blob: ba77f82540df64bd84b304f8e05c834ac6c04427 [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",
Artsiom Ablavatskibbe85062020-11-05 14:07:37 -080087 "src/subgraph/depth-to-space.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -080088 "src/subgraph/depthwise-convolution-2d.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",
Marat Dukhanad71b9a2020-11-20 00:01:51 -0800503 "src/x32-depthtospace2d-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 Barcharddb5c32d2020-11-16 23:58:42 -0800735 "src/f32-dwconv2d-chw/gen/3x3p1-wasmsimd-1x4-acc3-arm.c",
Frank Barcharddb5c32d2020-11-16 23:58:42 -0800736 "src/f32-dwconv2d-chw/gen/3x3p1-wasmsimd-1x4-acc3-x86.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -0800737 "src/f32-dwconv2d-chw/gen/3x3s2p1-wasmsimd-1x4-acc3-arm.c",
Frank Barcharddb5c32d2020-11-16 23:58:42 -0800738 "src/f32-dwconv2d-chw/gen/3x3s2p1-wasmsimd-1x4-acc3-x86.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -0800739 "src/f32-dwconv2d-chw/gen/5x5p2-wasmsimd-3x4-arm.c",
Frank Barcharddb5c32d2020-11-16 23:58:42 -0800740 "src/f32-dwconv2d-chw/gen/5x5p2-wasmsimd-3x4-x86.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -0800741 "src/f32-dwconv2d-chw/gen/5x5s2p2-wasmsimd-1x4-acc2-arm.c",
Frank Barcharddb5c32d2020-11-16 23:58:42 -0800742 "src/f32-dwconv2d-chw/gen/5x5s2p2-wasmsimd-1x4-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700743 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
744 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700745 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
746 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
747 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
748 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700749 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
750 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
751 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
752 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
753 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
754 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
755 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
756 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
757 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
758 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
759 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
760 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
761 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
762 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
763 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
764 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
765 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
766 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
767 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
768 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
769 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
770 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
771 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
772 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
773 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
774 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
775 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
776 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
777 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
778 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
779 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
780 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
781 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
782 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
783 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
784 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
785 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
786 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
787 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
788 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
789 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
790 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
791 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
792 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
793 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
794 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
795 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
796 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
797 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
798 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
799 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
800 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
801 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
802 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
803 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
804 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
805 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
806 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
807 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
808 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
809 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
810 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
811 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
812 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
813 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
814 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
815 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
816 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
817 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
818 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700819 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
820 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
821 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
XNNPACK Team965272b2020-10-23 21:10:15 -0700822 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
823 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700824 "src/f32-ibilinear/gen/wasmsimd-c4.c",
825 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700826 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
827 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
828 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
829 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
830 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
831 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
832 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
833 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
834 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
835 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
836 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
837 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
838 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
839 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
840 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
841 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
842 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
843 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
844 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
845 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
846 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
847 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
848 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
849 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
850 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
851 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
852 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
853 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
854 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
855 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
856 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
857 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
858 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
859 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
860 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
861 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
862 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
863 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
864 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
865 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700866 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
867 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700868 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
869 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
870 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
871 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700872 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
873 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700874 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
875 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
876 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700877 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
878 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700879 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
880 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
881 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
882 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
883 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
884 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
885 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700886 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
887 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700888 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
889 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
890 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
891 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700892 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700893 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700894 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700895 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
896 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700897 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700898 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
899 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700900 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700901 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
902 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700903 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700904 "src/f32-relu/gen/wasmsimd-x4.c",
905 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700906 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700907 "src/f32-rmax/wasmsimd-arm.c",
908 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700909 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
910 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
911 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
912 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
913 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
914 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700915 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
916 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
917 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
918 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
919 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
920 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800921 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined-x2.c",
922 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700923 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c",
924 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700925 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800926 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined-x2.c",
927 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700928 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c",
929 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700930 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800931 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined-x2.c",
932 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700933 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c",
934 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700935 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800936 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined-x2.c",
937 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700938 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c",
939 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700940 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800941 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined-x2.c",
942 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700943 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c",
944 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700945 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800946 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined-x2.c",
947 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700948 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c",
949 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700950 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800951 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined-x2.c",
952 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700953 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c",
954 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -0700955 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800956 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined-x2.c",
957 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700958 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c",
959 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -0700960 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700961 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
962 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700963 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700964 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
965 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700966 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700967 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
968 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700969 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700970 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
971 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700972 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700973 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
974 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700975 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700976 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
977 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700978 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700979 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
980 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700981 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700982 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
983 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700984 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700985 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
986 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700987 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700988 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
989 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700990 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700991 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
992 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700993 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700994 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
995 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700996 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700997 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
998 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700999 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001000 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
1001 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001002 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001003 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
1004 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001005 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001006 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
1007 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001008 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001009 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
1010 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001011 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001012 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
1013 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001014 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001015 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
1016 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001017 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001018 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
1019 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001020 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001021 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
1022 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001023 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001024 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
1025 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001026 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001027 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
1028 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001029 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001030 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
1031 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001032 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001033 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
1034 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001035 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001036 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
