blob: a5630c8a9ef436f70fd3987968bac3d47779d1e4 [file] [log] [blame]
Frank Barchard1a953052020-11-16 18:44:58 -08001# Copyright 2020 Google LLC
Marat Dukhan08c4a432019-10-03 09:29:21 -07002#
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",
Marat Dukhan13b68f22020-11-12 11:55:19 -080062 "src/operators/depth-to-space-nchw2nhwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070063 "src/operators/fully-connected-nc.c",
64 "src/operators/global-average-pooling-ncw.c",
65 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070066 "src/operators/leaky-relu-nc.c",
67 "src/operators/max-pooling-nhwc.c",
68 "src/operators/prelu-nc.c",
Artsiom Ablavatski97918102020-10-27 15:52:59 -070069 "src/operators/resize-bilinear-nchw.c",
Marat Dukhane8265432020-04-28 18:42:59 -070070 "src/operators/resize-bilinear-nhwc.c",
71 "src/operators/sigmoid-nc.c",
72 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070073 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070074 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070075]
76
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070078 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070079 "src/subgraph/add2.c",
80 "src/subgraph/argmax-pooling-2d.c",
81 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070082 "src/subgraph/bankers-rounding.c",
83 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070084 "src/subgraph/clamp.c",
85 "src/subgraph/convolution-2d.c",
86 "src/subgraph/deconvolution-2d.c",
87 "src/subgraph/depthwise-convolution-2d.c",
Artsiom Ablavatskibbe85062020-11-05 14:07:37 -080088 "src/subgraph/depth-to-space.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070089 "src/subgraph/divide.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070090 "src/subgraph/floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -070091 "src/subgraph/fully-connected.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070092 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070093 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070094 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070096 "src/subgraph/maximum2.c",
97 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070098 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070099 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700100 "src/subgraph/prelu.c",
101 "src/subgraph/sigmoid.c",
102 "src/subgraph/softmax.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700103 "src/subgraph/square-root.c",
104 "src/subgraph/square.c",
105 "src/subgraph/squared-difference.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700106 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700107 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700108 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700109 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700110 "src/subgraph/unpooling-2d.c",
111]
112
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800113TABLE_SRCS = [
114 "src/tables/exp2-k-over-64.c",
115 "src/tables/exp2-k-over-2048.c",
Marat Dukhan1f256fc2020-09-24 21:27:14 -0700116 "src/tables/exp2minus-k-over-64.c",
117 "src/tables/exp2minus-k-over-2048.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800118]
119
Marat Dukhan08c4a432019-10-03 09:29:21 -0700120SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800121 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800122 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800123 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700124 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
125 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700126 "src/f32-clamp/gen/scalar-x1.c",
127 "src/f32-clamp/gen/scalar-x2.c",
128 "src/f32-clamp/gen/scalar-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700129 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700130 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700131 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700132 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700133 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
134 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
135 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700136 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700137 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
138 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
139 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700140 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700141 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
142 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
143 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700144 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700145 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
146 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
147 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700148 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700149 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
150 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
151 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700152 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700153 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
154 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
155 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700156 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c",
157 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c",
158 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700159 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700160 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700161 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c",
162 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c",
163 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c",
164 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c",
165 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700166 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c",
167 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c",
168 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700169 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700170 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700171 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c",
172 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c",
173 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700174 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c",
175 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c",
176 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c",
177 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700178 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700179 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c",
180 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700181 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700182 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700183 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c",
Marat Dukhan29c0c332020-10-28 22:11:00 -0700184 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c",
185 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c",
186 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c",
187 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c",
188 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c",
189 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c",
190 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c",
191 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c",
192 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c",
193 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700194 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700195 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
196 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700197 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
198 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
199 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700200 "src/f32-gemm/gen/1x4-minmax-scalar.c",
201 "src/f32-gemm/gen/1x4-relu-scalar.c",
202 "src/f32-gemm/gen/1x4-scalar.c",
203 "src/f32-gemm/gen/2x4-minmax-scalar.c",
204 "src/f32-gemm/gen/2x4-relu-scalar.c",
205 "src/f32-gemm/gen/2x4-scalar.c",
206 "src/f32-gemm/gen/4x2-minmax-scalar.c",
207 "src/f32-gemm/gen/4x2-relu-scalar.c",
208 "src/f32-gemm/gen/4x2-scalar.c",
209 "src/f32-gemm/gen/4x4-minmax-scalar.c",
210 "src/f32-gemm/gen/4x4-relu-scalar.c",
211 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700212 "src/f32-hswish/gen/hswish-scalar-x1.c",
213 "src/f32-hswish/gen/hswish-scalar-x2.c",
214 "src/f32-hswish/gen/hswish-scalar-x4.c",
XNNPACK Team21432672020-10-19 19:58:48 -0700215 "src/f32-ibilinear-chw/gen/scalar-p1.c",
216 "src/f32-ibilinear-chw/gen/scalar-p2.c",
217 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700218 "src/f32-ibilinear/gen/scalar-c1.c",
219 "src/f32-ibilinear/gen/scalar-c2.c",
220 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700221 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700222 "src/f32-igemm/gen/1x4-relu-scalar.c",
223 "src/f32-igemm/gen/1x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700224 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700225 "src/f32-igemm/gen/2x4-relu-scalar.c",
226 "src/f32-igemm/gen/2x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700227 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700228 "src/f32-igemm/gen/4x2-relu-scalar.c",
229 "src/f32-igemm/gen/4x2-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700230 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700231 "src/f32-igemm/gen/4x4-relu-scalar.c",
232 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700233 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
234 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
235 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700236 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
237 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
238 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
239 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800240 "src/f32-prelu/gen/scalar-2x1.c",
241 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800242 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800243 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700244 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800245 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
246 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700247 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800248 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800249 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700250 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800251 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
252 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700253 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700254 "src/f32-relu/gen/scalar-x1.c",
255 "src/f32-relu/gen/scalar-x2.c",
256 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700257 "src/f32-relu/gen/scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700258 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800259 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
260 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
261 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700262 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
263 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
264 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800265 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
266 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
267 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700268 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
269 "src/f32-spmm/gen/1x1-minmax-scalar.c",
270 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
271 "src/f32-spmm/gen/2x1-minmax-scalar.c",
272 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
273 "src/f32-spmm/gen/4x1-minmax-scalar.c",
274 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
275 "src/f32-spmm/gen/8x1-minmax-scalar.c",
276 "src/f32-spmm/gen/8x2-minmax-scalar.c",
277 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700278 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
279 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
280 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700281 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700282 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
283 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
284 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700285 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700286 "src/f32-vbinary/gen/vadd-scalar-x1.c",
287 "src/f32-vbinary/gen/vadd-scalar-x2.c",
288 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700289 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700290 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
291 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
292 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700293 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700294 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
295 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
296 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700297 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700298 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
299 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
300 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700301 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700302 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
303 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
304 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700305 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700306 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
307 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
308 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700309 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700310 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
311 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
312 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700313 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700314 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
315 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
316 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700317 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700318 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
319 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
320 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700321 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700322 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
323 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
324 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700325 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800326 "src/f32-vbinary/gen/vmax-scalar-x1.c",
327 "src/f32-vbinary/gen/vmax-scalar-x2.c",
328 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700329 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800330 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
331 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
332 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700333 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800334 "src/f32-vbinary/gen/vmin-scalar-x1.c",
335 "src/f32-vbinary/gen/vmin-scalar-x2.c",
336 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700337 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800338 "src/f32-vbinary/gen/vminc-scalar-x1.c",
339 "src/f32-vbinary/gen/vminc-scalar-x2.c",
340 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700341 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700342 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
343 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
344 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700345 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700346 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
347 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
348 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700349 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700350 "src/f32-vbinary/gen/vmul-scalar-x1.c",
351 "src/f32-vbinary/gen/vmul-scalar-x2.c",
352 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700353 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700354 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
355 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
356 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700357 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700358 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
359 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
360 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700361 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700362 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
363 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
364 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700365 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700366 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
367 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
368 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700369 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700370 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
371 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
372 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700373 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700374 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
375 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
376 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700377 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700378 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
379 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
380 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700381 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700382 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
383 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
384 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700385 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700386 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
387 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
388 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700389 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700390 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
391 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
392 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700393 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700394 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
395 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
396 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700397 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700398 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
399 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
400 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700401 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700402 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
403 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
404 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700405 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700406 "src/f32-vbinary/gen/vsub-scalar-x1.c",
407 "src/f32-vbinary/gen/vsub-scalar-x2.c",
408 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700409 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700410 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
411 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
412 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700413 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700414 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
415 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
416 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700417 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700418 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
419 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
420 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700421 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700422 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
423 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
424 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700425 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
426 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
427 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700428 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
429 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
430 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700431 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
432 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
433 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
434 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
435 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
436 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
437 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
438 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
439 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700440 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
441 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
442 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700443 "src/f32-vunary/gen/vabs-scalar-x1.c",
444 "src/f32-vunary/gen/vabs-scalar-x2.c",
445 "src/f32-vunary/gen/vabs-scalar-x4.c",
446 "src/f32-vunary/gen/vneg-scalar-x1.c",
447 "src/f32-vunary/gen/vneg-scalar-x2.c",
448 "src/f32-vunary/gen/vneg-scalar-x4.c",
449 "src/f32-vunary/gen/vsqr-scalar-x1.c",
450 "src/f32-vunary/gen/vsqr-scalar-x2.c",
451 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800452 "src/math/expminus-scalar-lut64-p2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700453 "src/math/expminus-scalar-lut2048-p1.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800454 "src/math/expminus-scalar-p5.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700455 "src/math/roundd-scalar-addsub.c",
456 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700457 "src/math/roundd-scalar-floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700458 "src/math/roundne-scalar-addsub.c",
459 "src/math/roundne-scalar-nearbyint.c",
460 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700461 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700462 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700463 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700464 "src/math/roundz-scalar-addsub.c",
465 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700466 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700467 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700468 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700469 "src/math/sigmoid-scalar-rr2-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700470 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
471 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700472 "src/qs8-requantization/fp32-scalar-lrintf.c",
473 "src/qs8-requantization/fp32-scalar-magic.c",
474 "src/qs8-requantization/precise-scalar-signed64.c",
475 "src/qs8-requantization/precise-scalar-unsigned32.c",
476 "src/qs8-requantization/precise-scalar-unsigned64.c",
477 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700478 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
479 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
480 "src/qu8-dwconv/up1x9-minmax-scalar.c",
481 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
482 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700483 "src/qu8-gemm/2x2-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700484 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
485 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700486 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700487 "src/qu8-requantization/fp32-scalar-lrintf.c",
488 "src/qu8-requantization/fp32-scalar-magic.c",
489 "src/qu8-requantization/precise-scalar-signed64.c",
490 "src/qu8-requantization/precise-scalar-unsigned32.c",
491 "src/qu8-requantization/precise-scalar-unsigned64.c",
492 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700493 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700494 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700495 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700496 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700497 "src/u8-rmax/scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700498 "src/x8-lut/scalar.c",
499 "src/x8-zip/x2-scalar.c",
500 "src/x8-zip/x3-scalar.c",
501 "src/x8-zip/x4-scalar.c",
502 "src/x8-zip/xm-scalar.c",
Yury Kartynnike7841862020-11-04 18:22:18 -0800503 "src/x32-depth-to-space-chw2hwc/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700504 "src/x32-fill/scalar-float.c",
505 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700506 "src/x32-packx/x2-scalar.c",
507 "src/x32-packx/x3-scalar.c",
508 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700509 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700510 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700511 "src/x32-unpool/scalar.c",
512 "src/x32-zip/x2-scalar.c",
513 "src/x32-zip/x3-scalar.c",
514 "src/x32-zip/x4-scalar.c",
515 "src/x32-zip/xm-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700516]
517
Marat Dukhan436ebe62019-12-04 15:10:12 -0800518WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700519 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
520 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700521 "src/f32-clamp/gen/wasm-x1.c",
522 "src/f32-clamp/gen/wasm-x2.c",
523 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700524 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
525 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700526 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
527 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700528 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
529 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700530 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
531 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700532 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
533 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700534 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
535 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700536 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
537 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700538 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
539 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700540 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
541 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700542 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
543 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700544 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
545 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700546 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
547 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700548 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
549 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700550 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
551 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
552 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
553 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700554 "src/f32-gemm/gen/1x4-relu-wasm.c",
555 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700556 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700557 "src/f32-gemm/gen/2x4-relu-wasm.c",
558 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700559 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700560 "src/f32-gemm/gen/4x2-relu-wasm.c",
561 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700562 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700563 "src/f32-gemm/gen/4x4-relu-wasm.c",
564 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700565 "src/f32-hswish/gen/hswish-wasm-x1.c",
566 "src/f32-hswish/gen/hswish-wasm-x2.c",
567 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700568 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700569 "src/f32-igemm/gen/1x4-relu-wasm.c",
570 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700571 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700572 "src/f32-igemm/gen/2x4-relu-wasm.c",
573 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700574 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700575 "src/f32-igemm/gen/4x2-relu-wasm.c",
576 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700577 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700578 "src/f32-igemm/gen/4x4-relu-wasm.c",
579 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700580 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
581 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
582 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700583 "src/f32-prelu/gen/wasm-2x1.c",
584 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700585 "src/f32-relu/gen/wasm-x1.c",
586 "src/f32-relu/gen/wasm-x2.c",
587 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700588 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700589 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
590 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
591 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700592 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700593 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
594 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
595 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700596 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700597 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
598 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
599 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
600 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700601 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
602 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
603 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700604 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700605 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
606 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
607 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
608 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700609 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
610 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
611 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700612 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700613 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
614 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
615 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
616 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700617 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
618 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
619 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700620 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800621 "src/f32-vbinary/gen/vmax-wasm-x1.c",
622 "src/f32-vbinary/gen/vmax-wasm-x2.c",
623 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700624 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800625 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
626 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
627 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700628 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800629 "src/f32-vbinary/gen/vmin-wasm-x1.c",
630 "src/f32-vbinary/gen/vmin-wasm-x2.c",
631 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700632 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800633 "src/f32-vbinary/gen/vminc-wasm-x1.c",
634 "src/f32-vbinary/gen/vminc-wasm-x2.c",
635 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700636 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700637 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
638 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
639 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700640 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700641 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
642 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
643 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700644 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700645 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
646 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
647 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
648 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700649 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
650 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
651 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700652 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700653 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
654 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
655 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
656 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700657 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
658 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
659 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700660 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700661 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
662 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
663 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
664 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700665 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
666 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
667 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700668 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700669 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
670 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
671 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
672 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700673 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
674 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
675 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700676 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700677 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
678 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
679 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
680 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700681 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
682 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
683 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700684 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700685 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
686 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
687 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700688 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
689 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
690 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800691]
692
Marat Dukhan290055c2020-06-09 12:24:29 -0700693WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700694 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
695 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
696 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700697 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
698 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
699 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
700 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700701 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
702 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
703 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
704 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700705 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700706 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700707 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700708 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700709 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
710 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700711 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700712 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700713 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700714 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
715 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
716 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
717 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
718 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
719 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
720 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700721 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700722 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700723 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700724 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
725 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700726 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700727 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700728 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700729 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
730 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
731 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
732 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
733 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
734 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
Frank Barchard1a953052020-11-16 18:44:58 -0800735 "src/f32-dwconv2d-chw/3x3p1-wasmsimd-1x4-acc3.c",
736 "src/f32-dwconv2d-chw/3x3s2p1-wasmsimd-1x4-acc3.c",
737 "src/f32-dwconv2d-chw/5x5p2-wasmsimd-3x4.c",
738 "src/f32-dwconv2d-chw/5x5s2p2-wasmsimd-1x4-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700739 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
740 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700741 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
742 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
743 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
744 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700745 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
746 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
747 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
748 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
749 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
750 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
751 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
752 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
753 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
754 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
755 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
756 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
757 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
758 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
759 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
760 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
761 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
762 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
763 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
764 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
765 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
766 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
767 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
768 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
769 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
770 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
771 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
772 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
773 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
774 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
775 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
776 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
777 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
778 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
779 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
780 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
781 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
782 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
783 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
784 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
785 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
786 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
787 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
788 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
789 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
790 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
791 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
792 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
793 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
794 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
795 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
796 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
797 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
798 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
799 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
800 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
801 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
802 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
803 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
804 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
805 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
806 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
807 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
808 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
809 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
810 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
811 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
812 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
813 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
814 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700815 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
816 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
817 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
XNNPACK Team965272b2020-10-23 21:10:15 -0700818 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
819 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700820 "src/f32-ibilinear/gen/wasmsimd-c4.c",
821 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700822 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
823 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
824 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
825 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
826 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
827 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
828 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
829 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
830 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
831 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
832 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
833 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
834 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
835 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
836 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
837 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
838 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
839 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
840 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
841 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
842 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
843 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
844 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
845 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
846 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
847 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
848 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
849 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
850 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
851 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
852 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
853 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
854 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
855 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
856 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
857 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
858 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
859 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
860 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
861 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700862 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
863 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700864 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
865 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
866 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
867 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700868 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
869 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700870 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
871 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
872 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700873 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
874 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700875 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
876 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
877 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
878 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
879 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
880 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
881 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700882 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
883 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700884 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
885 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
886 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
887 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700888 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700889 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700890 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700891 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
892 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700893 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700894 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
895 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700896 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700897 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
898 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700899 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700900 "src/f32-relu/gen/wasmsimd-x4.c",
901 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700902 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700903 "src/f32-rmax/wasmsimd-arm.c",
904 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700905 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
906 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
907 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
908 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
909 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
910 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700911 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
912 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
913 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
914 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
915 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
916 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800917 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined-x2.c",
918 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700919 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c",
920 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700921 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800922 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined-x2.c",
923 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700924 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c",
925 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700926 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800927 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined-x2.c",
928 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700929 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c",
930 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700931 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800932 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined-x2.c",
933 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700934 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c",
935 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700936 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800937 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined-x2.c",
938 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700939 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c",
940 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700941 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800942 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined-x2.c",
