 ],
6726)
6727
Marat Dukhan08c4a432019-10-03 09:29:21 -07006728############################# Build configurations #############################
6729
Marat Dukhanb8642352019-10-30 15:43:02 -07006730# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006731config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006732 name = "xnn_enable_assembly_explicit_true",
6733 define_values = {"xnn_enable_assembly": "true"},
6734)
6735
6736# Disables usage of assembly kernels.
6737config_setting(
6738 name = "xnn_enable_assembly_explicit_false",
6739 define_values = {"xnn_enable_assembly": "false"},
6740)
6741
Marat Dukhan9de90e02020-06-18 16:04:12 -07006742# Enables usage of sparse inference.
6743config_setting(
6744 name = "xnn_enable_sparse_explicit_true",
6745 define_values = {"xnn_enable_sparse": "true"},
6746)
6747
6748# Disables usage of sparse inference.
6749config_setting(
6750 name = "xnn_enable_sparse_explicit_false",
6751 define_values = {"xnn_enable_sparse": "false"},
6752)
6753
Marat Dukhan05702cf2020-03-26 15:41:33 -07006754# Disables usage of HMP-aware optimizations.
6755config_setting(
6756 name = "xnn_enable_hmp_explicit_false",
6757 define_values = {"xnn_enable_hmp": "false"},
6758)
6759
Chao Mei6ddfc602020-05-13 22:29:36 -07006760# Enable usage of optimized memory allocation
6761config_setting(
6762 name = "xnn_enable_memopt_explicit_true",
6763 define_values = {"xnnpack_enable_memopt": "true"},
6764)
6765
6766# Disable usage of optimized memory allocation
6767config_setting(
6768 name = "xnn_enable_memopt_explicit_false",
6769 define_values = {"xnnpack_enable_memopt": "false"},
6770)
6771
Marat Dukhanb8642352019-10-30 15:43:02 -07006772# Builds with -c dbg
6773config_setting(
6774 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006775 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006776 "compilation_mode": "dbg",
6777 },
6778)
6779
6780# Builds with -c opt
6781config_setting(
6782 name = "optimized_build",
6783 values = {
6784 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006785 },
6786)
6787
6788config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006789 name = "linux_k8",
6790 values = {"cpu": "k8"},
6791)
6792
6793config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006794 name = "linux_arm",
6795 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006796)
6797
6798config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006799 name = "linux_armeabi",
6800 values = {"cpu": "armeabi"},
6801)
6802
6803config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006804 name = "linux_armhf",
6805 values = {"cpu": "armhf"},
6806)
6807
6808config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006809 name = "linux_armv7a",
6810 values = {"cpu": "armv7a"},
6811)
6812
6813config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006814 name = "linux_aarch64",
6815 values = {"cpu": "aarch64"},
6816)
6817
6818config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006819 name = "android",
6820 values = {"crosstool_top": "//external:android/crosstool"},
6821)
6822
6823config_setting(
6824 name = "android_armv7",
6825 values = {
6826 "crosstool_top": "//external:android/crosstool",
6827 "cpu": "armeabi-v7a",
6828 },
6829)
6830
6831config_setting(
6832 name = "android_arm64",
6833 values = {
6834 "crosstool_top": "//external:android/crosstool",
6835 "cpu": "arm64-v8a",
6836 },
6837)
6838
6839config_setting(
6840 name = "android_x86",
6841 values = {
6842 "crosstool_top": "//external:android/crosstool",
6843 "cpu": "x86",
6844 },
6845)
6846
6847config_setting(
6848 name = "android_x86_64",
6849 values = {
6850 "crosstool_top": "//external:android/crosstool",
6851 "cpu": "x86_64",
6852 },
6853)
6854
6855config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006856 name = "windows_x86_64",
6857 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006858)
6859
6860config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006861 name = "windows_x86_64_clang",
6862 values = {
6863 "compiler": "clang-cl",
6864 "cpu": "x64_windows",
6865 },
6866)
6867
6868config_setting(
6869 name = "windows_x86_64_mingw",
6870 values = {
6871 "compiler": "mingw-gcc",
6872 "cpu": "x64_windows",
6873 },
6874)
6875
6876config_setting(
6877 name = "windows_x86_64_msys",
6878 values = {
6879 "compiler": "msys-gcc",
6880 "cpu": "x64_windows",
6881 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006882)
6883
6884config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006885 name = "macos_x86_64",
6886 values = {
6887 "apple_platform_type": "macos",
6888 "cpu": "darwin",
6889 },
6890)
6891
6892config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006893 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006894 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006895)
6896
6897config_setting(
6898 name = "emscripten_wasm",
6899 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006900 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006901 "cpu": "wasm",
6902 },
6903)
6904
6905config_setting(
6906 name = "emscripten_wasmsimd",
6907 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006908 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006909 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006910 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006911 },
6912)
6913
6914config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006915 name = "ios_armv7",
6916 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006917 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006918 "cpu": "ios_armv7",
6919 },
6920)
6921
6922config_setting(
6923 name = "ios_arm64",
6924 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006925 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006926 "cpu": "ios_arm64",
6927 },
6928)
6929
6930config_setting(
6931 name = "ios_arm64e",
6932 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006933 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006934 "cpu": "ios_arm64e",
6935 },
6936)
6937
6938config_setting(
6939 name = "ios_x86",
6940 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006941 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006942 "cpu": "ios_i386",
6943 },
6944)
6945
6946config_setting(
6947 name = "ios_x86_64",
6948 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006949 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006950 "cpu": "ios_x86_64",
6951 },
6952)
6953
6954config_setting(
6955 name = "watchos_armv7k",
6956 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006957 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006958 "cpu": "watchos_armv7k",
6959 },
6960)
6961
6962config_setting(
6963 name = "watchos_arm64_32",
6964 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006965 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006966 "cpu": "watchos_arm64_32",
6967 },
6968)
6969
6970config_setting(
6971 name = "watchos_x86",
6972 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006973 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006974 "cpu": "watchos_i386",
6975 },
6976)
6977
6978config_setting(
6979 name = "watchos_x86_64",
6980 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006981 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006982 "cpu": "watchos_x86_64",
6983 },
6984)
6985
6986config_setting(
6987 name = "tvos_arm64",
6988 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006989 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006990 "cpu": "tvos_arm64",
6991 },
6992)
6993
6994config_setting(
6995 name = "tvos_x86_64",
6996 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006997 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006998 "cpu": "tvos_x86_64",
6999 },
7000)