1037 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001038 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001039 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
1040 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001041 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001042 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
1043 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001044 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001045 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
1046 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001047 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001048 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
1049 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001050 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001051 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
1052 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001053 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001054 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
1055 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001056 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001057 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
1058 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001059 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001060 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
1061 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001062 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001063 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
1064 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001065 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001066 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
1067 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001068 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001069 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1070 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001071 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001072 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1073 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001074 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001075 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1076 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001077 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001078 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1079 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001080 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001081 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1082 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001083 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001084 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1085 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001086 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001087 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1088 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001089 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001090 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1091 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001092 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001093 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1094 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001095 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001096 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1097 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001098 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001099 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1100 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001101 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001102 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1103 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001104 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001105 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1106 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001107 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001108 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1109 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001110 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001111 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1112 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1113 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1114 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001115 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001116 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001117 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001118 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001119 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1120 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1121 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1122 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001123 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1124 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1125 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1126 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1127 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1128 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1129 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1130 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1131 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1132 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001133 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1134 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001135 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1136 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1137 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1138 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1139 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1140 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001141 "src/math/roundd-wasmsimd-addsub.c",
1142 "src/math/roundd-wasmsimd-cvt.c",
1143 "src/math/roundne-wasmsimd-addsub.c",
1144 "src/math/roundu-wasmsimd-addsub.c",
1145 "src/math/roundu-wasmsimd-cvt.c",
1146 "src/math/roundz-wasmsimd-addsub.c",
1147 "src/math/roundz-wasmsimd-cvt.c",
1148 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1149 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001150 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1151 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1152 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001153 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1154 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1155 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001156 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1157 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1158 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001159 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001160 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001161 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001162 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1163 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001164 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001165 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1166 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001167 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1168 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001169 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001170 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001171 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001172 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001173 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001174 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001175 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001176 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1177 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1178 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1179 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1180 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1181 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1182 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1183 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001184 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001185 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001186 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001187 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001188 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001189 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001190 "src/x32-zip/x2-wasmsimd.c",
1191 "src/x32-zip/x3-wasmsimd.c",
1192 "src/x32-zip/x4-wasmsimd.c",
1193 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001194]
1195
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001196PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -07001197 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001198]
1199
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001200PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001201 "src/qs8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001202 "src/qs8-requantization/precise-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001203 "src/qu8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001204 "src/qu8-requantization/precise-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001205]
1206
Marat Dukhan08c4a432019-10-03 09:29:21 -07001207# ISA-specific micro-kernels
1208NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001209 "src/f32-argmaxpool/4x-neon-c4.c",
1210 "src/f32-argmaxpool/9p8x-neon-c4.c",
1211 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001212 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1213 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001214 "src/f32-clamp/gen/neon-x4.c",
1215 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001216 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001217 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001218 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001219 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001220 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001221 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001222 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001223 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001224 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001225 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001226 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001227 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001228 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001229 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1230 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1231 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1232 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1233 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001234 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001235 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001236 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1237 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1238 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001239 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001240 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001241 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1242 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1243 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1244 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1245 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001246 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1247 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1248 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001249 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001250 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001251 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1252 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1253 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001254 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1255 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1256 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1257 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001258 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001259 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1260 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001261 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001262 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001263 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001264 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001265 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1266 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001267 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1268 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1269 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1270 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1271 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1272 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1273 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1274 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001275 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001276 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001277 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001278 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1279 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001280 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001281 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1282 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001283 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001284 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1285 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1286 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1287 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1288 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001289 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1290 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001291 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1292 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001293 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1294 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001295 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1296 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1297 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1298 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1299 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1300 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1301 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1302 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1303 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1304 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1305 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1306 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1307 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1308 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1309 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1310 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001311 "src/f32-hswish/gen/hswish-neon-x4.c",
1312 "src/f32-hswish/gen/hswish-neon-x8.c",
1313 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001314 "src/f32-ibilinear/gen/neon-c4.c",
1315 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001316 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001317 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001318 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001319 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1320 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001321 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001322 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1323 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1324 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1325 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001326 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1327 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001328 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1329 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001330 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1331 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001332 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1333 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1334 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001335 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1336 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001337 "src/f32-prelu/gen/neon-1x4.c",
1338 "src/f32-prelu/gen/neon-1x8.c",
1339 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001340 "src/f32-prelu/gen/neon-2x4.c",
1341 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001342 "src/f32-prelu/gen/neon-2x16.c",
1343 "src/f32-prelu/gen/neon-4x4.c",
1344 "src/f32-prelu/gen/neon-4x8.c",
1345 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001346 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001347 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001348 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001349 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1350 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001351 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001352 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1353 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001354 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001355 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1356 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001357 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1358 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1359 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1360 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1361 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1362 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1363 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1364 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1365 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1366 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1367 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1368 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1369 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001370 "src/f32-relu/gen/neon-x4.c",