943 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700944 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c",
945 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700946 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800947 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined-x2.c",
948 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700949 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c",
950 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -0700951 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800952 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined-x2.c",
953 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700954 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c",
955 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -0700956 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700957 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
958 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700959 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700960 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
961 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700962 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700963 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
964 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700965 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700966 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
967 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700968 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700969 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
970 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700971 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700972 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
973 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700974 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700975 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
976 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700977 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700978 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
979 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700980 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700981 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
982 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700983 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700984 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
985 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700986 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700987 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
988 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700989 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700990 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
991 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700992 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700993 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
994 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700995 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700996 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
997 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700998 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700999 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
1000 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001001 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001002 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
1003 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001004 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001005 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
1006 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001007 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001008 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
1009 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001010 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001011 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
1012 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001013 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001014 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
1015 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001016 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001017 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
1018 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001019 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001020 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
1021 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001022 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001023 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
1024 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001025 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001026 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
1027 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001028 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001029 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
1030 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001031 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001032 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
1033 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001034 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001035 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
1036 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001037 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001038 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
1039 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001040 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001041 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
1042 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001043 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001044 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
1045 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001046 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001047 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
1048 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001049 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001050 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
1051 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001052 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001053 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
1054 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001055 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001056 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
1057 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001058 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001059 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
1060 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001061 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001062 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
1063 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001064 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001065 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1066 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001067 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001068 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1069 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001070 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001071 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1072 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001073 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001074 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1075 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001076 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001077 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1078 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001079 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001080 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1081 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001082 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001083 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1084 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001085 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001086 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1087 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001088 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001089 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1090 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001091 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001092 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1093 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001094 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001095 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1096 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001097 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001098 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1099 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001100 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001101 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1102 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001103 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001104 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1105 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001106 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001107 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1108 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1109 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1110 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001111 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001112 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001113 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001114 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001115 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1116 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1117 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1118 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001119 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1120 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1121 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1122 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1123 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1124 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1125 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1126 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1127 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1128 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001129 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1130 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001131 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1132 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1133 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1134 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1135 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1136 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001137 "src/math/roundd-wasmsimd-addsub.c",
1138 "src/math/roundd-wasmsimd-cvt.c",
1139 "src/math/roundne-wasmsimd-addsub.c",
1140 "src/math/roundu-wasmsimd-addsub.c",
1141 "src/math/roundu-wasmsimd-cvt.c",
1142 "src/math/roundz-wasmsimd-addsub.c",
1143 "src/math/roundz-wasmsimd-cvt.c",
1144 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1145 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001146 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1147 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1148 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001149 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1150 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1151 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001152 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1153 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1154 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001155 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001156 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001157 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001158 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1159 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001160 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001161 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1162 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001163 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1164 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001165 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001166 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001167 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001168 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001169 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001170 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001171 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001172 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1173 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1174 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1175 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1176 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1177 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1178 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1179 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001180 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001181 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001182 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001183 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001184 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001185 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001186 "src/x32-zip/x2-wasmsimd.c",
1187 "src/x32-zip/x3-wasmsimd.c",
1188 "src/x32-zip/x4-wasmsimd.c",
1189 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001190]
1191
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001192PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -07001193 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001194]
1195
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001196PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001197 "src/qs8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001198 "src/qs8-requantization/precise-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001199 "src/qu8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001200 "src/qu8-requantization/precise-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001201]
1202
Marat Dukhan08c4a432019-10-03 09:29:21 -07001203# ISA-specific micro-kernels
1204NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001205 "src/f32-argmaxpool/4x-neon-c4.c",
1206 "src/f32-argmaxpool/9p8x-neon-c4.c",
1207 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001208 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1209 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001210 "src/f32-clamp/gen/neon-x4.c",
1211 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001212 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001213 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001214 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001215 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001216 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001217 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001218 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001219 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001220 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001221 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001222 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001223 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001224 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001225 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1226 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1227 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1228 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1229 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001230 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001231 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001232 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1233 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1234 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001235 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001236 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001237 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1238 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1239 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1240 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1241 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001242 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1243 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1244 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001245 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001246 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001247 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1248 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1249 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001250 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1251 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1252 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1253 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001254 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001255 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1256 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001257 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001258 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001259 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001260 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001261 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1262 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001263 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1264 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1265 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1266 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1267 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1268 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1269 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1270 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001271 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001272 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001273 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001274 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1275 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001276 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001277 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1278 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001279 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001280 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1281 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1282 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1283 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1284 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001285 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1286 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001287 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1288 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001289 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1290 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001291 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1292 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1293 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1294 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1295 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1296 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1297 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1298 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1299 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1300 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1301 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1302 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1303 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1304 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1305 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1306 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001307 "src/f32-hswish/gen/hswish-neon-x4.c",
1308 "src/f32-hswish/gen/hswish-neon-x8.c",
1309 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001310 "src/f32-ibilinear/gen/neon-c4.c",
1311 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001312 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001313 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001314 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001315 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1316 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001317 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001318 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1319 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1320 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1321 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001322 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1323 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001324 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1325 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001326 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1327 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001328 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1329 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1330 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001331 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1332 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001333 "src/f32-prelu/gen/neon-1x4.c",
1334 "src/f32-prelu/gen/neon-1x8.c",
1335 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001336 "src/f32-prelu/gen/neon-2x4.c",
1337 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001338 "src/f32-prelu/gen/neon-2x16.c",
1339 "src/f32-prelu/gen/neon-4x4.c",
1340 "src/f32-prelu/gen/neon-4x8.c",
1341 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001342 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001343 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001344 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001345 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1346 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001347 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001348 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1349 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001350 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001351 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1352 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001353 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1354 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1355 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1356 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1357 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1358 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1359 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1360 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1361 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1362 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1363 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1364 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1365 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001366 "src/f32-relu/gen/neon-x4.c",
1367 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001368 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001369 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001370 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1371 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1372 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1373 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1374 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1375 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1376 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1377 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1378 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1379 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1380 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1381 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001382 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1383 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1384 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1385 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1386 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1387 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001388 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1389 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1390 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1391 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001392 "src/f32-vbinary/gen/vmax-neon-x4.c",
1393 "src/f32-vbinary/gen/vmax-neon-x8.c",
1394 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1395 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1396 "src/f32-vbinary/gen/vmin-neon-x4.c",
1397 "src/f32-vbinary/gen/vmin-neon-x8.c",
1398 "src/f32-vbinary/gen/vminc-neon-x4.c",
1399 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001400 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1401 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1402 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1403 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1404 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1405 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001406 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1407 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1408 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1409 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001410 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1411 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1412 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1413 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001414 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1415 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001416 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1417 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001418 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1419 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001420 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1421 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1422 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1423 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1424 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1425 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001426 "src/f32-vunary/gen/vabs-neon-x4.c",
1427 "src/f32-vunary/gen/vabs-neon-x8.c",
1428 "src/f32-vunary/gen/vneg-neon-x4.c",
1429 "src/f32-vunary/gen/vneg-neon-x8.c",
1430 "src/f32-vunary/gen/vsqr-neon-x4.c",
1431 "src/f32-vunary/gen/vsqr-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001432 "src/math/roundd-neon-addsub.c",
1433 "src/math/roundd-neon-cvt.c",
1434 "src/math/roundne-neon-addsub.c",
1435 "src/math/roundu-neon-addsub.c",
1436 "src/math/roundu-neon-cvt.c",
1437 "src/math/roundz-neon-addsub.c",
1438 "src/math/roundz-neon-cvt.c",
1439 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
1440 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1441 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1442 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1443 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1444 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1445 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1446 "src/math/sqrt-neon-nr1rsqrts.c",
1447 "src/math/sqrt-neon-nr2rsqrts.c",
1448 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001449 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1450 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1451 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1452 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001453 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1454 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1455 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1456 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001457 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1458 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1459 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1460 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001461 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001462 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001463 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001464 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1465 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001466 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001467 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001468 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001469 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001470 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001471 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001472 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1473 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1474 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1475 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1476 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1477 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1478 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1479 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001480 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1481 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1482 "src/qu8-dwconv/up8x9-minmax-neon.c",
1483 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1484 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1485 "src/qu8-gemm/4x8-minmax-neon.c",
1486 "src/qu8-gemm/8x8-minmax-neon.c",
1487 "src/qu8-igemm/4x8-minmax-neon.c",
1488 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001489 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001490 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001491 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001492 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001493 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001494 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001495 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001496 "src/x8-zip/x2-neon.c",
1497 "src/x8-zip/x3-neon.c",
1498 "src/x8-zip/x4-neon.c",
1499 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001500 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001501 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001502 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001503 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001504 "src/x32-zip/x2-neon.c",
1505 "src/x32-zip/x3-neon.c",
1506 "src/x32-zip/x4-neon.c",
1507 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001508]
1509
1510NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001511 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1512 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1513 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1514 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1515 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1516 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1517 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1518 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1519 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1520 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1521 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1522 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1523 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1524 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1525 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1526 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1527 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1528 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1529 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1530 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1531 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1532 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1533 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1534 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1535 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1536 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1537 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1538 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1539 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1540 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001541 "src/f32-ibilinear/gen/neonfma-c4.c",
1542 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001543 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001544 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001545 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001546 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1547 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001548 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1549 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001550 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1551 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001552 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1553 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001554 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001555 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001556 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001557 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1558 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001559 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001560 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1561 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001562 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001563 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1564 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001565 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1566 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1567 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1568 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1569 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1570 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1571 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1572 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1573 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1574 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1575 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1576 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1577 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001578 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1579 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1580 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1581 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1582 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1583 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001584 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1585 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1586 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1587 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1588 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1589 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001590 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1591 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1592 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1593 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1594 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1595 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001596 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1597 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1598 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1599 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1600 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1601 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001602 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1603 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1604 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1605 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1606 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1607 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001608 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1609 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1610 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1611 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1612 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1613 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001614 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1615 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1616 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1617 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1618 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1619 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1620 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1621 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1622 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1623 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1624 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1625 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1626 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1627 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1628 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1629 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1630 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1631 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001632 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1633 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001634 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1635 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1636 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1637 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1638 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1639 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1640 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1641 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1642 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1643 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1644 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1645 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1646 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1647 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1648 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1649 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1650 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1651 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1652 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1653 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001654 "src/math/exp-neonfma-lut64-p2.c",
1655 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001656 "src/math/expminus-neonfma-lut64-p2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001657 "src/math/expminus-neonfma-lut2048-p1.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001658 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001659 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1660 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1661 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001662 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1663 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1664 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001665 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1666 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1667 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001668 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1669 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1670 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001671 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1672 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1673 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001674 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1675 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1676 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001677 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001678 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001679 "src/math/sqrt-neonfma-nr2fma.c",
1680 "src/math/sqrt-neonfma-nr2fma1adj.c",
1681 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001682]
1683
1684AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001685 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001686 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001687 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001688 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001689 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001690 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001691 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001692 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001693 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001694 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
1695 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
1696 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001697 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001698 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001699 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
1700 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
1701 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
1702 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
1703 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001704 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
1705 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
1706 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001707 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001708 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001709 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
1710 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
1711 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001712 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
1713 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
1714 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
1715 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001716 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001717 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
1718 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001719 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001720 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001721 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001722 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001723 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
1724 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001725 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
1726 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
1727 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
1728 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
1729 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
1730 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
1731 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
1732 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001733 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001734 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001735 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1736 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1737 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1738 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1739 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1740 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1741 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1742 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1743 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1744 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1745 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1746 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1747 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1748 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1749 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1750 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1751 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1752 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1753 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1754 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001755 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1756 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1757 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1758 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1759 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1760 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1761 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1762 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1763 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1764 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1765 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1766 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001767 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1768 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1769 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1770 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1771 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1772 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001773 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001774 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001775 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1776 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1777 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1778 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001779 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001780 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1781 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1782 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001783 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1784 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1785 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1786 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001787 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001788 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1789 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1790 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001791 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001792 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001793 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
1794 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
1795 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001796 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1797 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1798 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1799 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1800 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1801 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001802 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1803 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001804 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001805 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001806 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001807 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001808 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001809 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001810]
1811
Marat Dukhan8853b822020-05-07 12:19:01 -07001812NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001813 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1814 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001815 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1816 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1817 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1818 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1819 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1820 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001821 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001822 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001823 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001824 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001825]
1826