1371 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001372 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001373 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001374 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1375 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1376 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1377 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1378 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1379 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1380 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1381 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1382 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1383 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1384 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1385 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001386 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1387 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1388 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1389 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1390 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1391 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001392 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1393 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1394 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1395 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001396 "src/f32-vbinary/gen/vmax-neon-x4.c",
1397 "src/f32-vbinary/gen/vmax-neon-x8.c",
1398 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1399 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1400 "src/f32-vbinary/gen/vmin-neon-x4.c",
1401 "src/f32-vbinary/gen/vmin-neon-x8.c",
1402 "src/f32-vbinary/gen/vminc-neon-x4.c",
1403 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001404 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1405 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1406 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1407 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1408 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1409 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001410 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1411 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1412 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1413 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001414 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1415 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1416 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1417 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001418 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1419 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001420 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1421 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001422 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1423 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001424 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1425 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1426 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1427 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1428 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1429 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001430 "src/f32-vunary/gen/vabs-neon-x4.c",
1431 "src/f32-vunary/gen/vabs-neon-x8.c",
1432 "src/f32-vunary/gen/vneg-neon-x4.c",
1433 "src/f32-vunary/gen/vneg-neon-x8.c",
1434 "src/f32-vunary/gen/vsqr-neon-x4.c",
1435 "src/f32-vunary/gen/vsqr-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001436 "src/math/roundd-neon-addsub.c",
1437 "src/math/roundd-neon-cvt.c",
1438 "src/math/roundne-neon-addsub.c",
1439 "src/math/roundu-neon-addsub.c",
1440 "src/math/roundu-neon-cvt.c",
1441 "src/math/roundz-neon-addsub.c",
1442 "src/math/roundz-neon-cvt.c",
1443 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
1444 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1445 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1446 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1447 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1448 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1449 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1450 "src/math/sqrt-neon-nr1rsqrts.c",
1451 "src/math/sqrt-neon-nr2rsqrts.c",
1452 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001453 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1454 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1455 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1456 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001457 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1458 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1459 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1460 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001461 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1462 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1463 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1464 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001465 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001466 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001467 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001468 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1469 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001470 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001471 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001472 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001473 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001474 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001475 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001476 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1477 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1478 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1479 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1480 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1481 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1482 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1483 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001484 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1485 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1486 "src/qu8-dwconv/up8x9-minmax-neon.c",
1487 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1488 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1489 "src/qu8-gemm/4x8-minmax-neon.c",
1490 "src/qu8-gemm/8x8-minmax-neon.c",
1491 "src/qu8-igemm/4x8-minmax-neon.c",
1492 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001493 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001494 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001495 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001496 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001497 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001498 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001499 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001500 "src/x8-zip/x2-neon.c",
1501 "src/x8-zip/x3-neon.c",
1502 "src/x8-zip/x4-neon.c",
1503 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001504 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001505 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001506 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001507 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001508 "src/x32-zip/x2-neon.c",
1509 "src/x32-zip/x3-neon.c",
1510 "src/x32-zip/x4-neon.c",
1511 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001512]
1513
1514NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001515 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1516 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1517 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1518 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1519 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1520 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1521 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1522 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1523 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1524 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1525 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1526 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1527 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1528 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1529 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1530 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1531 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1532 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1533 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1534 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1535 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1536 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1537 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1538 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1539 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1540 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1541 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1542 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1543 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1544 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001545 "src/f32-ibilinear/gen/neonfma-c4.c",
1546 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001547 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001548 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001549 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001550 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1551 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001552 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1553 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001554 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1555 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001556 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1557 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001558 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001559 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001560 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001561 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1562 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001563 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001564 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1565 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001566 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001567 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1568 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001569 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1570 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1571 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1572 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1573 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1574 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1575 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1576 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1577 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1578 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1579 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1580 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1581 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001582 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1583 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1584 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1585 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1586 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1587 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001588 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1589 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1590 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1591 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1592 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1593 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001594 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1595 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1596 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1597 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1598 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1599 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001600 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1601 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1602 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1603 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1604 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1605 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001606 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1607 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1608 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1609 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1610 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1611 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001612 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1613 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1614 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1615 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1616 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1617 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001618 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1619 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1620 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1621 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1622 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1623 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1624 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1625 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1626 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1627 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1628 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1629 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1630 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1631 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1632 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1633 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1634 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1635 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001636 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1637 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001638 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1639 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1640 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1641 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1642 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1643 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1644 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1645 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1646 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1647 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1648 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1649 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1650 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1651 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1652 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1653 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1654 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1655 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1656 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1657 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001658 "src/math/exp-neonfma-lut64-p2.c",
1659 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001660 "src/math/expminus-neonfma-lut64-p2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001661 "src/math/expminus-neonfma-lut2048-p1.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001662 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001663 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1664 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1665 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001666 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1667 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1668 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001669 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1670 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1671 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001672 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1673 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1674 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001675 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1676 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1677 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001678 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1679 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1680 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001681 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001682 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001683 "src/math/sqrt-neonfma-nr2fma.c",
1684 "src/math/sqrt-neonfma-nr2fma1adj.c",
1685 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001686]
1687
1688AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001689 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001690 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001691 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001692 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001693 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001694 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001695 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001696 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001697 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001698 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