Marat Dukhan08c4a432019-10-03 09:29:21 -07001827AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001828 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001829 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001830 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1831 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1832 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1833 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001834 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1835 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1836 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1837 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1838 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1839 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1840 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1841 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001842 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1843 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001844 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1845 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1846 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1847 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1848 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1849 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1850 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
1851 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1852 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1853 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1854 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1855 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1856 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1857 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1858 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1859 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001860 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001861 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1862 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1863 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1864 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1865 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1866 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1867 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1868 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1869 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001870 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001871 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001872 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001873 "src/f16-relu/gen/neonfp16arith-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001874 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001875 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001876 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001877 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001878 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001879 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001880 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001881 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1882 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1883 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1884 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1885 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1886 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1887 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1888 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1889 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1890 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1891 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1892 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1893 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1894 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1895 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1896 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1897 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1898 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1899 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1900 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1901 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1902 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1903 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1904 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1905 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1906 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
1907 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1908 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1909 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1910 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
1911 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001912]
1913
Benoit Jacoba9644732020-08-13 12:48:55 -07001914NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07001915 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001916 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001917 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001918 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001919 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001920 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001921 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001922 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001923 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001924 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001925 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001926 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001927 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001928 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001929 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001930 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001931 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001932 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001933]
1934
Marat Dukhan08c4a432019-10-03 09:29:21 -07001935SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001936 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1937 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001938 "src/f32-clamp/gen/sse-x4.c",
1939 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001940 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1941 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001942 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1943 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1944 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1945 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001946 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1947 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001948 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1949 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1950 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1951 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001952 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1953 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001954 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
1955 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
1956 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07001957 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001958 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07001959 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
1960 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
1961 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
1962 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
1963 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07001964 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
1965 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
1966 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001967 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07001968 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001969 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
1970 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
1971 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07001972 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
1973 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
1974 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
1975 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
1976 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
1977 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
1978 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
1979 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
1980 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
1981 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
1982 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
1983 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
1984 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07001985 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
1986 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
1987 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
1988 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
1989 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
1990 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
1991 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
1992 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07001993 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001994 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001995 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001996 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1997 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001998 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1999 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
2000 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
2001 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
2002 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
2003 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002004 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
2005 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
2006 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
2007 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
2008 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
2009 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
2010 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002011 "src/f32-hswish/gen/hswish-sse-x4.c",
2012 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07002013 "src/f32-ibilinear/gen/sse-c4.c",
2014 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002015 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
2016 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
2017 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
2018 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2019 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2020 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2021 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002022 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2023 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2024 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002025 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002026 "src/f32-prelu/gen/sse-2x4.c",
2027 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002028 "src/f32-relu/gen/sse-x4.c",
2029 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002030 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002031 "src/f32-spmm/gen/4x1-minmax-sse.c",
2032 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002033 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002034 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002035 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2036 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2037 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2038 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2039 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2040 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2041 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2042 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002043 "src/f32-vbinary/gen/vmax-sse-x4.c",
2044 "src/f32-vbinary/gen/vmax-sse-x8.c",
2045 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2046 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2047 "src/f32-vbinary/gen/vmin-sse-x4.c",
2048 "src/f32-vbinary/gen/vmin-sse-x8.c",
2049 "src/f32-vbinary/gen/vminc-sse-x4.c",
2050 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002051 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2052 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2053 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2054 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2055 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2056 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2057 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2058 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002059 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2060 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2061 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2062 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002063 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2064 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2065 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2066 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002067 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2068 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002069 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2070 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002071 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2072 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002073 "src/f32-vunary/gen/vabs-sse-x4.c",
2074 "src/f32-vunary/gen/vabs-sse-x8.c",
2075 "src/f32-vunary/gen/vneg-sse-x4.c",
2076 "src/f32-vunary/gen/vneg-sse-x8.c",
2077 "src/f32-vunary/gen/vsqr-sse-x4.c",
2078 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002079 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002080 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002081 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002082 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002083 "src/math/sqrt-sse-hh1mac.c",
2084 "src/math/sqrt-sse-nr1mac.c",
2085 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002086 "src/x32-fill/sse.c",
2087 "src/x32-packx/x4-sse.c",
2088 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002089]
2090
2091SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002092 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002093 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002094 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002095 "src/f32-prelu/gen/sse2-2x4.c",
2096 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002097 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002098 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002099 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002100 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2101 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002102 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002103 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2104 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002105 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002106 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2107 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002108 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002109 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
2110 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
2111 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
2112 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
2113 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
2114 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002115 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002116 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002117 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002118 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002119 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
2120 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002121 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2122 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002123 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2124 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002125 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2126 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2127 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2128 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2129 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2130 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
2131 "src/math/exp-sse2-lut64-p2.c",
2132 "src/math/exp-sse2-p5.c",
2133 "src/math/expminus-sse2-p5.c",
2134 "src/math/roundd-sse2-cvt.c",
2135 "src/math/roundne-sse2-cvt.c",
2136 "src/math/roundu-sse2-cvt.c",
2137 "src/math/roundz-sse2-cvt.c",
2138 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2139 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2140 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2141 "src/math/sigmoid-sse2-rr2-p5-div.c",
2142 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2143 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002144 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2145 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2146 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002147 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2148 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2149 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002150 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2151 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2152 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002153 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002154 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002155 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002156 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002157 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002158 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002159 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2160 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002161 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002162 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2163 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002164 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002165 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2166 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2167 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002168 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002169 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002170 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002171 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002172 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002173 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002174 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002175 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002176 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2177 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002178 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002179 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002180 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002181 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2182 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2183 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2184 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002185 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2186 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2187 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2188 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002189 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2190 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002191 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2192 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2193 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2194 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2195 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002196 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002197 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002198 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002199 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002200 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002201 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002202 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002203 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002204 "src/x8-zip/x2-sse2.c",
2205 "src/x8-zip/x3-sse2.c",
2206 "src/x8-zip/x4-sse2.c",
2207 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002208 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002209 "src/x32-zip/x2-sse2.c",
2210 "src/x32-zip/x3-sse2.c",
2211 "src/x32-zip/x4-sse2.c",
2212 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002213]
2214
2215SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002216 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2217 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2218 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002219 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002220 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002221 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2222 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2223 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2224 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2225 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002226 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2227 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2228 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002229 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2230 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2231 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002232 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2233 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2234 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002235 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002236 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002237 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002238 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002239 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002240 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002241 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2242 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002243 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002244 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2245 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002246 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002247 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2248 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2249 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002250 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002251 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002252 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002253 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002254 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002255 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002256 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002257 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002258 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2259 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002260 "src/qs8-requantization/precise-ssse3.c",
2261 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002262 "src/qu8-requantization/precise-ssse3.c",
2263 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002264]
2265
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002266SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002267 "src/f32-prelu/gen/sse41-2x4.c",
2268 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002269 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2270 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2271 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2272 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2273 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2274 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002275 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002276 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002277 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002278 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002279 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2280 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002281 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2282 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002283 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2284 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002285 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2286 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2287 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2288 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2289 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2290 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2291 "src/math/roundd-sse41.c",
2292 "src/math/roundne-sse41.c",
2293 "src/math/roundu-sse41.c",
2294 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002295 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2296 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2297 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002298 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2299 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2300 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002301 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2302 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2303 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002304 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002305 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002306 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002307 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002308 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002309 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002310 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2311 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002312 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002313 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2314 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002315 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002316 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2317 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2318 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002319 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002320 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002321 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002322 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002323 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002324 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002325 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002326 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002327 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2328 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002329 "src/qs8-requantization/fp32-sse4.c",
2330 "src/qs8-requantization/precise-sse4.c",
2331 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002332 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2333 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2334 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2335 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002336 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2337 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2338 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2339 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002340 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2341 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2342 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2343 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002344 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2345 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2346 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2347 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002348 "src/qu8-requantization/precise-sse4.c",
2349 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002350]
2351
Marat Dukhan08c4a432019-10-03 09:29:21 -07002352AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002353 "src/f32-clamp/gen/avx-x8.c",
2354 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002355 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2356 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002357 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2358 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002359 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2360 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002361 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2362 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2363 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2364 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2365 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2366 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002367 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002368 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2369 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002370 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002371 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002372 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002373 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002374 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2375 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2376 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2377 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2378 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2379 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2380 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2381 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2382 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2383 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2384 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002385 "src/f32-hswish/gen/hswish-avx-x8.c",
2386 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002387 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002388 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2389 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002390 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002391 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002392 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002393 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002394 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2395 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002396 "src/f32-prelu/gen/avx-2x8.c",
2397 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002398 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002399 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002400 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002401 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2402 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2403 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2404 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2405 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2406 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2407 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2408 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2409 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2410 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2411 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2412 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2413 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2414 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2415 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2416 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2417 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2418 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2419 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2420 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002421 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2422 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2423 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2424 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2425 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2426 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2427 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2428 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002429 "src/f32-vbinary/gen/vmax-avx-x8.c",
2430 "src/f32-vbinary/gen/vmax-avx-x16.c",
2431 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2432 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2433 "src/f32-vbinary/gen/vmin-avx-x8.c",
2434 "src/f32-vbinary/gen/vmin-avx-x16.c",
2435 "src/f32-vbinary/gen/vminc-avx-x8.c",
2436 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002437 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2438 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2439 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2440 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2441 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2442 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2443 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2444 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002445 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2446 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2447 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2448 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002449 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2450 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2451 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2452 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002453 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2454 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002455 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2456 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002457 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2458 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2459 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2460 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2461 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2462 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002463 "src/f32-vscale/avx-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002464 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2465 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002466 "src/f32-vunary/gen/vabs-avx-x8.c",
2467 "src/f32-vunary/gen/vabs-avx-x16.c",
2468 "src/f32-vunary/gen/vneg-avx-x8.c",
2469 "src/f32-vunary/gen/vneg-avx-x16.c",
2470 "src/f32-vunary/gen/vsqr-avx-x8.c",
2471 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhand08abcb2020-09-21 00:26:25 -07002472 "src/math/exp-avx-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002473 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002474 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002475 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2476 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002477]
2478
Marat Dukhan1566fee2020-08-02 21:55:41 -07002479XOP_UKERNELS = [
2480 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002481 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002482 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002483 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002484 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002485 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002486 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2487 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002488 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002489 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2490 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002491 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002492 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2493 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2494 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002495 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002496 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002497 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002498 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002499 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002500 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002501 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002502 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002503 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2504 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002505 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2506 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2507 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2508 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2509 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2510 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2511 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2512 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002513]
2514
Marat Dukhanfda12b82019-11-21 12:27:59 -08002515FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002516 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2517 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002518 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2519 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002520 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2521 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002522 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2523 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2524 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2525 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2526 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2527 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002528 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002529 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2530 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2531 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2532 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002533 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002534 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2535 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002536 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002537 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2538 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002539 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2540 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2541 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002542 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2543 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2544 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2545 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2546 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2547 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2548 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2549 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2550 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2551 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2552 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2553 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2554 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2555 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002556 "src/f32-hswish/gen/hswish-fma3-x8.c",
2557 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002558 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002559 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2560 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2561 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2562 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002563 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002564 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2565 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002566 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002567 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2568 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002569 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2570 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2571 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002572 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2573 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2574 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2575 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2576 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2577 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2578 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2579 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002580 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002581 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002582 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002583]
2584
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002585AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002586 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2587 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002588 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002589 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002590 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002591 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2592 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002593 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002594 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2595 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2596 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002597 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002598 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2599 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002600 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002601 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002602 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002603 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2604 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002605 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002606 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2607 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2608 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002609 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002610 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2611 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002612 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002613 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002614 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002615 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2616 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002617 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002618 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2619 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2620 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002621 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002622 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2623 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2624 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2625 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2626 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2627 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2628 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2629 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2630 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2631 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2632 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2633 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2634 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2635 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2636 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2637 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2638 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2639 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2640 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2641 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2642 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2643 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2644 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2645 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2646 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2647 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2648 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2649 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2650 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2651 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002652 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2653 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2654 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2655 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2656 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2657 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2658 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2659 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2660 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2661 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2662 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2663 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2664 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2665 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2666 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2667 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2668 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2669 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2670 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2671 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2672 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2673 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2674 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2675 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002676 "src/math/exp-avx2-lut8-p3-perm.c",
2677 "src/math/exp-avx2-lut8-p4-perm.c",
2678 "src/math/exp-avx2-p5.c",
2679 "src/math/expminus-avx2-p5.c",
2680 "src/math/extexp-avx2-p5.c",
2681 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2682 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2683 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2684 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2685 "src/math/sigmoid-avx2-rr1-p5-div.c",
2686 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
2687 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2688 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2689 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2690 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2691 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2692 "src/math/sigmoid-avx2-rr2-p5-div.c",
2693 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2694 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002695 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002696 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002697 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2698 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002699 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002700 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002701 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002702 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002703 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002704 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002705 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002706 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002707 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2708 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2709 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002710 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2711 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2712 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2713 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2714 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2715 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2716 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2717 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002718]