1699 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
1700 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001701 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001702 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001703 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
1704 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
1705 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
1706 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
1707 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001708 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
1709 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
1710 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001711 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001712 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001713 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
1714 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
1715 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001716 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
1717 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
1718 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
1719 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001720 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001721 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
1722 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001723 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001724 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001725 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001726 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001727 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
1728 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001729 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
1730 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
1731 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
1732 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
1733 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
1734 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
1735 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
1736 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001737 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001738 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001739 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1740 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1741 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1742 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1743 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1744 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1745 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1746 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1747 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1748 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1749 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1750 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1751 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1752 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1753 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1754 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1755 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1756 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1757 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1758 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001759 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1760 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1761 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1762 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1763 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1764 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1765 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1766 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1767 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1768 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1769 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1770 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001771 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1772 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1773 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1774 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1775 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1776 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001777 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001778 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001779 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1780 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1781 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1782 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001783 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001784 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1785 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1786 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001787 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1788 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1789 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1790 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001791 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001792 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1793 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1794 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001795 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001796 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001797 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
1798 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
1799 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001800 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1801 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1802 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1803 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1804 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1805 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001806 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1807 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001808 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001809 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001810 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001811 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001812 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001813 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001814]
1815
Marat Dukhan8853b822020-05-07 12:19:01 -07001816NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001817 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1818 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001819 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1820 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1821 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1822 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1823 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1824 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001825 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001826 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001827 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001828 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001829]
1830
Marat Dukhan08c4a432019-10-03 09:29:21 -07001831AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001832 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001833 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001834 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1835 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1836 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1837 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001838 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1839 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1840 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1841 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1842 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1843 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1844 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1845 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001846 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1847 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001848 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1849 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1850 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1851 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1852 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1853 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1854 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
1855 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1856 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1857 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1858 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1859 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1860 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1861 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1862 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1863 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001864 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001865 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1866 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1867 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1868 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1869 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1870 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1871 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1872 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1873 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001874 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001875 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001876 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001877 "src/f16-relu/gen/neonfp16arith-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001878 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001879 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001880 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001881 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001882 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001883 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001884 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001885 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1886 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1887 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1888 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1889 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1890 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1891 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1892 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1893 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1894 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1895 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1896 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1897 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1898 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1899 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1900 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1901 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1902 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1903 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1904 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1905 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1906 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1907 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1908 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1909 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1910 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
1911 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1912 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1913 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1914 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
1915 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001916]
1917
Benoit Jacoba9644732020-08-13 12:48:55 -07001918NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07001919 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001920 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001921 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001922 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001923 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001924 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001925 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001926 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001927 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001928 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001929 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001930 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001931 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001932 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001933 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001934 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001935 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001936 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001937]
1938
Marat Dukhan08c4a432019-10-03 09:29:21 -07001939SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001940 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1941 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001942 "src/f32-clamp/gen/sse-x4.c",
1943 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001944 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1945 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001946 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1947 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1948 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1949 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001950 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1951 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001952 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1953 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1954 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1955 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001956 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1957 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001958 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
1959 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
1960 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07001961 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001962 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07001963 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
1964 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
1965 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
1966 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
1967 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07001968 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
1969 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
1970 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001971 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07001972 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001973 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
1974 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
1975 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07001976 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
1977 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
1978 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
1979 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
1980 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
1981 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
1982 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
1983 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
1984 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
1985 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
1986 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
1987 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
1988 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07001989 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
1990 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
1991 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
1992 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
1993 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
1994 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
1995 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
1996 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07001997 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001998 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001999 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002000 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
2001 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002002 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
2003 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
2004 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
2005 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
2006 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
2007 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002008 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
2009 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
2010 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
2011 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
2012 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
2013 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
2014 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002015 "src/f32-hswish/gen/hswish-sse-x4.c",
2016 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07002017 "src/f32-ibilinear/gen/sse-c4.c",
2018 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002019 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
2020 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
2021 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
2022 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2023 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2024 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2025 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002026 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2027 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2028 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002029 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002030 "src/f32-prelu/gen/sse-2x4.c",
2031 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002032 "src/f32-relu/gen/sse-x4.c",
2033 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002034 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002035 "src/f32-spmm/gen/4x1-minmax-sse.c",
2036 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002037 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002038 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002039 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2040 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2041 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2042 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2043 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2044 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2045 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2046 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002047 "src/f32-vbinary/gen/vmax-sse-x4.c",
2048 "src/f32-vbinary/gen/vmax-sse-x8.c",
2049 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2050 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2051 "src/f32-vbinary/gen/vmin-sse-x4.c",
2052 "src/f32-vbinary/gen/vmin-sse-x8.c",
2053 "src/f32-vbinary/gen/vminc-sse-x4.c",
2054 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002055 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2056 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2057 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2058 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2059 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2060 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2061 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2062 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002063 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2064 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2065 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2066 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002067 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2068 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2069 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2070 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002071 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2072 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002073 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2074 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002075 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2076 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002077 "src/f32-vunary/gen/vabs-sse-x4.c",
2078 "src/f32-vunary/gen/vabs-sse-x8.c",
2079 "src/f32-vunary/gen/vneg-sse-x4.c",
2080 "src/f32-vunary/gen/vneg-sse-x8.c",
2081 "src/f32-vunary/gen/vsqr-sse-x4.c",
2082 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002083 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002084 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002085 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002086 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002087 "src/math/sqrt-sse-hh1mac.c",
2088 "src/math/sqrt-sse-nr1mac.c",
2089 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002090 "src/x32-fill/sse.c",
2091 "src/x32-packx/x4-sse.c",
2092 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002093]
2094
2095SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002096 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002097 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002098 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002099 "src/f32-prelu/gen/sse2-2x4.c",
2100 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002101 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002102 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002103 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002104 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2105 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002106 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002107 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2108 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002109 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002110 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2111 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002112 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002113 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
2114 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
2115 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
2116 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
2117 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
2118 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002119 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002120 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002121 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002122 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002123 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
2124 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002125 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2126 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002127 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2128 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002129 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2130 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2131 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2132 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2133 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2134 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
2135 "src/math/exp-sse2-lut64-p2.c",
2136 "src/math/exp-sse2-p5.c",
2137 "src/math/expminus-sse2-p5.c",
Marat Dukhana438aca2020-11-20 15:45:01 -08002138 "src/math/expm1minus-sse2-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002139 "src/math/roundd-sse2-cvt.c",
2140 "src/math/roundne-sse2-cvt.c",
2141 "src/math/roundu-sse2-cvt.c",
2142 "src/math/roundz-sse2-cvt.c",
2143 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2144 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2145 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2146 "src/math/sigmoid-sse2-rr2-p5-div.c",
2147 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2148 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002149 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2150 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2151 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002152 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2153 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2154 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002155 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2156 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2157 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002158 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002159 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002160 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002161 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002162 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002163 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002164 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2165 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002166 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002167 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2168 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002169 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002170 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2171 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2172 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002173 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002174 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002175 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002176 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002177 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002178 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002179 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002180 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002181 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2182 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002183 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002184 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002185 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002186 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2187 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2188 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2189 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002190 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2191 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2192 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2193 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002194 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2195 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002196 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2197 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2198 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2199 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2200 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002201 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002202 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002203 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002204 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002205 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002206 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002207 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002208 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002209 "src/x8-zip/x2-sse2.c",
2210 "src/x8-zip/x3-sse2.c",
2211 "src/x8-zip/x4-sse2.c",
2212 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002213 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002214 "src/x32-zip/x2-sse2.c",
2215 "src/x32-zip/x3-sse2.c",
2216 "src/x32-zip/x4-sse2.c",
2217 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002218]
2219
2220SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002221 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2222 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2223 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002224 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002225 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002226 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2227 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2228 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2229 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2230 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002231 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2232 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2233 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002234 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2235 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2236 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002237 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2238 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2239 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002240 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002241 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002242 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002243 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002244 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002245 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002246 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2247 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002248 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002249 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2250 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002251 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002252 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2253 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2254 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002255 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002256 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002257 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002258 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002259 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002260 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002261 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002262 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002263 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2264 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002265 "src/qs8-requantization/precise-ssse3.c",
2266 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002267 "src/qu8-requantization/precise-ssse3.c",
2268 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002269]
2270
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002271SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002272 "src/f32-prelu/gen/sse41-2x4.c",
2273 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002274 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2275 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2276 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2277 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2278 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2279 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002280 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002281 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002282 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002283 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002284 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2285 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002286 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2287 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002288 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2289 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002290 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2291 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2292 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2293 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2294 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2295 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2296 "src/math/roundd-sse41.c",
2297 "src/math/roundne-sse41.c",
2298 "src/math/roundu-sse41.c",
2299 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002300 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2301 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2302 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002303 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2304 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2305 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002306 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2307 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2308 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002309 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002310 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002311 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002312 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002313 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002314 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002315 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2316 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002317 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002318 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2319 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002320 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002321 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2322 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2323 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002324 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002325 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002326 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002327 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002328 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002329 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002330 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002331 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002332 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2333 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002334 "src/qs8-requantization/fp32-sse4.c",
2335 "src/qs8-requantization/precise-sse4.c",
2336 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002337 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2338 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2339 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2340 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002341 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2342 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2343 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2344 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002345 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2346 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2347 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2348 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002349 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2350 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2351 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2352 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002353 "src/qu8-requantization/precise-sse4.c",
2354 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002355]
2356