2719
Marat Dukhan08c4a432019-10-03 09:29:21 -07002720AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002721 "src/f32-clamp/gen/avx512f-x16.c",
2722 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002723 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2724 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002725 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2726 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002727 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2728 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002729 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2730 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2731 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2732 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2733 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2734 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002735 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2736 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2737 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2738 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2739 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2740 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002741 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2742 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2743 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2744 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2745 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2746 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002747 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2748 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002749 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2750 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2751 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2752 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2753 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2754 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002755 "src/f32-prelu/gen/avx512f-2x16.c",
2756 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002757 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2758 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002759 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002760 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002761 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002762 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2763 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002764 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002765 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2766 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2767 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002768 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002769 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2770 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002771 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002772 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002773 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002774 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2775 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002776 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002777 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2778 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2779 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002780 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002781 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2782 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002783 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002784 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002785 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002786 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2787 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002788 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002789 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2790 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2791 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002792 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002793 "src/f32-relu/gen/avx512f-x16.c",
2794 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002795 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002796 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2797 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2798 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2799 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2800 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2801 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2802 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2803 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2804 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2805 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2806 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2807 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2808 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2809 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2810 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2811 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002812 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2813 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2814 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2815 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2816 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2817 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2818 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2819 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2820 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2821 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2822 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2823 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2824 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2825 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2826 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2827 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002828 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2829 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2830 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2831 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2832 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2833 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2834 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2835 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2836 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2837 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2838 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2839 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2840 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2841 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2842 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2843 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002844 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2845 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2846 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2847 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2848 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2849 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2850 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2851 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002852 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2853 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2854 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2855 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2856 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2857 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2858 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2859 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002860 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2861 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2862 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2863 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2864 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2865 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2866 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2867 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002868 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2869 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2870 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2871 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002872 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2873 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2874 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2875 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002876 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2877 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002878 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2879 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
2880 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2881 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2882 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2883 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2884 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2885 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002886 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002887 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2888 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2889 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2890 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2891 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2892 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2893 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2894 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2895 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2896 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2897 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2898 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2899 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2900 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2901 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2902 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2903 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2904 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2905 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2906 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2907 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2908 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2909 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2910 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002911 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2912 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2913 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2914 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2915 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2916 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2917 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2918 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002919 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2920 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2921 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2922 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2923 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2924 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002925 "src/math/exp-avx512f-lut16-p3-perm-scalef.c",
2926 "src/math/exp-avx512f-lut16-p3-perm.c",
2927 "src/math/exp-avx512f-lut32-p2-perm2-scalef.c",
2928 "src/math/exp-avx512f-lut32-p2-perm2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002929 "src/math/exp-avx512f-p5-scalef.c",
2930 "src/math/exp-avx512f-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002931 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002932 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002933 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002934 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002935 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002936 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002937 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002938 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002939 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002940 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
2941 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
2942 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
2943 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
2944 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2945 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2946 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2947 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2948 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
2949 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002950 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002951 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002952 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
2953 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
2954 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
2955 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002956 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002957 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002958 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002959]
2960
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002961AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002962 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2963 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002964 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2965 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2966 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2967 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2968 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2969 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2970 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2971 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2972]
2973
Frank Barchardbcedc082020-08-17 18:00:51 -07002974WASM32_ASM_UKERNELS = [
2975 "src/f32-relu/wasm_shr_x1.S",
2976 "src/f32-relu/wasm_shr_x2.S",
2977 "src/f32-relu/wasm_shr_x4.S",
2978]
2979
Marat Dukhan08c4a432019-10-03 09:29:21 -07002980AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002981 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002982 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002983 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2984 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002985 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002986 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002987 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002988 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002989 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2990 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002991 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2992 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2993 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2994 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002995]
2996
2997AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002998 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07002999 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003000 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003001 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003002 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003003 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07003004 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003005 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
3006 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
3007 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
3008 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
3009 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
3010 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
3011 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003012 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
3013 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003014 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3015 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3016 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003017 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
3018 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003019 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3020 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3021 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3022 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003023 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003024 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3025 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003026 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3027 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3028 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3029 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003030 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003031 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003032 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003033 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003034 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3035 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3036 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3037 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3038 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3039 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3040 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3041 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3042 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3043 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3044 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3045 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3046 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3047 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3048 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3049 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3050 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3051 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3052 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3053 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3054 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3055 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003056 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003057 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003058 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3059 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003060 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3061 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3062 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3063 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3064 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3065 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003066 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3067 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3068 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3069 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003070 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3071 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003072 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07003073 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07003074 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003075]
3076
Marat Dukhan1b354632020-03-23 12:50:22 -07003077INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003078 "src/xnnpack/argmaxpool.h",
3079 "src/xnnpack/avgpool.h",
3080 "src/xnnpack/clamp.h",
3081 "src/xnnpack/common.h",
3082 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003083 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003084 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003085 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003086 "src/xnnpack/gavgpool.h",
3087 "src/xnnpack/gemm.h",
3088 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003089 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003090 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003091 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003092 "src/xnnpack/lut.h",
3093 "src/xnnpack/math.h",
3094 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003095 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003096 "src/xnnpack/packx.h",
3097 "src/xnnpack/pad.h",
3098 "src/xnnpack/params.h",
3099 "src/xnnpack/pavgpool.h",
3100 "src/xnnpack/ppmm.h",
3101 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003102 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003103 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003104 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003105 "src/xnnpack/rmax.h",
3106 "src/xnnpack/scalar-utils.h",
3107 "src/xnnpack/spmm.h",
3108 "src/xnnpack/unpool.h",
3109 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003110 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003111 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003112 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003113 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003114 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003115 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003116 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003117]
3118
3119INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003120 "include/xnnpack.h",
3121 "src/xnnpack/allocator.h",
3122 "src/xnnpack/compute.h",
3123 "src/xnnpack/im2col.h",
3124 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003125 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003126 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003127 "src/xnnpack/operator.h",
3128 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003129 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003130 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003131 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003132 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003133]
3134
Marat Dukhan1b354632020-03-23 12:50:22 -07003135ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003136 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003137]
3138
Marat Dukhan1b354632020-03-23 12:50:22 -07003139MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003140 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003141 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003142]
3143
Marat Dukhan1b354632020-03-23 12:50:22 -07003144MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003145 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003146 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003147 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003148 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003149]
3150
3151OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003152 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003153 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003154]
3155
3156WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003157 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003158 "src/xnnpack/operator.h",
3159 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003160]
3161
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003162LOGGING_COPTS = select({
3163 # No logging in optimized mode
3164 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3165 # Full logging in debug mode
3166 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3167 # Error-only logging in default (fastbuild) mode
3168 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3169})
3170
Marat Dukhan3b59de22020-06-03 20:15:19 -07003171LOGGING_SRCS = select({
3172 # No logging in optimized mode
3173 ":optimized_build": [],
3174 "//conditions:default": [
3175 "src/operator-strings.c",
3176 "src/subgraph-strings.c",
3177 ],
3178})
3179
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003180LOGGING_HDRS = [
3181 "src/xnnpack/log.h",
3182]
3183
Marat Dukhan08c4a432019-10-03 09:29:21 -07003184xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003185 name = "tables",
3186 srcs = TABLE_SRCS,
3187 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003188 gcc_copts = xnnpack_gcc_std_copts(),
3189 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003190)
3191
3192xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003193 name = "scalar_ukernels",
3194 srcs = SCALAR_UKERNELS,
3195 hdrs = INTERNAL_HDRS,
3196 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003197 gcc_copts = xnnpack_gcc_std_copts(),
3198 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003199 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003200 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003201 "@FP16",
3202 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003203 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003204 ],
3205)
3206
3207xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003208 name = "scalar_ukernels_test_mode",
3209 srcs = SCALAR_UKERNELS,
3210 hdrs = INTERNAL_HDRS,
3211 aarch32_copts = ["-marm"],
3212 copts = [
3213 "-UNDEBUG",
3214 "-DXNN_TEST_MODE=1",
3215 ],
3216 gcc_copts = xnnpack_gcc_std_copts(),
3217 msvc_copts = xnnpack_msvc_std_copts(),
3218 deps = [
3219 ":tables",
3220 "@FP16",
3221 "@FXdiv",
3222 "@pthreadpool",
3223 ],
3224)
3225
3226xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003227 name = "wasm_ukernels",
3228 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003229 gcc_copts = xnnpack_gcc_std_copts(),
3230 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003231 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003232 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003233 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003234 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003235 "@FP16",
3236 "@FXdiv",
3237 "@pthreadpool",
3238 ],
3239)
3240
3241xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003242 name = "wasm_ukernels_test_mode",
3243 hdrs = INTERNAL_HDRS,
3244 copts = [
3245 "-UNDEBUG",
3246 "-DXNN_TEST_MODE=1",
3247 ],
3248 gcc_copts = xnnpack_gcc_std_copts(),
3249 msvc_copts = xnnpack_msvc_std_copts(),
3250 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003251 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003252 deps = [
3253 ":tables",
3254 "@FP16",
3255 "@FXdiv",
3256 "@pthreadpool",
3257 ],
3258)
3259
3260xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003261 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003262 hdrs = INTERNAL_HDRS,
3263 aarch32_copts = [
3264 "-marm",
3265 "-mfpu=neon",
3266 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003267 gcc_copts = xnnpack_gcc_std_copts(),
3268 gcc_x86_copts = ["-msse2"],
3269 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003270 optimized_copts = [
3271 "-O3",
3272 "-ffast-math",
3273 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003274 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003275 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003276 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003277 "@FP16",
3278 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003279 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003280 ],
3281)
3282
3283xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003284 name = "psimd_fastmath_ukernels_test_mode",
3285 hdrs = INTERNAL_HDRS,
3286 aarch32_copts = [
3287 "-marm",
3288 "-mfpu=neon",
3289 ],
3290 copts = [
3291 "-UNDEBUG",
3292 "-DXNN_TEST_MODE=1",
3293 ],
3294 gcc_copts = xnnpack_gcc_std_copts(),
3295 gcc_x86_copts = ["-msse2"],
3296 msvc_copts = xnnpack_msvc_std_copts(),
3297 optimized_copts = [
3298 "-O3",
3299 "-ffast-math",
3300 ],
3301 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
3302 deps = [
3303 ":tables",
3304 "@FP16",
3305 "@psimd",
3306 "@pthreadpool",
3307 ],
3308)
3309
3310xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003311 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003312 hdrs = INTERNAL_HDRS,
3313 aarch32_copts = [
3314 "-marm",
3315 "-mfpu=neon",
3316 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003317 gcc_copts = xnnpack_gcc_std_copts(),
3318 gcc_x86_copts = ["-msse2"],
3319 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003320 optimized_copts = [
3321 "-O3",
3322 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003323 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003324 deps = [
3325 ":tables",
3326 "@FP16",
3327 "@psimd",
3328 "@pthreadpool",
3329 ],
3330)
3331
3332xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003333 name = "psimd_accmath_ukernels_test_mode",
3334 hdrs = INTERNAL_HDRS,
3335 aarch32_copts = [
3336 "-marm",
3337 "-mfpu=neon",
3338 ],
3339 copts = [
3340 "-UNDEBUG",
3341 "-DXNN_TEST_MODE=1",
3342 ],
3343 gcc_copts = xnnpack_gcc_std_copts(),
3344 gcc_x86_copts = ["-msse2"],
3345 msvc_copts = xnnpack_msvc_std_copts(),
3346 optimized_copts = [
3347 "-O3",
3348 ],
3349 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3350 deps = [
3351 ":tables",
3352 "@FP16",
3353 "@psimd",
3354 "@pthreadpool",
3355 ],
3356)
3357
3358xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003359 name = "neon_ukernels",
3360 hdrs = INTERNAL_HDRS,
3361 aarch32_copts = [
3362 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003363 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003364 "-mfpu=neon",
3365 ],
3366 aarch32_srcs = NEON_UKERNELS,
3367 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003368 gcc_copts = xnnpack_gcc_std_copts(),
3369 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003370 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003371 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003372 "@FP16",
3373 "@pthreadpool",
3374 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003375)
3376
3377xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003378 name = "neon_ukernels_test_mode",
3379 hdrs = INTERNAL_HDRS,
3380 aarch32_copts = [
3381 "-marm",
3382 "-march=armv7-a",
3383 "-mfpu=neon",
3384 ],
3385 aarch32_srcs = NEON_UKERNELS,
3386 aarch64_srcs = NEON_UKERNELS,
3387 copts = [
3388 "-UNDEBUG",
3389 "-DXNN_TEST_MODE=1",
3390 ],
3391 gcc_copts = xnnpack_gcc_std_copts(),
3392 msvc_copts = xnnpack_msvc_std_copts(),
3393 deps = [
3394 ":tables",
3395 "@FP16",
3396 "@pthreadpool",
3397 ],
3398)
3399
3400xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003401 name = "neonfma_ukernels",
3402 hdrs = INTERNAL_HDRS,
3403 aarch32_copts = [
3404 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003405 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003406 "-mfpu=neon-vfpv4",
3407 ],
3408 aarch32_srcs = NEONFMA_UKERNELS,
3409 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003410 apple_aarch32_copts = [
3411 "-mcpu=swift",
3412 "-mtune=generic",
3413 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003414 gcc_copts = xnnpack_gcc_std_copts(),
3415 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003416 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003417 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003418 "@FP16",
3419 "@pthreadpool",
3420 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003421)
3422
3423xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003424 name = "neonfma_ukernels_test_mode",
3425 hdrs = INTERNAL_HDRS,
3426 aarch32_copts = [
3427 "-marm",
3428 "-march=armv7-a",
3429 "-mfpu=neon-vfpv4",
3430 ],
3431 aarch32_srcs = NEONFMA_UKERNELS,
3432 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003433 apple_aarch32_copts = [
3434 "-mcpu=swift",
3435 "-mtune=generic",
3436 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003437 copts = [
3438 "-UNDEBUG",
3439 "-DXNN_TEST_MODE=1",
3440 ],
3441 gcc_copts = xnnpack_gcc_std_copts(),
3442 msvc_copts = xnnpack_msvc_std_copts(),
3443 deps = [
3444 ":tables",
3445 "@FP16",
3446 "@pthreadpool",
3447 ],
3448)
3449
3450xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003451 name = "neonv8_ukernels",
3452 hdrs = INTERNAL_HDRS,
3453 aarch32_copts = [
3454 "-marm",
3455 "-march=armv8-a",
3456 "-mfpu=neon-fp-armv8",
3457 ],
3458 aarch32_srcs = NEONV8_UKERNELS,
3459 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003460 apple_aarch32_copts = [
3461 "-mcpu=cyclone",
3462 "-mtune=generic",
3463 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003464 gcc_copts = xnnpack_gcc_std_copts(),
3465 msvc_copts = xnnpack_msvc_std_copts(),
3466 deps = [
3467 ":tables",
3468 "@FP16",
3469 "@pthreadpool",
3470 ],
3471)
3472
3473xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003474 name = "neonv8_ukernels_test_mode",
3475 hdrs = INTERNAL_HDRS,
3476 aarch32_copts = [
3477 "-marm",
3478 "-march=armv8-a",
3479 "-mfpu=neon-fp-armv8",
3480 ],
3481 aarch32_srcs = NEONV8_UKERNELS,
3482 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003483 apple_aarch32_copts = [
3484 "-mcpu=cyclone",
3485 "-mtune=generic",
3486 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003487 copts = [
3488 "-UNDEBUG",
3489 "-DXNN_TEST_MODE=1",
3490 ],
3491 gcc_copts = xnnpack_gcc_std_copts(),
3492 msvc_copts = xnnpack_msvc_std_copts(),
3493 deps = [
3494 ":tables",
3495 "@FP16",
3496 "@pthreadpool",
3497 ],
3498)
3499
3500xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003501 name = "neonfp16arith_ukernels",
3502 hdrs = INTERNAL_HDRS,
3503 aarch64_copts = ["-march=armv8.2-a+fp16"],
3504 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003505 gcc_copts = xnnpack_gcc_std_copts(),
3506 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003507 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003508 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003509 "@FP16",
3510 "@pthreadpool",
3511 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003512)
3513
3514xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003515 name = "neonfp16arith_ukernels_test_mode",
3516 hdrs = INTERNAL_HDRS,
3517 aarch64_copts = ["-march=armv8.2-a+fp16"],
3518 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3519 copts = [
3520 "-UNDEBUG",
3521 "-DXNN_TEST_MODE=1",
3522 ],
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(
Benoit Jacoba9644732020-08-13 12:48:55 -07003533 name = "neondot_ukernels",
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 gcc_copts = xnnpack_gcc_std_copts(),
3544 msvc_copts = xnnpack_msvc_std_copts(),
3545 deps = [
3546 ":tables",
3547 "@FP16",
3548 "@pthreadpool",
3549 ],
3550)
3551
3552xnnpack_cc_library(
3553 name = "neondot_ukernels_test_mode",
3554 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003555 aarch32_copts = [
3556 "-marm",
3557 "-march=armv8.2-a+dotprod",
3558 "-mfpu=neon-fp-armv8",
3559 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003560 aarch32_srcs = NEONDOT_UKERNELS,
3561 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3562 aarch64_srcs = NEONDOT_UKERNELS,
3563 copts = [
3564 "-UNDEBUG",
3565 "-DXNN_TEST_MODE=1",
3566 ],
3567 gcc_copts = xnnpack_gcc_std_copts(),
3568 msvc_copts = xnnpack_msvc_std_copts(),
3569 deps = [
3570 ":tables",
3571 "@FP16",
3572 "@pthreadpool",
3573 ],
3574)
3575
3576xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003577 name = "sse2_ukernels",
3578 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003579 gcc_copts = xnnpack_gcc_std_copts(),
3580 gcc_x86_copts = ["-msse2"],
3581 msvc_copts = xnnpack_msvc_std_copts(),
3582 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003583 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003584 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003585 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003586 "@FP16",
3587 "@pthreadpool",
3588 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003589)
3590
3591xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003592 name = "sse2_ukernels_test_mode",
3593 hdrs = INTERNAL_HDRS,
3594 copts = [
3595 "-UNDEBUG",
3596 "-DXNN_TEST_MODE=1",
3597 ],
3598 gcc_copts = xnnpack_gcc_std_copts(),
3599 gcc_x86_copts = ["-msse2"],
3600 msvc_copts = xnnpack_msvc_std_copts(),
3601 msvc_x86_32_copts = ["/arch:SSE2"],
3602 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3603 deps = [
3604 ":tables",
3605 "@FP16",
3606 "@pthreadpool",
3607 ],
3608)
3609
3610xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003611 name = "ssse3_ukernels",
3612 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003613 gcc_copts = xnnpack_gcc_std_copts(),
3614 gcc_x86_copts = ["-mssse3"],
3615 msvc_copts = xnnpack_msvc_std_copts(),
3616 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003617 x86_srcs = SSSE3_UKERNELS,
3618 deps = [
3619 ":tables",
3620 "@FP16",
3621 "@pthreadpool",
3622 ],
3623)
3624
3625xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003626 name = "ssse3_ukernels_test_mode",
3627 hdrs = INTERNAL_HDRS,
3628 copts = [
3629 "-UNDEBUG",
3630 "-DXNN_TEST_MODE=1",
3631 ],
3632 gcc_copts = xnnpack_gcc_std_copts(),
3633 gcc_x86_copts = ["-mssse3"],
3634 msvc_copts = xnnpack_msvc_std_copts(),
3635 msvc_x86_32_copts = ["/arch:SSE2"],
3636 x86_srcs = SSSE3_UKERNELS,
3637 deps = [
3638 ":tables",
3639 "@FP16",
3640 "@pthreadpool",
3641 ],
3642)
3643
3644xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003645 name = "sse41_ukernels",
3646 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003647 gcc_copts = xnnpack_gcc_std_copts(),
3648 gcc_x86_copts = ["-msse4.1"],
3649 msvc_copts = xnnpack_msvc_std_copts(),
3650 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003651 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003652 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003653 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003654 "@FP16",
3655 "@pthreadpool",
3656 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003657)
3658
3659xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003660 name = "sse41_ukernels_test_mode",
3661 hdrs = INTERNAL_HDRS,
3662 copts = [
3663 "-UNDEBUG",
3664 "-DXNN_TEST_MODE=1",
3665 ],
3666 gcc_copts = xnnpack_gcc_std_copts(),
3667 gcc_x86_copts = ["-msse4.1"],
3668 msvc_copts = xnnpack_msvc_std_copts(),
3669 msvc_x86_32_copts = ["/arch:SSE2"],
3670 x86_srcs = SSE41_UKERNELS,
3671 deps = [
3672 ":tables",
3673 "@FP16",
3674 "@pthreadpool",
3675 ],
3676)
3677
3678xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003679 name = "avx_ukernels",
3680 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003681 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"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003686 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003687 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003688 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003689 "@FP16",
3690 "@pthreadpool",
3691 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003692)
3693
3694xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003695 name = "avx_ukernels_test_mode",
3696 hdrs = INTERNAL_HDRS,
3697 copts = [
3698 "-UNDEBUG",
3699 "-DXNN_TEST_MODE=1",
3700 ],
3701 gcc_copts = xnnpack_gcc_std_copts(),
3702 gcc_x86_copts = ["-mavx"],
3703 msvc_copts = xnnpack_msvc_std_copts(),
3704 msvc_x86_32_copts = ["/arch:AVX"],
3705 msvc_x86_64_copts = ["/arch:AVX"],
3706 x86_srcs = AVX_UKERNELS,
3707 deps = [
3708 ":tables",
3709 "@FP16",
3710 "@pthreadpool",
3711 ],
3712)
3713
3714xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003715 name = "xop_ukernels",
3716 hdrs = INTERNAL_HDRS,
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(
3731 name = "xop_ukernels_test_mode",
3732 hdrs = INTERNAL_HDRS,
3733 copts = [
3734 "-UNDEBUG",
3735 "-DXNN_TEST_MODE=1",
3736 ],
3737 gcc_copts = xnnpack_gcc_std_copts(),
3738 gcc_x86_copts = ["-mxop"],
3739 msvc_copts = xnnpack_msvc_std_copts(),
3740 msvc_x86_32_copts = ["/arch:AVX"],
3741 msvc_x86_64_copts = ["/arch:AVX"],
3742 x86_srcs = XOP_UKERNELS,
3743 deps = [
3744 ":tables",
3745 "@FP16",
3746 "@pthreadpool",
3747 ],
3748)
3749
3750xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003751 name = "fma3_ukernels",
3752 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003753 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"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003758 x86_srcs = FMA3_UKERNELS,
3759 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003760 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003761 "@FP16",
3762 "@pthreadpool",
3763 ],
3764)
3765
3766xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003767 name = "fma3_ukernels_test_mode",
3768 hdrs = INTERNAL_HDRS,
3769 copts = [
3770 "-UNDEBUG",
3771 "-DXNN_TEST_MODE=1",
3772 ],
3773 gcc_copts = xnnpack_gcc_std_copts(),
3774 gcc_x86_copts = ["-mfma"],
3775 msvc_copts = xnnpack_msvc_std_copts(),
3776 msvc_x86_32_copts = ["/arch:AVX"],
3777 msvc_x86_64_copts = ["/arch:AVX"],
3778 x86_srcs = FMA3_UKERNELS,
3779 deps = [
3780 ":tables",
3781 "@FP16",
3782 "@pthreadpool",
3783 ],
3784)
3785
3786xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003787 name = "avx2_ukernels",
3788 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003789 gcc_copts = xnnpack_gcc_std_copts(),
3790 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003791 "-mfma",
3792 "-mavx2",
3793 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003794 msvc_copts = xnnpack_msvc_std_copts(),
3795 msvc_x86_32_copts = ["/arch:AVX2"],
3796 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003797 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003798 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003799 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003800 "@FP16",
3801 "@pthreadpool",
3802 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003803)
3804
3805xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003806 name = "avx2_ukernels_test_mode",
3807 hdrs = INTERNAL_HDRS,
3808 copts = [
3809 "-UNDEBUG",
3810 "-DXNN_TEST_MODE=1",
3811 ],
3812 gcc_copts = xnnpack_gcc_std_copts(),
3813 gcc_x86_copts = [
3814 "-mfma",
3815 "-mavx2",
3816 ],
3817 msvc_copts = xnnpack_msvc_std_copts(),
3818 msvc_x86_32_copts = ["/arch:AVX2"],
3819 msvc_x86_64_copts = ["/arch:AVX2"],
3820 x86_srcs = AVX2_UKERNELS,
3821 deps = [
3822 ":tables",
3823 "@FP16",
3824 "@pthreadpool",
3825 ],
3826)
3827
3828xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003829 name = "avx512f_ukernels",
3830 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003831 gcc_copts = xnnpack_gcc_std_copts(),
3832 gcc_x86_copts = ["-mavx512f"],
3833 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3834 msvc_copts = xnnpack_msvc_std_copts(),
3835 msvc_x86_32_copts = ["/arch:AVX512"],
3836 msvc_x86_64_copts = ["/arch:AVX512"],
3837 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003838 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003839 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003840 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003841 "@FP16",
3842 "@pthreadpool",
3843 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003844)
3845
3846xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003847 name = "avx512f_ukernels_test_mode",
3848 hdrs = INTERNAL_HDRS,
3849 copts = [
3850 "-UNDEBUG",
3851 "-DXNN_TEST_MODE=1",
3852 ],
3853 gcc_copts = xnnpack_gcc_std_copts(),
3854 gcc_x86_copts = ["-mavx512f"],
3855 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3856 msvc_copts = xnnpack_msvc_std_copts(),
3857 msvc_x86_32_copts = ["/arch:AVX512"],
3858 msvc_x86_64_copts = ["/arch:AVX512"],
3859 msys_copts = ["-fno-asynchronous-unwind-tables"],
3860 x86_srcs = AVX512F_UKERNELS,
3861 deps = [
3862 ":tables",
3863 "@FP16",
3864 "@pthreadpool",
3865 ],
3866)
3867
3868xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003869 name = "avx512skx_ukernels",
3870 hdrs = INTERNAL_HDRS,
3871 gcc_copts = xnnpack_gcc_std_copts(),
3872 gcc_x86_copts = [
3873 "-mavx512f",
3874 "-mavx512cd",
3875 "-mavx512bw",
3876 "-mavx512dq",
3877 "-mavx512vl",
3878 ],
3879 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3880 msvc_copts = xnnpack_msvc_std_copts(),
3881 msvc_x86_32_copts = ["/arch:AVX512"],
3882 msvc_x86_64_copts = ["/arch:AVX512"],
3883 msys_copts = ["-fno-asynchronous-unwind-tables"],
3884 x86_srcs = AVX512SKX_UKERNELS,
3885 deps = [
3886 ":tables",
3887 "@FP16",
3888 "@pthreadpool",
3889 ],
3890)
3891
3892xnnpack_cc_library(
3893 name = "avx512skx_ukernels_test_mode",
3894 hdrs = INTERNAL_HDRS,
3895 copts = [
3896 "-UNDEBUG",
3897 "-DXNN_TEST_MODE=1",
3898 ],
3899 gcc_copts = xnnpack_gcc_std_copts(),
3900 gcc_x86_copts = [
3901 "-mavx512f",
3902 "-mavx512cd",
3903 "-mavx512bw",
3904 "-mavx512dq",
3905 "-mavx512vl",
3906 ],
3907 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3908 msvc_copts = xnnpack_msvc_std_copts(),
3909 msvc_x86_32_copts = ["/arch:AVX512"],
3910 msvc_x86_64_copts = ["/arch:AVX512"],
3911 msys_copts = ["-fno-asynchronous-unwind-tables"],
3912 x86_srcs = AVX512SKX_UKERNELS,
3913 deps = [
3914 ":tables",
3915 "@FP16",
3916 "@pthreadpool",
3917 ],
3918)
3919
3920xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003921 name = "asm_ukernels",
3922 hdrs = ["src/xnnpack/assembly.h"],
3923 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07003924 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003925 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003926 wasm_srcs = WASM32_ASM_UKERNELS,
3927 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003928)
3929
Marat Dukhan3b59de22020-06-03 20:15:19 -07003930xnnpack_cc_library(
3931 name = "logging_utils",
3932 srcs = LOGGING_SRCS,
3933 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3934 copts = LOGGING_COPTS + [
3935 "-Isrc",
3936 "-Iinclude",
3937 ] + select({
3938 ":debug_build": [],
3939 "//conditions:default": xnnpack_min_size_copts(),
3940 }),
3941 gcc_copts = xnnpack_gcc_std_copts(),
3942 msvc_copts = xnnpack_msvc_std_copts(),
3943 visibility = xnnpack_visibility(),
3944 deps = [
3945 "@FP16",
3946 "@clog",
3947 "@pthreadpool",
3948 ],
3949)
3950
Marat Dukhan08c4a432019-10-03 09:29:21 -07003951xnnpack_aggregate_library(
3952 name = "ukernels",
3953 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003954 ":neon_ukernels",
3955 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003956 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003957 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003958 ":asm_ukernels",
3959 ],
3960 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003961 ":neon_ukernels",
3962 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003963 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003964 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003965 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003966 ":asm_ukernels",
3967 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003968 generic_deps = [
3969 ":scalar_ukernels",
3970 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003971 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003972 ":psimd_fastmath_ukernels",
3973 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003974 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003975 wasm_deps = [
3976 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003977 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003978 ],
3979 wasmsimd_deps = [
3980 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003981 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003982 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003983 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003984 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003985 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003986 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003987 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003988 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003989 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003990 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003991 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003992 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003993 ],
3994)
3995
Marat Dukhan33fcf782020-05-24 14:27:15 -07003996xnnpack_aggregate_library(
3997 name = "ukernels_test_mode",
3998 aarch32_deps = [
3999 ":neon_ukernels_test_mode",
4000 ":neonfma_ukernels_test_mode",
4001 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004002 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004003 ":asm_ukernels",
4004 ],
4005 aarch64_deps = [
4006 ":neon_ukernels_test_mode",
4007 ":neonfma_ukernels_test_mode",
4008 ":neonv8_ukernels_test_mode",
4009 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004010 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004011 ":asm_ukernels",
4012 ],
4013 generic_deps = [
4014 ":scalar_ukernels_test_mode",
4015 ],
4016 psimd_deps = [
4017 ":psimd_fastmath_ukernels_test_mode",
4018 ":psimd_accmath_ukernels_test_mode",
4019 ],
4020 wasm_deps = [
4021 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004022 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004023 ],
4024 wasmsimd_deps = [
4025 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004026 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004027 ],
4028 x86_deps = [
4029 ":sse2_ukernels_test_mode",
4030 ":ssse3_ukernels_test_mode",
4031 ":sse41_ukernels_test_mode",
4032 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004033 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004034 ":fma3_ukernels_test_mode",
4035 ":avx2_ukernels_test_mode",
4036 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004037 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004038 ],
4039)
4040
Marat Dukhan86df3842020-06-17 14:03:31 -07004041# This target must be compatible with Android NDK r17 and compile with Clang 6.