Marat Dukhan08c4a432019-10-03 09:29:21 -07002357AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002358 "src/f32-clamp/gen/avx-x8.c",
2359 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002360 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2361 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002362 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2363 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002364 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2365 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002366 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2367 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2368 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2369 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2370 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2371 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002372 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002373 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2374 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002375 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002376 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002377 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002378 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002379 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2380 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2381 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2382 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2383 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2384 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2385 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2386 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2387 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2388 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2389 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002390 "src/f32-hswish/gen/hswish-avx-x8.c",
2391 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002392 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002393 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2394 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002395 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002396 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002397 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002398 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002399 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2400 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002401 "src/f32-prelu/gen/avx-2x8.c",
2402 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002403 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002404 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002405 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002406 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2407 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2408 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2409 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2410 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2411 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2412 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2413 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2414 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2415 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2416 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2417 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2418 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2419 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2420 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2421 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2422 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2423 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2424 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2425 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002426 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2427 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2428 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2429 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2430 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2431 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2432 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2433 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002434 "src/f32-vbinary/gen/vmax-avx-x8.c",
2435 "src/f32-vbinary/gen/vmax-avx-x16.c",
2436 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2437 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2438 "src/f32-vbinary/gen/vmin-avx-x8.c",
2439 "src/f32-vbinary/gen/vmin-avx-x16.c",
2440 "src/f32-vbinary/gen/vminc-avx-x8.c",
2441 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002442 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2443 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2444 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2445 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2446 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2447 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2448 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2449 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002450 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2451 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2452 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2453 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002454 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2455 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2456 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2457 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002458 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2459 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002460 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2461 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002462 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2463 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2464 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2465 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2466 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2467 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002468 "src/f32-vscale/avx-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002469 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2470 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002471 "src/f32-vunary/gen/vabs-avx-x8.c",
2472 "src/f32-vunary/gen/vabs-avx-x16.c",
2473 "src/f32-vunary/gen/vneg-avx-x8.c",
2474 "src/f32-vunary/gen/vneg-avx-x16.c",
2475 "src/f32-vunary/gen/vsqr-avx-x8.c",
2476 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhand08abcb2020-09-21 00:26:25 -07002477 "src/math/exp-avx-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002478 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002479 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002480 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2481 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002482]
2483
Marat Dukhan1566fee2020-08-02 21:55:41 -07002484XOP_UKERNELS = [
2485 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002486 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002487 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002488 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002489 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002490 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002491 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2492 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002493 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002494 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2495 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002496 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002497 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2498 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2499 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002500 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002501 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002502 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002503 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002504 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002505 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002506 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002507 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002508 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2509 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002510 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2511 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2512 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2513 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2514 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2515 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2516 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2517 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002518]
2519
Marat Dukhanfda12b82019-11-21 12:27:59 -08002520FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002521 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2522 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002523 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2524 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002525 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2526 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002527 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2528 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2529 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2530 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2531 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2532 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002533 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002534 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2535 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2536 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2537 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002538 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002539 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2540 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002541 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002542 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2543 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002544 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2545 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2546 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002547 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2548 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2549 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2550 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2551 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2552 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2553 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2554 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2555 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2556 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2557 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2558 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2559 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2560 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002561 "src/f32-hswish/gen/hswish-fma3-x8.c",
2562 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002563 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002564 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2565 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2566 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2567 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002568 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002569 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2570 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002571 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002572 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2573 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002574 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2575 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2576 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002577 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2578 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2579 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2580 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2581 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2582 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2583 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2584 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002585 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002586 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002587 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002588]
2589
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002590AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002591 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2592 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002593 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002594 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002595 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002596 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2597 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002598 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002599 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2600 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2601 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002602 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002603 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2604 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002605 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002606 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002607 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002608 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2609 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002610 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002611 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2612 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2613 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002614 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002615 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2616 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002617 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002618 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002619 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002620 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2621 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002622 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002623 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2624 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2625 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002626 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002627 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2628 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2629 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2630 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2631 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2632 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2633 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2634 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2635 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2636 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2637 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2638 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2639 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2640 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2641 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2642 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2643 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2644 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2645 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2646 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2647 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2648 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2649 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2650 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2651 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2652 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2653 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2654 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2655 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2656 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002657 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2658 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2659 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2660 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2661 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2662 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2663 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2664 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2665 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2666 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2667 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2668 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2669 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2670 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2671 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2672 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2673 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2674 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2675 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2676 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2677 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2678 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2679 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2680 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002681 "src/math/exp-avx2-lut8-p3-perm.c",
2682 "src/math/exp-avx2-lut8-p4-perm.c",
2683 "src/math/exp-avx2-p5.c",