4042# It is similar to :ukernels target, but excludes microkernels which require
4043# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07004044xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07004045 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004046 aarch32_deps = [
4047 ":neon_ukernels",
4048 ":neonfma_ukernels",
4049 ":neonv8_ukernels",
4050 ":asm_ukernels",
4051 ],
4052 aarch64_deps = [
4053 ":neon_ukernels",
4054 ":neonfma_ukernels",
4055 ":neonv8_ukernels",
4056 ":asm_ukernels",
4057 ],
4058 generic_deps = [
4059 ":scalar_ukernels",
4060 ],
4061 psimd_deps = [
4062 ":psimd_fastmath_ukernels",
4063 ":psimd_accmath_ukernels",
4064 ],
4065 wasm_deps = [
4066 ":wasm_ukernels",
4067 ],
4068 wasmsimd_deps = [
4069 ":wasm_ukernels",
4070 ],
4071 x86_deps = [
4072 ":sse2_ukernels",
4073 ":ssse3_ukernels",
4074 ":sse41_ukernels",
4075 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004076 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004077 ":fma3_ukernels",
4078 ":avx2_ukernels",
4079 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004080 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004081 ],
4082)
4083
Marat Dukhan08c4a432019-10-03 09:29:21 -07004084xnnpack_cc_library(
4085 name = "im2col",
4086 srcs = ["src/im2col.c"],
4087 hdrs = [
4088 "src/xnnpack/common.h",
4089 "src/xnnpack/im2col.h",
4090 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004091 gcc_copts = xnnpack_gcc_std_copts(),
4092 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004093)
4094
4095xnnpack_cc_library(
4096 name = "indirection",
4097 srcs = ["src/indirection.c"],
4098 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004099 gcc_copts = xnnpack_gcc_std_copts(),
4100 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004101 deps = [
4102 "@FP16",
4103 "@FXdiv",
4104 "@pthreadpool",
4105 ],
4106)
4107
4108xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004109 name = "indirection_test_mode",
4110 srcs = ["src/indirection.c"],
4111 hdrs = INTERNAL_HDRS,
4112 copts = [
4113 "-UNDEBUG",
4114 "-DXNN_TEST_MODE=1",
4115 ],
4116 gcc_copts = xnnpack_gcc_std_copts(),
4117 msvc_copts = xnnpack_msvc_std_copts(),
4118 deps = [
4119 "@FP16",
4120 "@FXdiv",
4121 "@pthreadpool",
4122 ],
4123)
4124
4125xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004126 name = "packing",
4127 srcs = ["src/packing.c"],
4128 hdrs = INTERNAL_HDRS,
4129 gcc_copts = xnnpack_gcc_std_copts(),
4130 msvc_copts = xnnpack_msvc_std_copts(),
4131 deps = [
4132 "@FP16",
4133 "@FXdiv",
4134 "@pthreadpool",
4135 ],
4136)
4137
4138xnnpack_cc_library(
4139 name = "packing_test_mode",
4140 srcs = ["src/packing.c"],
4141 hdrs = INTERNAL_HDRS,
4142 copts = [
4143 "-UNDEBUG",
4144 "-DXNN_TEST_MODE=1",
4145 ],
4146 gcc_copts = xnnpack_gcc_std_copts(),
4147 msvc_copts = xnnpack_msvc_std_copts(),
4148 deps = [
4149 "@FP16",
4150 "@FXdiv",
4151 "@pthreadpool",
4152 ],
4153)
4154
4155xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004156 name = "operator_run",
4157 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004158 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004159 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004160 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4161 "//conditions:default": [],
4162 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004163 gcc_copts = xnnpack_gcc_std_copts(),
4164 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004165 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004166 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004167 "@FP16",
4168 "@FXdiv",
4169 "@clog",
4170 "@pthreadpool",
4171 ],
4172)
4173
Chao Mei6ddfc602020-05-13 22:29:36 -07004174xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004175 name = "operator_run_test_mode",
4176 srcs = ["src/operator-run.c"],
4177 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4178 copts = LOGGING_COPTS + [
4179 "-UNDEBUG",
4180 "-DXNN_TEST_MODE=1",
4181 ] + select({
4182 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4183 "//conditions:default": [],
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",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004189 "@FP16",
4190 "@FXdiv",
4191 "@clog",
4192 "@pthreadpool",
4193 ],
4194)
4195
4196xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004197 name = "memory_planner",
4198 srcs = ["src/memory-planner.c"],
4199 hdrs = INTERNAL_HDRS,
4200 defines = select({
4201 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4202 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4203 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4204 }),
4205 gcc_copts = xnnpack_gcc_std_copts(),
4206 msvc_copts = xnnpack_msvc_std_copts(),
4207 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004208 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004209 "@pthreadpool",
4210 ],
4211)
4212
Marat Dukhan33fcf782020-05-24 14:27:15 -07004213xnnpack_cc_library(
4214 name = "memory_planner_test_mode",
4215 srcs = ["src/memory-planner.c"],
4216 hdrs = INTERNAL_HDRS,
4217 copts = [
4218 "-UNDEBUG",
4219 "-DXNN_TEST_MODE=1",
4220 ],
4221 defines = select({
4222 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4223 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4224 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4225 }),
4226 gcc_copts = xnnpack_gcc_std_copts(),
4227 msvc_copts = xnnpack_msvc_std_copts(),
4228 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004229 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004230 "@pthreadpool",
4231 ],
4232)
4233
Marat Dukhan08c4a432019-10-03 09:29:21 -07004234cc_library(
4235 name = "enable_assembly",
4236 defines = select({
4237 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4238 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004239 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004240 }),
4241)
4242
Marat Dukhan9de90e02020-06-18 16:04:12 -07004243cc_library(
4244 name = "enable_sparse",
4245 defines = select({
4246 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4247 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07004248 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004249 }),
4250)
4251
Marat Dukhancf056b22019-10-07 10:26:29 -07004252xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004253 name = "operators",
4254 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004255 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004256 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004257 ],
4258 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004259 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004260 "-Isrc",
4261 "-Iinclude",
4262 ] + select({
4263 ":debug_build": [],
4264 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004265 }) + select({
4266 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4267 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004268 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004269 gcc_copts = xnnpack_gcc_std_copts(),
4270 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004271 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004272 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004273 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004274 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004275 "@FP16",
4276 "@FXdiv",
4277 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004278 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004279 ],
4280)
4281
Marat Dukhan10a38082020-04-17 03:58:35 -07004282xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004283 name = "operators_test_mode",
4284 srcs = OPERATOR_SRCS + [
4285 "src/memory.c",
4286 "src/operator-delete.c",
4287 ],
4288 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4289 copts = LOGGING_COPTS + [
4290 "-Isrc",
4291 "-Iinclude",
4292 "-UNDEBUG",
4293 "-DXNN_TEST_MODE=1",
4294 ] + select({
4295 ":debug_build": [],
4296 "//conditions:default": xnnpack_min_size_copts(),
4297 }) + select({
4298 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4299 "//conditions:default": [],
4300 }),
4301 gcc_copts = xnnpack_gcc_std_copts(),
4302 msvc_copts = xnnpack_msvc_std_copts(),
4303 deps = [
4304 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004305 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004306 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004307 "@FP16",
4308 "@FXdiv",
4309 "@clog",
4310 "@pthreadpool",
4311 ],
4312)
4313
4314xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004315 name = "XNNPACK",
4316 srcs = [
4317 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004318 "src/runtime.c",
4319 "src/subgraph.c",
4320 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004321 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004322 hdrs = ["include/xnnpack.h"],
4323 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004324 "-Isrc",
4325 "-Iinclude",
4326 ] + select({
4327 ":debug_build": [],
4328 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004329 }) + select({
4330 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4331 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004332 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004333 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004334 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004335 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004336 visibility = xnnpack_visibility(),
4337 deps = [
4338 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004339 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004340 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004341 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004342 ":operator_run",
4343 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004344 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004345 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004346 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004347 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004348 ] + select({
4349 ":emscripten": [],
4350 "//conditions:default": ["@cpuinfo"],
4351 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004352)
4353
Marat Dukhan10a38082020-04-17 03:58:35 -07004354xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004355 name = "XNNPACK_test_mode",
4356 srcs = [
4357 "src/init.c",
4358 "src/runtime.c",
4359 "src/subgraph.c",
4360 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004361 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004362 hdrs = ["include/xnnpack.h"],
4363 copts = LOGGING_COPTS + [
4364 "-Isrc",
4365 "-Iinclude",
4366 "-UNDEBUG",
4367 "-DXNN_TEST_MODE=1",
4368 ] + select({
4369 ":debug_build": [],
4370 "//conditions:default": xnnpack_min_size_copts(),
4371 }) + select({
4372 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4373 "//conditions:default": [],
4374 }),
4375 gcc_copts = xnnpack_gcc_std_copts(),
4376 includes = ["include"],
4377 msvc_copts = xnnpack_msvc_std_copts(),
4378 visibility = xnnpack_visibility(),
4379 deps = [
4380 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004381 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004382 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004383 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004384 ":operator_run_test_mode",
4385 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004386 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004387 "@clog",
4388 "@FP16",
4389 "@pthreadpool",
4390 ] + select({
4391 ":emscripten": [],
4392 "//conditions:default": ["@cpuinfo"],
4393 }),
4394)
4395
4396xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004397 name = "xnnpack_f32",
4398 srcs = [
4399 "src/init.c",
4400 "src/runtime.c",
4401 "src/subgraph.c",
4402 "src/tensor.c",
4403 ] + SUBGRAPH_SRCS,
4404 hdrs = ["include/xnnpack.h"],
4405 copts = LOGGING_COPTS + [
4406 "-Isrc",
4407 "-Iinclude",
4408 ] + select({
4409 ":debug_build": [],
4410 "//conditions:default": xnnpack_min_size_copts(),
4411 }) + select({
4412 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4413 "//conditions:default": [],
4414 }),
4415 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004416 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004417 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004418 "XNN_NO_U8_OPERATORS",
4419 "XNN_NO_X8_OPERATORS",
4420 "XNN_NO_F16_OPERATORS",
4421 "XNN_NO_X16_OPERATORS",
4422 ],
4423 gcc_copts = xnnpack_gcc_std_copts(),
4424 includes = ["include"],
4425 msvc_copts = xnnpack_msvc_std_copts(),
4426 visibility = xnnpack_visibility(),
4427 deps = [
4428 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004429 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004430 ":logging_utils",
4431 ":memory_planner",
4432 ":operator_run",
4433 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004434 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004435 "@clog",
4436 "@FP16",
4437 "@pthreadpool",
4438 ] + select({
4439 ":emscripten": [],
4440 "//conditions:default": ["@cpuinfo"],
4441 }),
4442)
4443
4444xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004445 name = "xnnpack_operators_nhwc_f32",
4446 srcs = [
4447 "src/init.c",
4448 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004449 hdrs = ["include/xnnpack.h"],
4450 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004451 "-Isrc",
4452 "-Iinclude",
4453 ] + select({
4454 ":debug_build": [],
4455 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004456 }) + select({
4457 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4458 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004459 }),
4460 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004461 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004462 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004463 "XNN_NO_U8_OPERATORS",
4464 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004465 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004466 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004467 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004468 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004469 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004470 visibility = xnnpack_visibility(),
4471 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004472 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004473 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004474 ":operator_run",
4475 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004476 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004477 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004478 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004479 ] + select({
4480 ":emscripten": [],
4481 "//conditions:default": ["@cpuinfo"],
4482 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004483)
4484
Marat Dukhancf056b22019-10-07 10:26:29 -07004485xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004486 name = "bench_utils",
4487 srcs = ["bench/utils.cc"],
4488 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004489 deps = [
4490 "@com_google_benchmark//:benchmark",
4491 "@cpuinfo",
4492 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004493)
4494
Frank Barchard7e955972019-10-11 10:34:25 -07004495######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004496
4497xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004498 name = "qs8_gemm_bench",
4499 srcs = [
4500 "bench/gemm.h",
4501 "bench/qs8-gemm.cc",
4502 "src/xnnpack/AlignedAllocator.h",
4503 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004504 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4505 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004506)
4507
4508xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004509 name = "qs8_requantization_bench",
4510 srcs = [
4511 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004512 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004513 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004514 ] + MICROKERNEL_BENCHMARK_HDRS,
4515 deps = MICROKERNEL_BENCHMARK_DEPS,
4516)
4517
4518xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004519 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004520 srcs = [
4521 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004522 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004523 "src/xnnpack/AlignedAllocator.h",
4524 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004525 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004526 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004527)
4528
4529xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004530 name = "qu8_requantization_bench",
4531 srcs = [
4532 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004533 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004534 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004535 ] + MICROKERNEL_BENCHMARK_HDRS,
4536 deps = MICROKERNEL_BENCHMARK_DEPS,
4537)
4538
4539xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004540 name = "f16_igemm_bench",
4541 srcs = [
4542 "bench/f16-igemm.cc",
4543 "bench/conv.h",
4544 "bench/google/conv.h",
4545 "src/xnnpack/AlignedAllocator.h",
4546 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004547 deps = MICROKERNEL_BENCHMARK_DEPS + [
4548 ":indirection",
4549 ":packing",
4550 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004551)
4552
4553xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004554 name = "f16_gemm_bench",
4555 srcs = [
4556 "bench/f16-gemm.cc",
4557 "bench/gemm.h",
4558 "src/xnnpack/AlignedAllocator.h",
4559 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004560 deps = MICROKERNEL_BENCHMARK_DEPS + [
4561 ":packing",
4562 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004563)
4564
4565xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004566 name = "f16_spmm_bench",
4567 srcs = [
4568 "bench/f16-spmm.cc",
4569 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004570 "src/xnnpack/AlignedAllocator.h",
4571 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004572 deps = MICROKERNEL_BENCHMARK_DEPS,
4573)
4574
4575xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004576 name = "f32_igemm_bench",
4577 srcs = [
4578 "bench/f32-igemm.cc",
4579 "bench/conv.h",
4580 "src/xnnpack/AlignedAllocator.h",
4581 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004582 deps = MICROKERNEL_BENCHMARK_DEPS + [
4583 ":indirection",
4584 ":packing",
4585 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004586)
4587
4588xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004589 name = "f16_relu_bench",
4590 srcs = [
4591 "bench/f16-relu.cc",
4592 "src/xnnpack/AlignedAllocator.h",
4593 ] + MICROKERNEL_BENCHMARK_HDRS,
4594 deps = MICROKERNEL_BENCHMARK_DEPS,
4595)
4596
4597xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004598 name = "f32_conv_hwc_bench",
4599 srcs = [
4600 "bench/f32-conv-hwc.cc",
4601 "bench/dconv.h",
4602 "src/xnnpack/AlignedAllocator.h",
4603 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004604 deps = MICROKERNEL_BENCHMARK_DEPS + [
4605 ":packing",
4606 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004607)
4608
4609xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004610 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004611 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004612 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004613 "bench/dconv.h",
4614 "src/xnnpack/AlignedAllocator.h",
4615 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004616 deps = MICROKERNEL_BENCHMARK_DEPS + [
4617 ":packing",
4618 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004619)
4620
4621xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004622 name = "f16_dwconv_bench",
4623 srcs = [
4624 "bench/f16-dwconv.cc",
4625 "bench/dwconv.h",
4626 "bench/google/dwconv.h",
4627 "src/xnnpack/AlignedAllocator.h",
4628 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004629 deps = MICROKERNEL_BENCHMARK_DEPS + [
4630 ":indirection",
4631 ":packing",
4632 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004633)
4634
4635xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004636 name = "f32_dwconv_bench",
4637 srcs = [
4638 "bench/f32-dwconv.cc",
4639 "bench/dwconv.h",
4640 "src/xnnpack/AlignedAllocator.h",
4641 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004642 deps = MICROKERNEL_BENCHMARK_DEPS + [
4643 ":indirection",
4644 ":packing",
4645 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004646)
4647
4648xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07004649 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004650 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07004651 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004652 "bench/dwconv.h",
4653 "src/xnnpack/AlignedAllocator.h",
4654 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004655 deps = MICROKERNEL_BENCHMARK_DEPS + [
4656 ":indirection",
4657 ":packing",
4658 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004659)
4660
4661xnnpack_benchmark(
4662 name = "f32_gemm_bench",
4663 srcs = [
4664 "bench/f32-gemm.cc",
4665 "bench/gemm.h",
4666 "src/xnnpack/AlignedAllocator.h",
4667 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004668 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004669 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004670)
4671
4672xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004673 name = "f32_hswish_bench",
4674 srcs = [
4675 "bench/f32-hswish.cc",
4676 "src/xnnpack/AlignedAllocator.h",
4677 ] + MICROKERNEL_BENCHMARK_HDRS,
4678 deps = MICROKERNEL_BENCHMARK_DEPS,
4679)
4680
4681xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004682 name = "f32_raddexpminusmax_bench",
4683 srcs = [
4684 "bench/f32-raddexpminusmax.cc",
4685 "src/xnnpack/AlignedAllocator.h",
4686 ] + MICROKERNEL_BENCHMARK_HDRS,
4687 deps = MICROKERNEL_BENCHMARK_DEPS,
4688)
4689
4690xnnpack_benchmark(
4691 name = "f32_raddextexp_bench",
4692 srcs = [
4693 "bench/f32-raddextexp.cc",
4694 "src/xnnpack/AlignedAllocator.h",
4695 ] + MICROKERNEL_BENCHMARK_HDRS,
4696 deps = MICROKERNEL_BENCHMARK_DEPS,
4697)
4698
4699xnnpack_benchmark(
4700 name = "f32_raddstoreexpminusmax_bench",
4701 srcs = [
4702 "bench/f32-raddstoreexpminusmax.cc",
4703 "src/xnnpack/AlignedAllocator.h",
4704 ] + MICROKERNEL_BENCHMARK_HDRS,
4705 deps = MICROKERNEL_BENCHMARK_DEPS,
4706)
4707
4708xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004709 name = "f32_relu_bench",
4710 srcs = [
4711 "bench/f32-relu.cc",
4712 "src/xnnpack/AlignedAllocator.h",
4713 ] + MICROKERNEL_BENCHMARK_HDRS,
4714 deps = MICROKERNEL_BENCHMARK_DEPS,
4715)
4716
4717xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004718 name = "f32_rmax_bench",
4719 srcs = [
4720 "bench/f32-rmax.cc",
4721 "src/xnnpack/AlignedAllocator.h",
4722 ] + MICROKERNEL_BENCHMARK_HDRS,
4723 deps = MICROKERNEL_BENCHMARK_DEPS,
4724)
4725
4726xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004727 name = "f32_sigmoid_bench",
4728 srcs = [
4729 "bench/f32-sigmoid.cc",
4730 "src/xnnpack/AlignedAllocator.h",
4731 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004732 deps = MICROKERNEL_BENCHMARK_DEPS,
4733)
4734
4735xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004736 name = "f32_spmm_bench",
4737 srcs = [
4738 "bench/f32-spmm.cc",
4739 "bench/gemm.h",
4740 "src/xnnpack/AlignedAllocator.h",
4741 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004742 deps = MICROKERNEL_BENCHMARK_DEPS,
4743)
4744
4745xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004746 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004747 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004748 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004749 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004750 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004751 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004752)
4753
4754xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004755 name = "f32_vscaleexpminusmax_bench",
4756 srcs = [
4757 "bench/f32-vscaleexpminusmax.cc",
4758 "src/xnnpack/AlignedAllocator.h",
4759 ] + MICROKERNEL_BENCHMARK_HDRS,
4760 deps = MICROKERNEL_BENCHMARK_DEPS,
4761)
4762
4763xnnpack_benchmark(
4764 name = "f32_vscaleextexp_bench",
4765 srcs = [
4766 "bench/f32-vscaleextexp.cc",
4767 "src/xnnpack/AlignedAllocator.h",
4768 ] + MICROKERNEL_BENCHMARK_HDRS,
4769 deps = MICROKERNEL_BENCHMARK_DEPS,
4770)
4771
4772xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004773 name = "f32_vsqrt_bench",
4774 srcs = [
4775 "bench/f32-vsqrt.cc",
4776 "src/xnnpack/AlignedAllocator.h",
4777 ] + MICROKERNEL_BENCHMARK_HDRS,
4778 deps = MICROKERNEL_BENCHMARK_DEPS,
4779)
4780
4781xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004782 name = "f32_im2col_gemm_bench",
4783 srcs = [
4784 "bench/f32-im2col-gemm.cc",
4785 "bench/conv.h",
4786 "src/xnnpack/AlignedAllocator.h",
4787 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004788 deps = MICROKERNEL_BENCHMARK_DEPS + [
4789 ":im2col",
4790 ":packing",
4791 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004792)
4793
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004794xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004795 name = "rounding_bench",
4796 srcs = [
4797 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004798 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004799 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004800 ] + MICROKERNEL_BENCHMARK_HDRS,
4801 deps = MICROKERNEL_BENCHMARK_DEPS,
4802)
4803
Marat Dukhan08c4a432019-10-03 09:29:21 -07004804########################### Benchmarks for operators ###########################
4805
4806xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004807 name = "average_pooling_bench",
4808 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004809 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004810 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004811 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004812)
4813
4814xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004815 name = "bankers_rounding_bench",
4816 srcs = ["bench/bankers-rounding.cc"],
4817 copts = xnnpack_optional_tflite_copts(),
4818 tags = ["nowin32"],
4819 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4820)
4821
4822xnnpack_benchmark(
4823 name = "ceiling_bench",
4824 srcs = ["bench/ceiling.cc"],
4825 copts = xnnpack_optional_tflite_copts(),
4826 tags = ["nowin32"],
4827 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4828)
4829
4830xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004831 name = "channel_shuffle_bench",
4832 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004833 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004834)
4835
4836xnnpack_benchmark(
4837 name = "convolution_bench",
4838 srcs = ["bench/convolution.cc"],
4839 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004840 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004841 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004842)
4843
4844xnnpack_benchmark(
4845 name = "deconvolution_bench",
4846 srcs = ["bench/deconvolution.cc"],
4847 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004848 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004849 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004850)
4851
4852xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004853 name = "floor_bench",
4854 srcs = ["bench/floor.cc"],
4855 copts = xnnpack_optional_tflite_copts(),
4856 tags = ["nowin32"],
4857 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4858)
4859
4860xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004861 name = "global_average_pooling_bench",
4862 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004863 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004864)
4865
4866xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004867 name = "hardswish_bench",
4868 srcs = ["bench/hardswish.cc"],
4869 copts = xnnpack_optional_tflite_copts(),
4870 tags = ["nowin32"],
4871 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4872)
4873
4874xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004875 name = "max_pooling_bench",
4876 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004877 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004878)
4879
4880xnnpack_benchmark(
4881 name = "sigmoid_bench",
4882 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004883 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004884 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004885 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004886)
4887
4888xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004889 name = "prelu_bench",
4890 srcs = ["bench/prelu.cc"],
4891 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004892 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004893 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004894)
4895
4896xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004897 name = "softmax_bench",
4898 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004899 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004900 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004901 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004902)
4903
Marat Dukhan87727142020-06-24 15:24:10 -07004904xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004905 name = "square_root_bench",
4906 srcs = ["bench/square-root.cc"],
4907 copts = xnnpack_optional_tflite_copts(),
4908 tags = ["nowin32"],
4909 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4910)
4911
4912xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004913 name = "truncation_bench",
4914 srcs = ["bench/truncation.cc"],
4915 deps = OPERATOR_BENCHMARK_DEPS,
4916)
4917
Marat Dukhanc068bb62019-10-04 13:24:39 -07004918############################# End-to-end benchmarks ############################
4919
4920cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004921 name = "fp32_mobilenet_v1",
4922 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004923 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004924 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004925 linkstatic = True,
4926 deps = [
4927 ":XNNPACK",
4928 "@pthreadpool",
4929 ],
4930)
4931
4932cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004933 name = "fp16_mobilenet_v1",
4934 srcs = ["models/fp16-mobilenet-v1.cc"],
4935 hdrs = ["models/models.h"],
4936 copts = xnnpack_std_cxxopts(),
4937 linkstatic = True,
4938 deps = [
4939 ":XNNPACK",
4940 "@FP16",
4941 "@pthreadpool",
4942 ],
4943)
4944
4945cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004946 name = "qs8_mobilenet_v1",
4947 srcs = ["models/qs8-mobilenet-v1.cc"],
4948 hdrs = ["models/models.h"],
4949 copts = xnnpack_std_cxxopts(),
4950 linkstatic = True,
4951 deps = [
4952 ":XNNPACK",
4953 "@pthreadpool",
4954 ],
4955)
4956
4957cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004958 name = "qs8_mobilenet_v2",
4959 srcs = ["models/qs8-mobilenet-v2.cc"],
4960 hdrs = ["models/models.h"],
4961 copts = xnnpack_std_cxxopts(),
4962 linkstatic = True,
4963 deps = [
4964 ":XNNPACK",
4965 "@pthreadpool",
4966 ],
4967)
4968
4969cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004970 name = "fp32_mobilenet_v2",
4971 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004972 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004973 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004974 linkstatic = True,
4975 deps = [
4976 ":XNNPACK",
4977 "@pthreadpool",
4978 ],
4979)
4980
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004981cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004982 name = "fp16_mobilenet_v2",
4983 srcs = ["models/fp16-mobilenet-v2.cc"],
4984 hdrs = ["models/models.h"],
4985 copts = xnnpack_std_cxxopts(),
4986 linkstatic = True,
4987 deps = [
4988 ":XNNPACK",
4989 "@FP16",
4990 "@pthreadpool",
4991 ],
4992)
4993
4994cc_library(
4995 name = "fp32_mobilenet_v3_large",
4996 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004997 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004998 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004999 linkstatic = True,
5000 deps = [
5001 ":XNNPACK",
5002 "@pthreadpool",
5003 ],
5004)
5005
5006cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005007 name = "fp16_mobilenet_v3_large",
5008 srcs = ["models/fp16-mobilenet-v3-large.cc"],
5009 hdrs = ["models/models.h"],
5010 copts = xnnpack_std_cxxopts(),
5011 linkstatic = True,
5012 deps = [
5013 ":XNNPACK",
5014 "@FP16",
5015 "@pthreadpool",
5016 ],
5017)
5018
5019cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005020 name = "fp32_mobilenet_v3_small",
5021 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005022 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005023 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005024 linkstatic = True,
5025 deps = [
5026 ":XNNPACK",
5027 "@pthreadpool",
5028 ],
5029)
5030
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005031cc_library(
5032 name = "fp16_mobilenet_v3_small",
5033 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5034 hdrs = ["models/models.h"],
5035 copts = xnnpack_std_cxxopts(),
5036 linkstatic = True,
5037 deps = [
5038 ":XNNPACK",
5039 "@FP16",
5040 "@pthreadpool",
5041 ],
5042)
5043
Marat Dukhanc068bb62019-10-04 13:24:39 -07005044xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005045 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005046 srcs = [
5047 "bench/f32-dwconv-e2e.cc",
5048 "bench/end2end.h",
5049 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005050 deps = MICROKERNEL_BENCHMARK_DEPS + [
5051 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005052 ":fp32_mobilenet_v1",
5053 ":fp32_mobilenet_v2",
5054 ":fp32_mobilenet_v3_large",
5055 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005056 ],
5057)
5058
5059xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005060 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005061 srcs = [
5062 "bench/f32-gemm-e2e.cc",
5063 "bench/end2end.h",
5064 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005065 deps = MICROKERNEL_BENCHMARK_DEPS + [
5066 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005067 ":fp32_mobilenet_v1",
5068 ":fp32_mobilenet_v2",
5069 ":fp32_mobilenet_v3_large",
5070 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005071 ],
5072)
5073
5074xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005075 name = "end2end_bench",
5076 srcs = ["bench/end2end.cc"],
5077 deps = [
5078 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005079 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005080 ":fp16_mobilenet_v1",
5081 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005082 ":fp16_mobilenet_v3_large",
5083 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005084 ":fp32_mobilenet_v1",
5085 ":fp32_mobilenet_v2",
5086 ":fp32_mobilenet_v3_large",
5087 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005088 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005089 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005090 "@pthreadpool",
5091 ],
5092)
5093
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005094#################### Accuracy evaluation for math functions ####################
5095
5096xnnpack_benchmark(
5097 name = "f32_exp_eval",
5098 srcs = [
5099 "eval/f32-exp.cc",
5100 "src/xnnpack/AlignedAllocator.h",
5101 ] + ACCURACY_EVAL_HDRS,
5102 deps = ACCURACY_EVAL_DEPS,
5103)
5104
Marat Dukhan515c9772019-10-17 18:07:57 -07005105xnnpack_benchmark(
5106 name = "f32_expminus_eval",
5107 srcs = [
5108 "eval/f32-expminus.cc",
5109 "src/xnnpack/AlignedAllocator.h",
5110 ] + ACCURACY_EVAL_HDRS,
5111 deps = ACCURACY_EVAL_DEPS,
5112)
5113
Marat Dukhan98ba4412019-10-23 02:14:28 -07005114xnnpack_benchmark(
5115 name = "f32_extexp_eval",
5116 srcs = [
5117 "eval/f32-extexp.cc",
5118 "src/xnnpack/AlignedAllocator.h",
5119 ] + ACCURACY_EVAL_HDRS,
5120 deps = ACCURACY_EVAL_DEPS,
5121)
5122
Marat Dukhan8853b822020-05-07 12:19:01 -07005123xnnpack_unit_test(
5124 name = "f32_roundne_eval",
5125 srcs = [
5126 "eval/f32-roundne.cc",
5127 "src/xnnpack/AlignedAllocator.h",
5128 "src/xnnpack/math-stubs.h",
5129 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005130 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005131 deps = MICROKERNEL_TEST_DEPS,
5132)
5133
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005134xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005135 name = "f32_roundd_eval",
5136 srcs = [
5137 "eval/f32-roundd.cc",
5138 "src/xnnpack/AlignedAllocator.h",
5139 "src/xnnpack/math-stubs.h",
5140 ] + MICROKERNEL_TEST_HDRS,
5141 automatic = False,
5142 deps = MICROKERNEL_TEST_DEPS,
5143)
5144
5145xnnpack_unit_test(
5146 name = "f32_roundu_eval",
5147 srcs = [
5148 "eval/f32-roundu.cc",
5149 "src/xnnpack/AlignedAllocator.h",
5150 "src/xnnpack/math-stubs.h",
5151 ] + MICROKERNEL_TEST_HDRS,
5152 automatic = False,
5153 deps = MICROKERNEL_TEST_DEPS,
5154)
5155
5156xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005157 name = "f32_roundz_eval",
5158 srcs = [
5159 "eval/f32-roundz.cc",
5160 "src/xnnpack/AlignedAllocator.h",
5161 "src/xnnpack/math-stubs.h",
5162 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005163 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005164 deps = MICROKERNEL_TEST_DEPS,
5165)
5166
Marat Dukhan346a9e52019-11-15 09:06:30 -08005167xnnpack_benchmark(
5168 name = "f32_sigmoid_eval",
5169 srcs = [
5170 "eval/f32-sigmoid.cc",
5171 "src/xnnpack/AlignedAllocator.h",
5172 ] + ACCURACY_EVAL_HDRS,
5173 deps = ACCURACY_EVAL_DEPS,
5174)
5175
Marat Dukhan84000762020-06-29 18:38:43 -07005176xnnpack_benchmark(
5177 name = "f32_sqrt_eval",
5178 srcs = [
5179 "eval/f32-sqrt.cc",
5180 "src/xnnpack/AlignedAllocator.h",
5181 ] + ACCURACY_EVAL_HDRS,
5182 deps = ACCURACY_EVAL_DEPS + [
5183 ":bench_utils",
5184 ],
5185)
5186
Marat Dukhan08c4a432019-10-03 09:29:21 -07005187######################### Unit tests for micro-kernels #########################
5188
5189xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005190 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005191 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005192 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005193 "test/gemm-microkernel-tester.h",
5194 "src/xnnpack/AlignedAllocator.h",
5195 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005196 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005197)
5198
5199xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005200 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005201 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005202 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005203 "test/spmm-microkernel-tester.h",
5204 "src/xnnpack/AlignedAllocator.h",
5205 ] + MICROKERNEL_TEST_HDRS,
5206 deps = MICROKERNEL_TEST_DEPS,
5207)
5208
5209xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005210 name = "f32_argmaxpool_test",
5211 srcs = [
5212 "test/f32-argmaxpool.cc",
5213 "test/argmaxpool-microkernel-tester.h",
5214 "src/xnnpack/AlignedAllocator.h",
5215 ] + MICROKERNEL_TEST_HDRS,
5216 deps = MICROKERNEL_TEST_DEPS,
5217)
5218
5219xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005220 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005221 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005222 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005223 "test/avgpool-microkernel-tester.h",
5224 "src/xnnpack/AlignedAllocator.h",
5225 ] + MICROKERNEL_TEST_HDRS,
5226 deps = MICROKERNEL_TEST_DEPS,
5227)
5228
5229xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005230 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005231 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005232 "test/f32-ibilinear.cc",
5233 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005234 "src/xnnpack/AlignedAllocator.h",
5235 ] + MICROKERNEL_TEST_HDRS,
5236 deps = MICROKERNEL_TEST_DEPS,
5237)
5238
5239xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005240 name = "f32_ibilinear_chw_test",
5241 srcs = [
5242 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005243 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005244 "src/xnnpack/AlignedAllocator.h",
5245 ] + MICROKERNEL_TEST_HDRS,
5246 deps = MICROKERNEL_TEST_DEPS,
5247)
5248
5249xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005250 name = "f16_clamp_test",
5251 srcs = [
5252 "test/f16-clamp.cc",
5253 "test/clamp-microkernel-tester.h",
5254 ] + MICROKERNEL_TEST_HDRS,
5255 deps = MICROKERNEL_TEST_DEPS,
5256)
5257
5258xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005259 name = "f32_clamp_test",
5260 srcs = [
5261 "test/f32-clamp.cc",
5262 "test/clamp-microkernel-tester.h",
5263 ] + MICROKERNEL_TEST_HDRS,
5264 deps = MICROKERNEL_TEST_DEPS,
5265)
5266
5267xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005268 name = "f32_igemm_test",
5269 srcs = [
5270 "test/f32-igemm.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 Dukhan163a7e62020-04-09 04:19:26 -07005275)
5276
5277xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005278 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005279 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005280 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005281 "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"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005285)
5286
5287xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005288 name = "f32_igemm_minmax_test",
5289 srcs = [
5290 "test/f32-igemm-minmax.cc",
5291 "test/gemm-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 Dukhane207b7b2020-05-28 16:27:42 -07005295)
5296
5297xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005298 name = "f16_igemm_minmax_test",
5299 srcs = [
5300 "test/f16-igemm-minmax.cc",
5301 "test/gemm-microkernel-tester.h",
5302 "src/xnnpack/AlignedAllocator.h",
5303 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005304 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005305)
5306
5307xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005308 name = "f32_conv_hwc_test",
5309 srcs = [
5310 "test/f32-conv-hwc.cc",
5311 "test/conv-hwc-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 Dukhan08c4a432019-10-03 09:29:21 -07005315)
5316
5317xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005318 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005319 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005320 "test/f32-conv-hwc2chw.cc",
5321 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005322 "src/xnnpack/AlignedAllocator.h",
5323 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005324 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005325)
5326
5327xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005328 name = "f32_dwconv_test",
5329 srcs = [
5330 "test/f32-dwconv.cc",
5331 "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 Dukhan163a7e62020-04-09 04:19:26 -07005335)
5336
5337xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005338 name = "f16_dwconv_minmax_test",
5339 srcs = [
5340 "test/f16-dwconv-minmax.cc",
5341 "test/dwconv-microkernel-tester.h",
5342 "src/xnnpack/AlignedAllocator.h",
5343 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005344 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005345)
5346
5347xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005348 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005349 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005350 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005351 "test/dwconv-microkernel-tester.h",
5352 "src/xnnpack/AlignedAllocator.h",
5353 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005354 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005355)
5356
5357xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005358 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005359 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005360 "test/f32-dwconv2d-chw.cc",
5361 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005362 "src/xnnpack/AlignedAllocator.h",
5363 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005364 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005365)
5366
5367xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005368 name = "f16_gavgpool_minmax_test",
5369 srcs = [
5370 "test/f16-gavgpool-minmax.cc",
5371 "test/gavgpool-microkernel-tester.h",
5372 "src/xnnpack/AlignedAllocator.h",
5373 ] + MICROKERNEL_TEST_HDRS,
5374 deps = MICROKERNEL_TEST_DEPS,
5375)
5376
5377xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005378 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005379 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005380 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005381 "test/gavgpool-microkernel-tester.h",
5382 "src/xnnpack/AlignedAllocator.h",
5383 ] + MICROKERNEL_TEST_HDRS,
5384 deps = MICROKERNEL_TEST_DEPS,
5385)
5386
5387xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005388 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005389 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005390 "test/f32-gavgpool-cw.cc",
5391 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005392 "src/xnnpack/AlignedAllocator.h",
5393 ] + MICROKERNEL_TEST_HDRS,
5394 deps = MICROKERNEL_TEST_DEPS,
5395)
5396
5397xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005398 name = "f32_gemm_test",
5399 srcs = [
5400 "test/f32-gemm.cc",
5401 "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 Dukhan163a7e62020-04-09 04:19:26 -07005405)
5406
5407xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005408 name = "f32_gemm_relu_test",
5409 srcs = [
5410 "test/f32-gemm-relu.cc",
5411 "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 Dukhan467f6362020-05-22 23:21:55 -07005415)
5416
5417xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005418 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005419 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005420 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005421 "test/gemm-microkernel-tester.h",
5422 "src/xnnpack/AlignedAllocator.h",
5423 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005424 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005425)
5426
5427xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005428 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005429 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005430 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005431 "test/gemm-microkernel-tester.h",
5432 "src/xnnpack/AlignedAllocator.h",
5433 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005434 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005435)
5436
5437xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005438 name = "f16_hswish_test",
5439 srcs = [
5440 "test/f16-hswish.cc",
5441 "test/hswish-microkernel-tester.h",
5442 ] + MICROKERNEL_TEST_HDRS,
5443 deps = MICROKERNEL_TEST_DEPS,
5444)
5445
5446xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005447 name = "f32_hswish_test",
5448 srcs = [
5449 "test/f32-hswish.cc",
5450 "test/hswish-microkernel-tester.h",
5451 ] + MICROKERNEL_TEST_HDRS,
5452 deps = MICROKERNEL_TEST_DEPS,
5453)
5454
5455xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005456 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005457 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005458 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005459 "test/maxpool-microkernel-tester.h",
5460 ] + MICROKERNEL_TEST_HDRS,
5461 deps = MICROKERNEL_TEST_DEPS,
5462)
5463
5464xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005465 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005466 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005467 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005468 "test/avgpool-microkernel-tester.h",
5469 "src/xnnpack/AlignedAllocator.h",
5470 ] + MICROKERNEL_TEST_HDRS,
5471 deps = MICROKERNEL_TEST_DEPS,
5472)
5473
5474xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005475 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005476 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005477 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005478 "test/gemm-microkernel-tester.h",
5479 "src/xnnpack/AlignedAllocator.h",
5480 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005481 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005482)
5483
5484xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005485 name = "f16_prelu_test",