2684 "src/math/expminus-avx2-p5.c",
2685 "src/math/extexp-avx2-p5.c",
2686 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2687 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2688 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2689 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2690 "src/math/sigmoid-avx2-rr1-p5-div.c",
2691 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
2692 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2693 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2694 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2695 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2696 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2697 "src/math/sigmoid-avx2-rr2-p5-div.c",
2698 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2699 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002700 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002701 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002702 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2703 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002704 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002705 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002706 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002707 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002708 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002709 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002710 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002711 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002712 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2713 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2714 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002715 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2716 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2717 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2718 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2719 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2720 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2721 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2722 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002723]
2724
Marat Dukhan08c4a432019-10-03 09:29:21 -07002725AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002726 "src/f32-clamp/gen/avx512f-x16.c",
2727 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002728 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2729 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002730 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2731 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002732 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2733 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002734 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2735 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2736 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2737 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2738 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2739 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002740 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2741 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2742 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2743 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2744 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2745 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002746 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2747 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2748 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2749 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2750 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2751 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002752 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2753 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002754 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2755 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2756 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2757 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2758 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2759 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002760 "src/f32-prelu/gen/avx512f-2x16.c",
2761 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002762 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2763 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002764 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002765 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002766 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002767 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2768 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002769 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002770 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2771 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2772 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002773 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002774 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2775 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002776 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002777 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002778 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002779 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2780 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002781 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002782 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2783 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2784 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002785 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002786 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2787 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002788 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002789 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002790 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002791 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2792 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002793 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002794 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2795 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2796 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002797 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002798 "src/f32-relu/gen/avx512f-x16.c",
2799 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002800 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002801 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2802 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2803 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2804 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2805 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2806 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2807 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2808 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2809 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2810 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2811 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2812 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2813 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2814 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2815 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2816 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002817 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2818 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2819 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2820 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2821 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2822 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2823 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2824 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2825 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2826 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2827 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2828 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2829 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2830 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2831 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2832 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002833 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2834 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2835 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2836 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2837 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2838 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2839 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2840 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2841 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2842 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2843 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2844 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2845 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2846 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2847 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2848 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002849 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2850 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2851 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2852 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2853 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2854 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2855 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2856 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002857 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2858 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2859 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2860 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2861 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2862 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2863 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2864 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002865 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2866 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2867 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2868 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2869 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2870 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2871 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2872 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002873 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2874 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2875 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2876 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002877 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2878 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2879 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2880 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002881 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2882 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002883 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2884 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
2885 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2886 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2887 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2888 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2889 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2890 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002891 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002892 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2893 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2894 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2895 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2896 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2897 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2898 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2899 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2900 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2901 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2902 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2903 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2904 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2905 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2906 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2907 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2908 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2909 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2910 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2911 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2912 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2913 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2914 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2915 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002916 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2917 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2918 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2919 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2920 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2921 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2922 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2923 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002924 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2925 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2926 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2927 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2928 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2929 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002930 "src/math/exp-avx512f-lut16-p3-perm-scalef.c",
2931 "src/math/exp-avx512f-lut16-p3-perm.c",
2932 "src/math/exp-avx512f-lut32-p2-perm2-scalef.c",
2933 "src/math/exp-avx512f-lut32-p2-perm2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002934 "src/math/exp-avx512f-p5-scalef.c",
2935 "src/math/exp-avx512f-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002936 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002937 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002938 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002939 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002940 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002941 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002942 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002943 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002944 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002945 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
2946 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
2947 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
2948 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
2949 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2950 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2951 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2952 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2953 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
2954 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002955 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002956 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002957 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
2958 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
2959 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
2960 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002961 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002962 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002963 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002964]
2965
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002966AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002967 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2968 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002969 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2970 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2971 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2972 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2973 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2974 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2975 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2976 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2977]
2978
Frank Barchardbcedc082020-08-17 18:00:51 -07002979WASM32_ASM_UKERNELS = [
2980 "src/f32-relu/wasm_shr_x1.S",
2981 "src/f32-relu/wasm_shr_x2.S",
2982 "src/f32-relu/wasm_shr_x4.S",
2983]
2984
Marat Dukhan08c4a432019-10-03 09:29:21 -07002985AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002986 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002987 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002988 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2989 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002990 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002991 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002992 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002993 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002994 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2995 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002996 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2997 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2998 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2999 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003000]
3001
3002AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003003 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003004 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003005 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003006 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003007 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003008 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07003009 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003010 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
3011 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
3012 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