5486 srcs = [
5487 "test/f16-prelu.cc",
5488 "test/prelu-microkernel-tester.h",
5489 "src/xnnpack/AlignedAllocator.h",
5490 ] + MICROKERNEL_TEST_HDRS,
5491 deps = MICROKERNEL_TEST_DEPS,
5492)
5493
5494xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005495 name = "f32_prelu_test",
5496 srcs = [
5497 "test/f32-prelu.cc",
5498 "test/prelu-microkernel-tester.h",
5499 "src/xnnpack/AlignedAllocator.h",
5500 ] + MICROKERNEL_TEST_HDRS,
5501 deps = MICROKERNEL_TEST_DEPS,
5502)
5503
5504xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005505 name = "f32_raddexpminusmax_test",
5506 srcs = [
5507 "test/f32-raddexpminusmax.cc",
5508 "test/raddexpminusmax-microkernel-tester.h",
5509 ] + MICROKERNEL_TEST_HDRS,
5510 deps = MICROKERNEL_TEST_DEPS,
5511)
5512
5513xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005514 name = "f32_raddextexp_test",
5515 srcs = [
5516 "test/f32-raddextexp.cc",
5517 "test/raddextexp-microkernel-tester.h",
5518 ] + MICROKERNEL_TEST_HDRS,
5519 deps = MICROKERNEL_TEST_DEPS,
5520)
5521
5522xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005523 name = "f32_raddstoreexpminusmax_test",
5524 srcs = [
5525 "test/f32-raddstoreexpminusmax.cc",
5526 "test/raddstoreexpminusmax-microkernel-tester.h",
5527 ] + MICROKERNEL_TEST_HDRS,
5528 deps = MICROKERNEL_TEST_DEPS,
5529)
5530
5531xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005532 name = "f32_relu_test",
5533 srcs = [
5534 "test/f32-relu.cc",
5535 "test/vunary-microkernel-tester.h",
5536 ] + MICROKERNEL_TEST_HDRS,
5537 deps = MICROKERNEL_TEST_DEPS,
5538)
5539
5540xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005541 name = "f32_rmax_test",
5542 srcs = [
5543 "test/f32-rmax.cc",
5544 "test/rmax-microkernel-tester.h",
5545 ] + MICROKERNEL_TEST_HDRS,
5546 deps = MICROKERNEL_TEST_DEPS,
5547)
5548
5549xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005550 name = "f32_sigmoid_test",
5551 srcs = [
5552 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005553 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005554 ] + MICROKERNEL_TEST_HDRS,
5555 deps = MICROKERNEL_TEST_DEPS,
5556)
5557
5558xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005559 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005560 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005561 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005562 "test/spmm-microkernel-tester.h",
5563 "src/xnnpack/AlignedAllocator.h",
5564 ] + MICROKERNEL_TEST_HDRS,
5565 deps = MICROKERNEL_TEST_DEPS,
5566)
5567
5568xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005569 name = "f16_vadd_minmax_test",
5570 srcs = [
5571 "test/f16-vadd-minmax.cc",
5572 "test/vbinary-microkernel-tester.h",
5573 ] + MICROKERNEL_TEST_HDRS,
5574 deps = MICROKERNEL_TEST_DEPS,
5575)
5576
5577xnnpack_unit_test(
5578 name = "f16_vaddc_minmax_test",
5579 srcs = [
5580 "test/f16-vaddc-minmax.cc",
5581 "test/vbinaryc-microkernel-tester.h",
5582 ] + MICROKERNEL_TEST_HDRS,
5583 deps = MICROKERNEL_TEST_DEPS,
5584)
5585
5586xnnpack_unit_test(
5587 name = "f16_vdiv_minmax_test",
5588 srcs = [
5589 "test/f16-vdiv-minmax.cc",
5590 "test/vbinary-microkernel-tester.h",
5591 ] + MICROKERNEL_TEST_HDRS,
5592 deps = MICROKERNEL_TEST_DEPS,
5593)
5594
5595xnnpack_unit_test(
5596 name = "f16_vdivc_minmax_test",
5597 srcs = [
5598 "test/f16-vdivc-minmax.cc",
5599 "test/vbinaryc-microkernel-tester.h",
5600 ] + MICROKERNEL_TEST_HDRS,
5601 deps = MICROKERNEL_TEST_DEPS,
5602)
5603
5604xnnpack_unit_test(
5605 name = "f16_vrdivc_minmax_test",
5606 srcs = [
5607 "test/f16-vrdivc-minmax.cc",
5608 "test/vbinaryc-microkernel-tester.h",
5609 ] + MICROKERNEL_TEST_HDRS,
5610 deps = MICROKERNEL_TEST_DEPS,
5611)
5612
5613xnnpack_unit_test(
5614 name = "f16_vmax_test",
5615 srcs = [
5616 "test/f16-vmax.cc",
5617 "test/vbinary-microkernel-tester.h",
5618 ] + MICROKERNEL_TEST_HDRS,
5619 deps = MICROKERNEL_TEST_DEPS,
5620)
5621
5622xnnpack_unit_test(
5623 name = "f16_vmaxc_test",
5624 srcs = [
5625 "test/f16-vmaxc.cc",
5626 "test/vbinaryc-microkernel-tester.h",
5627 ] + MICROKERNEL_TEST_HDRS,
5628 deps = MICROKERNEL_TEST_DEPS,
5629)
5630
5631xnnpack_unit_test(
5632 name = "f16_vmin_test",
5633 srcs = [
5634 "test/f16-vmin.cc",
5635 "test/vbinary-microkernel-tester.h",
5636 ] + MICROKERNEL_TEST_HDRS,
5637 deps = MICROKERNEL_TEST_DEPS,
5638)
5639
5640xnnpack_unit_test(
5641 name = "f16_vminc_test",
5642 srcs = [
5643 "test/f16-vminc.cc",
5644 "test/vbinaryc-microkernel-tester.h",
5645 ] + MICROKERNEL_TEST_HDRS,
5646 deps = MICROKERNEL_TEST_DEPS,
5647)
5648
5649xnnpack_unit_test(
5650 name = "f16_vmul_minmax_test",
5651 srcs = [
5652 "test/f16-vmul-minmax.cc",
5653 "test/vbinary-microkernel-tester.h",
5654 ] + MICROKERNEL_TEST_HDRS,
5655 deps = MICROKERNEL_TEST_DEPS,
5656)
5657
5658xnnpack_unit_test(
5659 name = "f16_vmulc_minmax_test",
5660 srcs = [
5661 "test/f16-vmulc-minmax.cc",
5662 "test/vbinaryc-microkernel-tester.h",
5663 ] + MICROKERNEL_TEST_HDRS,
5664 deps = MICROKERNEL_TEST_DEPS,
5665)
5666
5667xnnpack_unit_test(
5668 name = "f16_vsub_minmax_test",
5669 srcs = [
5670 "test/f16-vsub-minmax.cc",
5671 "test/vbinary-microkernel-tester.h",
5672 ] + MICROKERNEL_TEST_HDRS,
5673 deps = MICROKERNEL_TEST_DEPS,
5674)
5675
5676xnnpack_unit_test(
5677 name = "f16_vsubc_minmax_test",
5678 srcs = [
5679 "test/f16-vsubc-minmax.cc",
5680 "test/vbinaryc-microkernel-tester.h",
5681 ] + MICROKERNEL_TEST_HDRS,
5682 deps = MICROKERNEL_TEST_DEPS,
5683)
5684
5685xnnpack_unit_test(
5686 name = "f16_vrsubc_minmax_test",
5687 srcs = [
5688 "test/f16-vrsubc-minmax.cc",
5689 "test/vbinaryc-microkernel-tester.h",
5690 ] + MICROKERNEL_TEST_HDRS,
5691 deps = MICROKERNEL_TEST_DEPS,
5692)
5693
5694xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005695 name = "f32_vabs_test",
5696 srcs = [
5697 "test/f32-vabs.cc",
5698 "test/vunary-microkernel-tester.h",
5699 ] + MICROKERNEL_TEST_HDRS,
5700 deps = MICROKERNEL_TEST_DEPS,
5701)
5702
5703xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005704 name = "f32_vadd_test",
5705 srcs = [
5706 "test/f32-vadd.cc",
5707 "test/vbinary-microkernel-tester.h",
5708 ] + MICROKERNEL_TEST_HDRS,
5709 deps = MICROKERNEL_TEST_DEPS,
5710)
5711
5712xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005713 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005714 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005715 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005716 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005717 ] + MICROKERNEL_TEST_HDRS,
5718 deps = MICROKERNEL_TEST_DEPS,
5719)
5720
5721xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005722 name = "f32_vadd_relu_test",
5723 srcs = [
5724 "test/f32-vadd-relu.cc",
5725 "test/vbinary-microkernel-tester.h",
5726 ] + MICROKERNEL_TEST_HDRS,
5727 deps = MICROKERNEL_TEST_DEPS,
5728)
5729
5730xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005731 name = "f32_vaddc_test",
5732 srcs = [
5733 "test/f32-vaddc.cc",
5734 "test/vbinaryc-microkernel-tester.h",
5735 ] + MICROKERNEL_TEST_HDRS,
5736 deps = MICROKERNEL_TEST_DEPS,
5737)
5738
5739xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005740 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005741 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005742 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005743 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005744 ] + MICROKERNEL_TEST_HDRS,
5745 deps = MICROKERNEL_TEST_DEPS,
5746)
5747
5748xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005749 name = "f32_vaddc_relu_test",
5750 srcs = [
5751 "test/f32-vaddc-relu.cc",
5752 "test/vbinaryc-microkernel-tester.h",
5753 ] + MICROKERNEL_TEST_HDRS,
5754 deps = MICROKERNEL_TEST_DEPS,
5755)
5756
5757xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005758 name = "f32_vdiv_test",
5759 srcs = [
5760 "test/f32-vdiv.cc",
5761 "test/vbinary-microkernel-tester.h",
5762 ] + MICROKERNEL_TEST_HDRS,
5763 deps = MICROKERNEL_TEST_DEPS,
5764)
5765
5766xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005767 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005768 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005769 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005770 "test/vbinary-microkernel-tester.h",
5771 ] + MICROKERNEL_TEST_HDRS,
5772 deps = MICROKERNEL_TEST_DEPS,
5773)
5774
5775xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005776 name = "f32_vdiv_relu_test",
5777 srcs = [
5778 "test/f32-vdiv-relu.cc",
5779 "test/vbinary-microkernel-tester.h",
5780 ] + MICROKERNEL_TEST_HDRS,
5781 deps = MICROKERNEL_TEST_DEPS,
5782)
5783
5784xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005785 name = "f32_vdivc_test",
5786 srcs = [
5787 "test/f32-vdivc.cc",
5788 "test/vbinaryc-microkernel-tester.h",
5789 ] + MICROKERNEL_TEST_HDRS,
5790 deps = MICROKERNEL_TEST_DEPS,
5791)
5792
5793xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005794 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005795 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005796 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005797 "test/vbinaryc-microkernel-tester.h",
5798 ] + MICROKERNEL_TEST_HDRS,
5799 deps = MICROKERNEL_TEST_DEPS,
5800)
5801
5802xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005803 name = "f32_vdivc_relu_test",
5804 srcs = [
5805 "test/f32-vdivc-relu.cc",
5806 "test/vbinaryc-microkernel-tester.h",
5807 ] + MICROKERNEL_TEST_HDRS,
5808 deps = MICROKERNEL_TEST_DEPS,
5809)
5810
5811xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005812 name = "f32_vrdivc_test",
5813 srcs = [
5814 "test/f32-vrdivc.cc",
5815 "test/vbinaryc-microkernel-tester.h",
5816 ] + MICROKERNEL_TEST_HDRS,
5817 deps = MICROKERNEL_TEST_DEPS,
5818)
5819
5820xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005821 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005822 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005823 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005824 "test/vbinaryc-microkernel-tester.h",
5825 ] + MICROKERNEL_TEST_HDRS,
5826 deps = MICROKERNEL_TEST_DEPS,
5827)
5828
5829xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005830 name = "f32_vrdivc_relu_test",
5831 srcs = [
5832 "test/f32-vrdivc-relu.cc",
5833 "test/vbinaryc-microkernel-tester.h",
5834 ] + MICROKERNEL_TEST_HDRS,
5835 deps = MICROKERNEL_TEST_DEPS,
5836)
5837
5838xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005839 name = "f32_vmax_test",
5840 srcs = [
5841 "test/f32-vmax.cc",
5842 "test/vbinary-microkernel-tester.h",
5843 ] + MICROKERNEL_TEST_HDRS,
5844 deps = MICROKERNEL_TEST_DEPS,
5845)
5846
5847xnnpack_unit_test(
5848 name = "f32_vmaxc_test",
5849 srcs = [
5850 "test/f32-vmaxc.cc",
5851 "test/vbinaryc-microkernel-tester.h",
5852 ] + MICROKERNEL_TEST_HDRS,
5853 deps = MICROKERNEL_TEST_DEPS,
5854)
5855
5856xnnpack_unit_test(
5857 name = "f32_vmin_test",
5858 srcs = [
5859 "test/f32-vmin.cc",
5860 "test/vbinary-microkernel-tester.h",
5861 ] + MICROKERNEL_TEST_HDRS,
5862 deps = MICROKERNEL_TEST_DEPS,
5863)
5864
5865xnnpack_unit_test(
5866 name = "f32_vminc_test",
5867 srcs = [
5868 "test/f32-vminc.cc",
5869 "test/vbinaryc-microkernel-tester.h",
5870 ] + MICROKERNEL_TEST_HDRS,
5871 deps = MICROKERNEL_TEST_DEPS,
5872)
5873
5874xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005875 name = "f32_vmul_test",
5876 srcs = [
5877 "test/f32-vmul.cc",
5878 "test/vbinary-microkernel-tester.h",
5879 ] + MICROKERNEL_TEST_HDRS,
5880 deps = MICROKERNEL_TEST_DEPS,
5881)
5882
5883xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005884 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005885 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005886 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005887 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005888 ] + MICROKERNEL_TEST_HDRS,
5889 deps = MICROKERNEL_TEST_DEPS,
5890)
5891
5892xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005893 name = "f32_vmul_relu_test",
5894 srcs = [
5895 "test/f32-vmul-relu.cc",
5896 "test/vbinary-microkernel-tester.h",
5897 ] + MICROKERNEL_TEST_HDRS,
5898 deps = MICROKERNEL_TEST_DEPS,
5899)
5900
5901xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005902 name = "f32_vmulc_test",
5903 srcs = [
5904 "test/f32-vmulc.cc",
5905 "test/vbinaryc-microkernel-tester.h",
5906 ] + MICROKERNEL_TEST_HDRS,
5907 deps = MICROKERNEL_TEST_DEPS,
5908)
5909
5910xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005911 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005912 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005913 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005914 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005915 ] + MICROKERNEL_TEST_HDRS,
5916 deps = MICROKERNEL_TEST_DEPS,
5917)
5918
5919xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005920 name = "f32_vmulc_relu_test",
5921 srcs = [
5922 "test/f32-vmulc-relu.cc",
5923 "test/vbinaryc-microkernel-tester.h",
5924 ] + MICROKERNEL_TEST_HDRS,
5925 deps = MICROKERNEL_TEST_DEPS,
5926)
5927
5928xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005929 name = "f16_vmulcaddc_minmax_test",
5930 srcs = [
5931 "test/f16-vmulcaddc-minmax.cc",
5932 "test/vmulcaddc-microkernel-tester.h",
5933 "src/xnnpack/AlignedAllocator.h",
5934 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005935 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005936)
5937
5938xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005939 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005940 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005941 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005942 "test/vmulcaddc-microkernel-tester.h",
5943 "src/xnnpack/AlignedAllocator.h",
5944 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005945 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005946)
5947
5948xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005949 name = "f32_vlrelu_test",
5950 srcs = [
5951 "test/f32-vlrelu.cc",
5952 "test/vunary-microkernel-tester.h",
5953 ] + MICROKERNEL_TEST_HDRS,
5954 deps = MICROKERNEL_TEST_DEPS,
5955)
5956
5957xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005958 name = "f32_vneg_test",
5959 srcs = [
5960 "test/f32-vneg.cc",
5961 "test/vunary-microkernel-tester.h",
5962 ] + MICROKERNEL_TEST_HDRS,
5963 deps = MICROKERNEL_TEST_DEPS,
5964)
5965
5966xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005967 name = "f32_vrndne_test",
5968 srcs = [
5969 "test/f32-vrndne.cc",
5970 "test/vunary-microkernel-tester.h",
5971 ] + MICROKERNEL_TEST_HDRS,
5972 deps = MICROKERNEL_TEST_DEPS,
5973)
5974
5975xnnpack_unit_test(
5976 name = "f32_vrndz_test",
5977 srcs = [
5978 "test/f32-vrndz.cc",
5979 "test/vunary-microkernel-tester.h",
5980 ] + MICROKERNEL_TEST_HDRS,
5981 deps = MICROKERNEL_TEST_DEPS,
5982)
5983
5984xnnpack_unit_test(
5985 name = "f32_vrndu_test",
5986 srcs = [
5987 "test/f32-vrndu.cc",
5988 "test/vunary-microkernel-tester.h",
5989 ] + MICROKERNEL_TEST_HDRS,
5990 deps = MICROKERNEL_TEST_DEPS,
5991)
5992
5993xnnpack_unit_test(
5994 name = "f32_vrndd_test",
5995 srcs = [
5996 "test/f32-vrndd.cc",
5997 "test/vunary-microkernel-tester.h",
5998 ] + MICROKERNEL_TEST_HDRS,
5999 deps = MICROKERNEL_TEST_DEPS,
6000)
6001
6002xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07006003 name = "f32_vscale_test",
6004 srcs = [
6005 "test/f32-vscale.cc",
6006 "test/vscale-microkernel-tester.h",
6007 ] + MICROKERNEL_TEST_HDRS,
6008 deps = MICROKERNEL_TEST_DEPS,
6009)
6010
6011xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006012 name = "f32_vscaleexpminusmax_test",
6013 srcs = [
6014 "test/f32-vscaleexpminusmax.cc",
6015 "test/vscaleexpminusmax-microkernel-tester.h",
6016 ] + MICROKERNEL_TEST_HDRS,
6017 deps = MICROKERNEL_TEST_DEPS,
6018)
6019
6020xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006021 name = "f32_vscaleextexp_test",
6022 srcs = [
6023 "test/f32-vscaleextexp.cc",
6024 "test/vscaleextexp-microkernel-tester.h",
6025 ] + MICROKERNEL_TEST_HDRS,
6026 deps = MICROKERNEL_TEST_DEPS,
6027)
6028
6029xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006030 name = "f32_vsqr_test",
6031 srcs = [
6032 "test/f32-vsqr.cc",
6033 "test/vunary-microkernel-tester.h",
6034 ] + MICROKERNEL_TEST_HDRS,
6035 deps = MICROKERNEL_TEST_DEPS,
6036)
6037
6038xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006039 name = "f32_vsqrdiff_test",
6040 srcs = [
6041 "test/f32-vsqrdiff.cc",
6042 "test/vbinary-microkernel-tester.h",
6043 ] + MICROKERNEL_TEST_HDRS,
6044 deps = MICROKERNEL_TEST_DEPS,
6045)
6046
6047xnnpack_unit_test(
6048 name = "f32_vsqrdiffc_test",
6049 srcs = [
6050 "test/f32-vsqrdiffc.cc",
6051 "test/vbinaryc-microkernel-tester.h",
6052 ] + MICROKERNEL_TEST_HDRS,
6053 deps = MICROKERNEL_TEST_DEPS,
6054)
6055
6056xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006057 name = "f32_vsqrt_test",
6058 srcs = [
6059 "test/f32-vsqrt.cc",
6060 "test/vunary-microkernel-tester.h",
6061 ] + MICROKERNEL_TEST_HDRS,
6062 deps = MICROKERNEL_TEST_DEPS,
6063)
6064
6065xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006066 name = "f32_vsub_test",
6067 srcs = [
6068 "test/f32-vsub.cc",
6069 "test/vbinary-microkernel-tester.h",
6070 ] + MICROKERNEL_TEST_HDRS,
6071 deps = MICROKERNEL_TEST_DEPS,
6072)
6073
6074xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006075 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006076 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006077 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006078 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006079 ] + MICROKERNEL_TEST_HDRS,
6080 deps = MICROKERNEL_TEST_DEPS,
6081)
6082
6083xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006084 name = "f32_vsub_relu_test",
6085 srcs = [
6086 "test/f32-vsub-relu.cc",
6087 "test/vbinary-microkernel-tester.h",
6088 ] + MICROKERNEL_TEST_HDRS,
6089 deps = MICROKERNEL_TEST_DEPS,
6090)
6091
6092xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006093 name = "f32_vsubc_test",
6094 srcs = [
6095 "test/f32-vsubc.cc",
6096 "test/vbinaryc-microkernel-tester.h",
6097 ] + MICROKERNEL_TEST_HDRS,
6098 deps = MICROKERNEL_TEST_DEPS,
6099)
6100
6101xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006102 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006103 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006104 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006105 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006106 ] + MICROKERNEL_TEST_HDRS,
6107 deps = MICROKERNEL_TEST_DEPS,
6108)
6109
6110xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006111 name = "f32_vsubc_relu_test",
6112 srcs = [
6113 "test/f32-vsubc-relu.cc",
6114 "test/vbinaryc-microkernel-tester.h",
6115 ] + MICROKERNEL_TEST_HDRS,
6116 deps = MICROKERNEL_TEST_DEPS,
6117)
6118
6119xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006120 name = "f32_vrsubc_test",
6121 srcs = [
6122 "test/f32-vrsubc.cc",
6123 "test/vbinaryc-microkernel-tester.h",
6124 ] + MICROKERNEL_TEST_HDRS,
6125 deps = MICROKERNEL_TEST_DEPS,
6126)
6127
6128xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006129 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006130 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006131 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006132 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006133 ] + MICROKERNEL_TEST_HDRS,
6134 deps = MICROKERNEL_TEST_DEPS,
6135)
6136
6137xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006138 name = "f32_vrsubc_relu_test",
6139 srcs = [
6140 "test/f32-vrsubc-relu.cc",
6141 "test/vbinaryc-microkernel-tester.h",
6142 ] + MICROKERNEL_TEST_HDRS,
6143 deps = MICROKERNEL_TEST_DEPS,
6144)
6145
6146xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006147 name = "qs8_dwconv_minmax_test",
6148 srcs = [
6149 "test/qs8-dwconv-minmax.cc",
6150 "test/dwconv-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 Dukhan4ed53f42020-08-06 01:12:55 -07006157 name = "qs8_gavgpool_minmax_test",
6158 srcs = [
6159 "test/qs8-gavgpool-minmax.cc",
6160 "test/gavgpool-microkernel-tester.h",
6161 "src/xnnpack/AlignedAllocator.h",
6162 ] + MICROKERNEL_TEST_HDRS,
6163 deps = MICROKERNEL_TEST_DEPS,
6164)
6165
6166xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006167 name = "qs8_gemm_minmax_test",
6168 srcs = [
6169 "test/qs8-gemm-minmax.cc",
6170 "test/gemm-microkernel-tester.h",
6171 "src/xnnpack/AlignedAllocator.h",
6172 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6173 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6174)
6175
6176xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006177 name = "qs8_igemm_minmax_test",
6178 srcs = [
6179 "test/qs8-igemm-minmax.cc",
6180 "test/gemm-microkernel-tester.h",
6181 "src/xnnpack/AlignedAllocator.h",
6182 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6183 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6184)
6185
6186xnnpack_unit_test(
6187 name = "qs8_requantization_test",
6188 srcs = [
6189 "src/xnnpack/requantization-stubs.h",
6190 "test/qs8-requantization.cc",
6191 "test/requantization-tester.h",
6192 ] + MICROKERNEL_TEST_HDRS,
6193 deps = MICROKERNEL_TEST_DEPS,
6194)
6195
6196xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006197 name = "qs8_vadd_minmax_test",
6198 srcs = [
6199 "test/qs8-vadd-minmax.cc",
6200 "test/vadd-microkernel-tester.h",
6201 ] + MICROKERNEL_TEST_HDRS,
6202 deps = MICROKERNEL_TEST_DEPS,
6203)
6204
6205xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006206 name = "qs8_vaddc_minmax_test",
6207 srcs = [
6208 "test/qs8-vaddc-minmax.cc",
6209 "test/vaddc-microkernel-tester.h",
6210 ] + MICROKERNEL_TEST_HDRS,
6211 deps = MICROKERNEL_TEST_DEPS,
6212)
6213
6214xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006215 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006216 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006217 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006218 "test/avgpool-microkernel-tester.h",
6219 "src/xnnpack/AlignedAllocator.h",
6220 ] + MICROKERNEL_TEST_HDRS,
6221 deps = MICROKERNEL_TEST_DEPS,
6222)
6223
6224xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006225 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006226 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006227 "test/qu8-dwconv-minmax.cc",
6228 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006229 "src/xnnpack/AlignedAllocator.h",
6230 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006231 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006232)
6233
6234xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006235 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006236 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006237 "test/qu8-igemm-minmax.cc",
6238 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006239 "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 Dukhan08b7a972020-07-14 18:17:29 -07006245 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006246 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006247 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006248 "test/gavgpool-microkernel-tester.h",
6249 "src/xnnpack/AlignedAllocator.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_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006256 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006257 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006258 "test/gemm-microkernel-tester.h",
6259 "src/xnnpack/AlignedAllocator.h",
6260 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006261 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006262)
6263
6264xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006265 name = "qu8_requantization_test",
6266 srcs = [
6267 "src/xnnpack/requantization-stubs.h",
6268 "test/qu8-requantization.cc",
6269 "test/requantization-tester.h",
6270 ] + MICROKERNEL_TEST_HDRS,
6271 deps = MICROKERNEL_TEST_DEPS,
6272)
6273
6274xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006275 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006276 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006277 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006278 "test/vadd-microkernel-tester.h",
6279 ] + MICROKERNEL_TEST_HDRS,
6280 deps = MICROKERNEL_TEST_DEPS,
6281)
6282
6283xnnpack_unit_test(
6284 name = "u8_clamp_test",
6285 srcs = [
6286 "test/u8-clamp.cc",
6287 "test/clamp-microkernel-tester.h",
6288 ] + MICROKERNEL_TEST_HDRS,
6289 deps = MICROKERNEL_TEST_DEPS,
6290)
6291
6292xnnpack_unit_test(
6293 name = "u8_lut32norm_test",
6294 srcs = [
6295 "test/u8-lut32norm.cc",
6296 "test/lut-norm-microkernel-tester.h",
6297 ] + MICROKERNEL_TEST_HDRS,
6298 deps = MICROKERNEL_TEST_DEPS,
6299)
6300
6301xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006302 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006303 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006304 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006305 "test/maxpool-microkernel-tester.h",
6306 ] + MICROKERNEL_TEST_HDRS,
6307 deps = MICROKERNEL_TEST_DEPS,
6308)
6309
6310xnnpack_unit_test(
6311 name = "u8_rmax_test",
6312 srcs = [
6313 "test/u8-rmax.cc",
6314 "test/rmax-microkernel-tester.h",
6315 ] + MICROKERNEL_TEST_HDRS,
6316 deps = MICROKERNEL_TEST_DEPS,
6317)
6318
6319xnnpack_unit_test(
Yury Kartynnike7841862020-11-04 18:22:18 -08006320 name = "x32_depth_to_space_chw2hwc_test",
6321 srcs = [
6322 "test/x32-depth-to-space-chw2hwc.cc",
6323 "test/depth-to-space-microkernel-tester.h",
6324 ] + MICROKERNEL_TEST_HDRS,
6325 deps = MICROKERNEL_TEST_DEPS,
6326)
6327
6328xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006329 name = "x32_fill_test",
6330 srcs = [
6331 "test/x32-fill.cc",
6332 "test/fill-microkernel-tester.h",
6333 ] + MICROKERNEL_TEST_HDRS,
6334 deps = MICROKERNEL_TEST_DEPS,
6335)
6336
6337xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006338 name = "x32_packx_test",
6339 srcs = [
6340 "test/x32-packx.cc",
6341 "test/pack-microkernel-tester.h",
6342 "src/xnnpack/AlignedAllocator.h",
6343 ] + MICROKERNEL_TEST_HDRS,
6344 deps = MICROKERNEL_TEST_DEPS,
6345)
6346
6347xnnpack_unit_test(
6348 name = "x32_pad_test",
6349 srcs = [
6350 "test/x32-pad.cc",
6351 "test/pad-microkernel-tester.h",
6352 ] + MICROKERNEL_TEST_HDRS,
6353 deps = MICROKERNEL_TEST_DEPS,
6354)
6355
6356xnnpack_unit_test(
6357 name = "x32_unpool_test",
6358 srcs = [
6359 "test/x32-unpool.cc",
6360 "test/unpool-microkernel-tester.h",
6361 ] + MICROKERNEL_TEST_HDRS,
6362 deps = MICROKERNEL_TEST_DEPS,
6363)
6364
6365xnnpack_unit_test(
6366 name = "x32_zip_test",
6367 srcs = [
6368 "test/x32-zip.cc",
6369 "test/zip-microkernel-tester.h",
6370 ] + MICROKERNEL_TEST_HDRS,
6371 deps = MICROKERNEL_TEST_DEPS,
6372)
6373
6374xnnpack_unit_test(
6375 name = "x8_lut_test",
6376 srcs = [
6377 "test/x8-lut.cc",
6378 "test/lut-microkernel-tester.h",
6379 ] + MICROKERNEL_TEST_HDRS,
6380 deps = MICROKERNEL_TEST_DEPS,
6381)
6382
6383xnnpack_unit_test(
6384 name = "x8_zip_test",
6385 srcs = [
6386 "test/x8-zip.cc",
6387 "test/zip-microkernel-tester.h",
6388 ] + MICROKERNEL_TEST_HDRS,
6389 deps = MICROKERNEL_TEST_DEPS,
6390)
6391
Marat Dukhan20c3b922020-03-10 03:45:06 -07006392########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006393
6394xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006395 name = "operator_size_test",
6396 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006397 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006398)
6399
Marat Dukhan20c3b922020-03-10 03:45:06 -07006400xnnpack_binary(
6401 name = "subgraph_size_test",
6402 srcs = ["test/subgraph-size.c"],
6403 deps = [":XNNPACK"],
6404)
6405
6406########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006407
6408xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006409 name = "abs_nc_test",
6410 srcs = [
6411 "test/abs-nc.cc",
6412 "test/abs-operator-tester.h",
6413 ],
6414 deps = OPERATOR_TEST_DEPS,
6415)
6416
6417xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006418 name = "add_nd_test",
6419 srcs = [
6420 "test/add-nd.cc",
6421 "test/binary-elementwise-operator-tester.h",
6422 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006423 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006424)
6425
6426xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006427 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006428 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006429 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006430 "test/argmax-pooling-operator-tester.h",
6431 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006432 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006433)
6434
6435xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006436 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006437 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006438 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006439 "test/average-pooling-operator-tester.h",
6440 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006441 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006442)
6443
6444xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006445 name = "bankers_rounding_nc_test",
6446 srcs = [
6447 "test/bankers-rounding-nc.cc",
6448 "test/bankers-rounding-operator-tester.h",
6449 ],
6450 deps = OPERATOR_TEST_DEPS,
6451)
6452
6453xnnpack_unit_test(
6454 name = "ceiling_nc_test",
6455 srcs = [
6456 "test/ceiling-nc.cc",
6457 "test/ceiling-operator-tester.h",
6458 ],
6459 deps = OPERATOR_TEST_DEPS,
6460)
6461
6462xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006463 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006464 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006465 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006466 "test/channel-shuffle-operator-tester.h",
6467 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006468 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006469)
6470
6471xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006472 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006473 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006474 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006475 "test/clamp-operator-tester.h",
6476 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006477 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006478)
6479
6480xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006481 name = "constant_pad_nd_test",
6482 srcs = [
6483 "test/constant-pad-nd.cc",
6484 "test/constant-pad-operator-tester.h",
6485 ],
6486 deps = OPERATOR_TEST_DEPS,
6487)
6488
6489xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006490 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006491 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006492 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006493 "test/convolution-operator-tester.h",
6494 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006495 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006496)
6497
6498xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006499 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006500 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006501 "test/convolution-nchw.cc",
6502 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006503 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006504 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006505)
6506
6507xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006508 name = "copy_nc_test",
6509 srcs = [
6510 "test/copy-nc.cc",
6511 "test/copy-operator-tester.h",
6512 ],
6513 deps = OPERATOR_TEST_DEPS,
6514)
6515
6516xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006517 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006518 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006519 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006520 "test/deconvolution-operator-tester.h",
6521 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006522 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006523)
6524
6525xnnpack_unit_test(
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08006526 name = "depth_to_space_test",
6527 srcs = [
6528 "test/depth-to-space.cc",
6529 "test/depth-to-space-operator-tester.h",
6530 ] + OPERATOR_TEST_PARAMS_HDRS,
6531 deps = OPERATOR_TEST_DEPS,
6532)
6533
6534xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006535 name = "divide_nd_test",
6536 srcs = [
6537 "test/binary-elementwise-operator-tester.h",
6538 "test/divide-nd.cc",
6539 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006540 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006541)
6542
6543xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006544 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006545 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006546 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006547 "test/fully-connected-operator-tester.h",
6548 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006549 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006550)
6551
6552xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006553 name = "floor_nc_test",
6554 srcs = [
6555 "test/floor-nc.cc",
6556 "test/floor-operator-tester.h",
6557 ],
6558 deps = OPERATOR_TEST_DEPS,
6559)
6560
6561xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006562 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006563 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006564 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006565 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006566 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006567 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006568)
6569
6570xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006571 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006572 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006573 "test/global-average-pooling-ncw.cc",
6574 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006575 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006576 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006577)
6578
6579xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006580 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006581 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006582 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006583 "test/hardswish-operator-tester.h",
6584 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006585 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006586)
6587
6588xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006589 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006590 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006591 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006592 "test/leaky-relu-operator-tester.h",
6593 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006594 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006595)
6596
6597xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006598 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006599 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006600 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006601 "test/max-pooling-operator-tester.h",
6602 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006603 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006604)
6605
6606xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006607 name = "maximum_nd_test",
6608 srcs = [
6609 "test/binary-elementwise-operator-tester.h",
6610 "test/maximum-nd.cc",
6611 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006612 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006613)
6614
6615xnnpack_unit_test(
6616 name = "minimum_nd_test",
6617 srcs = [
6618 "test/binary-elementwise-operator-tester.h",
6619 "test/minimum-nd.cc",
6620 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006621 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006622)
6623
6624xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006625 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006626 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006627 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006628 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006629 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006630 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006631)
6632
6633xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006634 name = "negate_nc_test",
6635 srcs = [
6636 "test/negate-nc.cc",
6637 "test/negate-operator-tester.h",
6638 ],
6639 deps = OPERATOR_TEST_DEPS,
6640)
6641
6642xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006643 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006644 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006645 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006646 "test/prelu-operator-tester.h",
6647 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006648 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006649)
6650
6651xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006652 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006653 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006654 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006655 "test/resize-bilinear-operator-tester.h",
6656 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006657 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006658)
6659
6660xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07006661 name = "resize_bilinear_nchw_test",
6662 srcs = [
6663 "test/resize-bilinear-nchw.cc",
6664 "test/resize-bilinear-operator-tester.h",
6665 ] + OPERATOR_TEST_PARAMS_HDRS,
6666 deps = OPERATOR_TEST_DEPS,
6667)
6668
6669xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006670 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006671 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006672 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006673 "test/sigmoid-operator-tester.h",
6674 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006675 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006676)
6677
6678xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006679 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006680 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006681 "test/softmax-nc.cc",
6682 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006683 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006684 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006685)
6686
6687xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006688 name = "square_nc_test",
6689 srcs = [
6690 "test/square-nc.cc",
6691 "test/square-operator-tester.h",
6692 ],
6693 deps = OPERATOR_TEST_DEPS,
6694)
6695
6696xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006697 name = "square_root_nc_test",
6698 srcs = [
6699 "test/square-root-nc.cc",
6700 "test/square-root-operator-tester.h",
6701 ],
6702 deps = OPERATOR_TEST_DEPS,
6703)
6704
6705xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006706 name = "squared_difference_nd_test",
6707 srcs = [
6708 "test/binary-elementwise-operator-tester.h",
6709 "test/squared-difference-nd.cc",
6710 ],
6711 deps = OPERATOR_TEST_DEPS,
6712)
6713
6714xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006715 name = "subtract_nd_test",
6716 srcs = [
6717 "test/binary-elementwise-operator-tester.h",
6718 "test/subtract-nd.cc",
6719 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006720 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006721)
6722
6723xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006724 name = "truncation_nc_test",
6725 srcs = [
6726 "test/truncation-nc.cc",
6727 "test/truncation-operator-tester.h",
6728 ],
6729 deps = OPERATOR_TEST_DEPS,
6730)
6731
6732xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006733 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006734 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006735 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006736 "test/unpooling-operator-tester.h",
6737 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006738 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006739)
6740
Chao Mei6ddfc602020-05-13 22:29:36 -07006741############################### Misc unit tests ###############################
6742
6743xnnpack_unit_test(
6744 name = "memory_planner_test",
6745 srcs = [
6746 "test/memory-planner-test.cc",
6747 ],
6748 deps = [
6749 ":XNNPACK",
6750 ":memory_planner",
6751 ],
6752)
6753
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07006754xnnpack_unit_test(
6755 name = "subgraph_nchw_test",
6756 srcs = [
6757 "src/xnnpack/subgraph.h",
6758 "test/subgraph-nchw.cc",
6759 "test/subgraph-tester.h",
6760 ],
6761 deps = [
6762 ":XNNPACK",
6763 ],
6764)
6765
Marat Dukhan08c4a432019-10-03 09:29:21 -07006766############################# Build configurations #############################
6767
Marat Dukhanb8642352019-10-30 15:43:02 -07006768# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006769config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006770 name = "xnn_enable_assembly_explicit_true",
6771 define_values = {"xnn_enable_assembly": "true"},
6772)
6773
6774# Disables usage of assembly kernels.
6775config_setting(
6776 name = "xnn_enable_assembly_explicit_false",
6777 define_values = {"xnn_enable_assembly": "false"},
6778)
6779
Marat Dukhan9de90e02020-06-18 16:04:12 -07006780# Enables usage of sparse inference.
6781config_setting(
6782 name = "xnn_enable_sparse_explicit_true",
6783 define_values = {"xnn_enable_sparse": "true"},
6784)
6785
6786# Disables usage of sparse inference.
6787config_setting(
6788 name = "xnn_enable_sparse_explicit_false",
6789 define_values = {"xnn_enable_sparse": "false"},
6790)
6791
Marat Dukhan05702cf2020-03-26 15:41:33 -07006792# Disables usage of HMP-aware optimizations.
6793config_setting(
6794 name = "xnn_enable_hmp_explicit_false",
6795 define_values = {"xnn_enable_hmp": "false"},
6796)
6797
Chao Mei6ddfc602020-05-13 22:29:36 -07006798# Enable usage of optimized memory allocation
6799config_setting(
6800 name = "xnn_enable_memopt_explicit_true",
6801 define_values = {"xnnpack_enable_memopt": "true"},
6802)
6803
6804# Disable usage of optimized memory allocation
6805config_setting(
6806 name = "xnn_enable_memopt_explicit_false",
6807 define_values = {"xnnpack_enable_memopt": "false"},
6808)
6809
Marat Dukhanb8642352019-10-30 15:43:02 -07006810# Builds with -c dbg
6811config_setting(
6812 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006813 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006814 "compilation_mode": "dbg",
6815 },
6816)
6817
6818# Builds with -c opt
6819config_setting(
6820 name = "optimized_build",
6821 values = {
6822 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006823 },
6824)
6825
6826config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006827 name = "linux_k8",
6828 values = {"cpu": "k8"},
6829)
6830
6831config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006832 name = "linux_arm",
6833 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006834)
6835
6836config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006837 name = "linux_armeabi",
6838 values = {"cpu": "armeabi"},
6839)
6840
6841config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006842 name = "linux_armhf",
6843 values = {"cpu": "armhf"},
6844)
6845
6846config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006847 name = "linux_armv7a",
6848 values = {"cpu": "armv7a"},
6849)
6850
6851config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006852 name = "linux_aarch64",
6853 values = {"cpu": "aarch64"},
6854)
6855
6856config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006857 name = "android",
6858 values = {"crosstool_top": "//external:android/crosstool"},
6859)
6860
6861config_setting(
6862 name = "android_armv7",
6863 values = {
6864 "crosstool_top": "//external:android/crosstool",
6865 "cpu": "armeabi-v7a",
6866 },
6867)
6868
6869config_setting(
6870 name = "android_arm64",
6871 values = {
6872 "crosstool_top": "//external:android/crosstool",
6873 "cpu": "arm64-v8a",
6874 },
6875)
6876
6877config_setting(
6878 name = "android_x86",
6879 values = {
6880 "crosstool_top": "//external:android/crosstool",
6881 "cpu": "x86",
6882 },
6883)
6884
6885config_setting(
6886 name = "android_x86_64",
6887 values = {
6888 "crosstool_top": "//external:android/crosstool",
6889 "cpu": "x86_64",
6890 },
6891)
6892
6893config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006894 name = "windows_x86_64",
6895 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006896)
6897
6898config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006899 name = "windows_x86_64_clang",
6900 values = {
6901 "compiler": "clang-cl",
6902 "cpu": "x64_windows",
6903 },
6904)
6905
6906config_setting(
6907 name = "windows_x86_64_mingw",
6908 values = {
6909 "compiler": "mingw-gcc",
6910 "cpu": "x64_windows",
6911 },
6912)
6913
6914config_setting(
6915 name = "windows_x86_64_msys",
6916 values = {
6917 "compiler": "msys-gcc",
6918 "cpu": "x64_windows",
6919 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006920)
6921
6922config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006923 name = "macos_x86_64",
6924 values = {
6925 "apple_platform_type": "macos",
6926 "cpu": "darwin",
6927 },
6928)
6929
6930config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006931 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006932 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006933)
6934
6935config_setting(
6936 name = "emscripten_wasm",
6937 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006938 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006939 "cpu": "wasm",
6940 },
6941)
6942
6943config_setting(
6944 name = "emscripten_wasmsimd",
6945 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006946 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006947 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006948 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006949 },
6950)
6951
6952config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006953 name = "ios_armv7",
6954 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006955 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006956 "cpu": "ios_armv7",
6957 },
6958)
6959
6960config_setting(
6961 name = "ios_arm64",
6962 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006963 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006964 "cpu": "ios_arm64",
6965 },
6966)
6967
6968config_setting(
6969 name = "ios_arm64e",
6970 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006971 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006972 "cpu": "ios_arm64e",
6973 },
6974)
6975
6976config_setting(
6977 name = "ios_x86",
6978 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006979 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006980 "cpu": "ios_i386",
6981 },
6982)
6983
6984config_setting(
6985 name = "ios_x86_64",
6986 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006987 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006988 "cpu": "ios_x86_64",
6989 },
6990)
6991
6992config_setting(
6993 name = "watchos_armv7k",
6994 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006995 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006996 "cpu": "watchos_armv7k",
6997 },
6998)
6999
7000config_setting(
7001 name = "watchos_arm64_32",
7002 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007003 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007004 "cpu": "watchos_arm64_32",
7005 },
7006)
7007
7008config_setting(
7009 name = "watchos_x86",
7010 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007011 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007012 "cpu": "watchos_i386",
7013 },
7014)
7015
7016config_setting(
7017 name = "watchos_x86_64",
7018 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007019 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007020 "cpu": "watchos_x86_64",
7021 },
7022)
7023
7024config_setting(
7025 name = "tvos_arm64",
7026 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007027 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007028 "cpu": "tvos_arm64",
7029 },
7030)
7031
7032config_setting(
7033 name = "tvos_x86_64",
7034 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007035 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007036 "cpu": "tvos_x86_64",
7037 },
7038)