3013 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
3014 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
3015 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
3016 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003017 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
3018 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003019 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3020 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3021 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003022 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
3023 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003024 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3025 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3026 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3027 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003028 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003029 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3030 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003031 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3032 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3033 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3034 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003035 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003036 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003037 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003038 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003039 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3040 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3041 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3042 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3043 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3044 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3045 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3046 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3047 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3048 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3049 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3050 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3051 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3052 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3053 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3054 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3055 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3056 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3057 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3058 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3059 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3060 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003061 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003062 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003063 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3064 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003065 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3066 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3067 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3068 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3069 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3070 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003071 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3072 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3073 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3074 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003075 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3076 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003077 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07003078 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07003079 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003080]
3081
Marat Dukhan1b354632020-03-23 12:50:22 -07003082INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003083 "src/xnnpack/argmaxpool.h",
3084 "src/xnnpack/avgpool.h",
3085 "src/xnnpack/clamp.h",
3086 "src/xnnpack/common.h",
3087 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003088 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003089 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003090 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003091 "src/xnnpack/gavgpool.h",
3092 "src/xnnpack/gemm.h",
3093 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003094 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003095 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003096 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003097 "src/xnnpack/lut.h",
3098 "src/xnnpack/math.h",
3099 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003100 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003101 "src/xnnpack/packx.h",
3102 "src/xnnpack/pad.h",
3103 "src/xnnpack/params.h",
3104 "src/xnnpack/pavgpool.h",
3105 "src/xnnpack/ppmm.h",
3106 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003107 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003108 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003109 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003110 "src/xnnpack/rmax.h",
3111 "src/xnnpack/scalar-utils.h",
3112 "src/xnnpack/spmm.h",
3113 "src/xnnpack/unpool.h",
3114 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003115 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003116 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003117 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003118 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003119 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003120 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003121 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003122]
3123
3124INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003125 "include/xnnpack.h",
3126 "src/xnnpack/allocator.h",
3127 "src/xnnpack/compute.h",
3128 "src/xnnpack/im2col.h",
3129 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003130 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003131 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003132 "src/xnnpack/operator.h",
3133 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003134 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003135 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003136 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003137 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003138]
3139
Marat Dukhan1b354632020-03-23 12:50:22 -07003140ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003141 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003142]
3143
Marat Dukhan1b354632020-03-23 12:50:22 -07003144MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003145 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003146 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003147]
3148
Marat Dukhan1b354632020-03-23 12:50:22 -07003149MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003150 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003151 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003152 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003153 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003154]
3155
3156OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003157 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003158 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003159]
3160
3161WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003162 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003163 "src/xnnpack/operator.h",
3164 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003165]
3166
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003167LOGGING_COPTS = select({
3168 # No logging in optimized mode
3169 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3170 # Full logging in debug mode
3171 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3172 # Error-only logging in default (fastbuild) mode
3173 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3174})
3175
Marat Dukhan3b59de22020-06-03 20:15:19 -07003176LOGGING_SRCS = select({
3177 # No logging in optimized mode
3178 ":optimized_build": [],
3179 "//conditions:default": [
3180 "src/operator-strings.c",
3181 "src/subgraph-strings.c",
3182 ],
3183})
3184
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003185LOGGING_HDRS = [
3186 "src/xnnpack/log.h",
3187]
3188
Marat Dukhan08c4a432019-10-03 09:29:21 -07003189xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003190 name = "tables",
3191 srcs = TABLE_SRCS,
3192 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003193 gcc_copts = xnnpack_gcc_std_copts(),
3194 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003195)
3196
3197xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003198 name = "scalar_ukernels",
3199 srcs = SCALAR_UKERNELS,
3200 hdrs = INTERNAL_HDRS,
3201 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003202 gcc_copts = xnnpack_gcc_std_copts(),
3203 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003204 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003205 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003206 "@FP16",
3207 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003208 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003209 ],
3210)
3211
3212xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003213 name = "scalar_ukernels_test_mode",
3214 srcs = SCALAR_UKERNELS,
3215 hdrs = INTERNAL_HDRS,
3216 aarch32_copts = ["-marm"],
3217 copts = [
3218 "-UNDEBUG",
3219 "-DXNN_TEST_MODE=1",
3220 ],
3221 gcc_copts = xnnpack_gcc_std_copts(),
3222 msvc_copts = xnnpack_msvc_std_copts(),
3223 deps = [
3224 ":tables",
3225 "@FP16",
3226 "@FXdiv",
3227 "@pthreadpool",
3228 ],
3229)
3230
3231xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003232 name = "wasm_ukernels",
3233 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003234 gcc_copts = xnnpack_gcc_std_copts(),
3235 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003236 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003237 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003238 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003239 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003240 "@FP16",
3241 "@FXdiv",
3242 "@pthreadpool",
3243 ],
3244)
3245
3246xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003247 name = "wasm_ukernels_test_mode",
3248 hdrs = INTERNAL_HDRS,
3249 copts = [
3250 "-UNDEBUG",
3251 "-DXNN_TEST_MODE=1",
3252 ],
3253 gcc_copts = xnnpack_gcc_std_copts(),
3254 msvc_copts = xnnpack_msvc_std_copts(),
3255 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003256 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003257 deps = [
3258 ":tables",
3259 "@FP16",
3260 "@FXdiv",
3261 "@pthreadpool",
3262 ],
3263)
3264
3265xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003266 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003267 hdrs = INTERNAL_HDRS,
3268 aarch32_copts = [
3269 "-marm",
3270 "-mfpu=neon",
3271 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003272 gcc_copts = xnnpack_gcc_std_copts(),
3273 gcc_x86_copts = ["-msse2"],
3274 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003275 optimized_copts = [
3276 "-O3",
3277 "-ffast-math",
3278 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003279 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003280 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003281 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003282 "@FP16",
3283 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003284 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003285 ],
3286)
3287
3288xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003289 name = "psimd_fastmath_ukernels_test_mode",
3290 hdrs = INTERNAL_HDRS,
3291 aarch32_copts = [
3292 "-marm",
3293 "-mfpu=neon",
3294 ],
3295 copts = [
3296 "-UNDEBUG",
3297 "-DXNN_TEST_MODE=1",
3298 ],
3299 gcc_copts = xnnpack_gcc_std_copts(),
3300 gcc_x86_copts = ["-msse2"],
3301 msvc_copts = xnnpack_msvc_std_copts(),
3302 optimized_copts = [
3303 "-O3",
3304 "-ffast-math",
3305 ],
3306 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
3307 deps = [
3308 ":tables",
3309 "@FP16",
3310 "@psimd",
3311 "@pthreadpool",
3312 ],
3313)
3314
3315xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003316 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003317 hdrs = INTERNAL_HDRS,
3318 aarch32_copts = [
3319 "-marm",
3320 "-mfpu=neon",
3321 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003322 gcc_copts = xnnpack_gcc_std_copts(),
3323 gcc_x86_copts = ["-msse2"],
3324 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003325 optimized_copts = [
3326 "-O3",
3327 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003328 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003329 deps = [
3330 ":tables",
3331 "@FP16",
3332 "@psimd",
3333 "@pthreadpool",
3334 ],
3335)
3336
3337xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003338 name = "psimd_accmath_ukernels_test_mode",
3339 hdrs = INTERNAL_HDRS,
3340 aarch32_copts = [
3341 "-marm",
3342 "-mfpu=neon",
3343 ],
3344 copts = [
3345 "-UNDEBUG",
3346 "-DXNN_TEST_MODE=1",
3347 ],
3348 gcc_copts = xnnpack_gcc_std_copts(),
3349 gcc_x86_copts = ["-msse2"],
3350 msvc_copts = xnnpack_msvc_std_copts(),
3351 optimized_copts = [
3352 "-O3",
3353 ],
3354 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3355 deps = [
3356 ":tables",
3357 "@FP16",
3358 "@psimd",
3359 "@pthreadpool",
3360 ],
3361)
3362
3363xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003364 name = "neon_ukernels",
3365 hdrs = INTERNAL_HDRS,
3366 aarch32_copts = [
3367 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003368 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003369 "-mfpu=neon",
3370 ],
3371 aarch32_srcs = NEON_UKERNELS,
3372 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003373 gcc_copts = xnnpack_gcc_std_copts(),
3374 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003375 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003376 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003377 "@FP16",
3378 "@pthreadpool",
3379 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003380)
3381
3382xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003383 name = "neon_ukernels_test_mode",
3384 hdrs = INTERNAL_HDRS,
3385 aarch32_copts = [
3386 "-marm",
3387 "-march=armv7-a",
3388 "-mfpu=neon",
3389 ],
3390 aarch32_srcs = NEON_UKERNELS,
3391 aarch64_srcs = NEON_UKERNELS,
3392 copts = [
3393 "-UNDEBUG",
3394 "-DXNN_TEST_MODE=1",
3395 ],
3396 gcc_copts = xnnpack_gcc_std_copts(),
3397 msvc_copts = xnnpack_msvc_std_copts(),
3398 deps = [
3399 ":tables",
3400 "@FP16",
3401 "@pthreadpool",
3402 ],
3403)
3404
3405xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003406 name = "neonfma_ukernels",
3407 hdrs = INTERNAL_HDRS,
3408 aarch32_copts = [
3409 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003410 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003411 "-mfpu=neon-vfpv4",
3412 ],
3413 aarch32_srcs = NEONFMA_UKERNELS,
3414 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003415 apple_aarch32_copts = [
3416 "-mcpu=swift",
3417 "-mtune=generic",
3418 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003419 gcc_copts = xnnpack_gcc_std_copts(),
3420 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003421 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003422 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003423 "@FP16",
3424 "@pthreadpool",
3425 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003426)
3427
3428xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003429 name = "neonfma_ukernels_test_mode",
3430 hdrs = INTERNAL_HDRS,
3431 aarch32_copts = [
3432 "-marm",
3433 "-march=armv7-a",
3434 "-mfpu=neon-vfpv4",
3435 ],
3436 aarch32_srcs = NEONFMA_UKERNELS,
3437 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003438 apple_aarch32_copts = [
3439 "-mcpu=swift",
3440 "-mtune=generic",
3441 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003442 copts = [
3443 "-UNDEBUG",
3444 "-DXNN_TEST_MODE=1",
3445 ],
3446 gcc_copts = xnnpack_gcc_std_copts(),
3447 msvc_copts = xnnpack_msvc_std_copts(),
3448 deps = [
3449 ":tables",
3450 "@FP16",
3451 "@pthreadpool",
3452 ],
3453)
3454
3455xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003456 name = "neonv8_ukernels",
3457 hdrs = INTERNAL_HDRS,
3458 aarch32_copts = [
3459 "-marm",
3460 "-march=armv8-a",
3461 "-mfpu=neon-fp-armv8",
3462 ],
3463 aarch32_srcs = NEONV8_UKERNELS,
3464 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003465 apple_aarch32_copts = [
3466 "-mcpu=cyclone",
3467 "-mtune=generic",
3468 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003469 gcc_copts = xnnpack_gcc_std_copts(),
3470 msvc_copts = xnnpack_msvc_std_copts(),
3471 deps = [
3472 ":tables",
3473 "@FP16",
3474 "@pthreadpool",
3475 ],
3476)
3477
3478xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003479 name = "neonv8_ukernels_test_mode",
3480 hdrs = INTERNAL_HDRS,
3481 aarch32_copts = [
3482 "-marm",
3483 "-march=armv8-a",
3484 "-mfpu=neon-fp-armv8",
3485 ],
3486 aarch32_srcs = NEONV8_UKERNELS,
3487 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003488 apple_aarch32_copts = [
3489 "-mcpu=cyclone",
3490 "-mtune=generic",
3491 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003492 copts = [
3493 "-UNDEBUG",
3494 "-DXNN_TEST_MODE=1",
3495 ],
3496 gcc_copts = xnnpack_gcc_std_copts(),
3497 msvc_copts = xnnpack_msvc_std_copts(),
3498 deps = [
3499 ":tables",
3500 "@FP16",
3501 "@pthreadpool",
3502 ],
3503)
3504
3505xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003506 name = "neonfp16arith_ukernels",
3507 hdrs = INTERNAL_HDRS,
3508 aarch64_copts = ["-march=armv8.2-a+fp16"],
3509 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003510 gcc_copts = xnnpack_gcc_std_copts(),
3511 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003512 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003513 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003514 "@FP16",
3515 "@pthreadpool",
3516 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003517)
3518
3519xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003520 name = "neonfp16arith_ukernels_test_mode",
3521 hdrs = INTERNAL_HDRS,
3522 aarch64_copts = ["-march=armv8.2-a+fp16"],
3523 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3524 copts = [
3525 "-UNDEBUG",
3526 "-DXNN_TEST_MODE=1",
3527 ],
3528 gcc_copts = xnnpack_gcc_std_copts(),
3529 msvc_copts = xnnpack_msvc_std_copts(),
3530 deps = [
3531 ":tables",
3532 "@FP16",
3533 "@pthreadpool",
3534 ],
3535)
3536
3537xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003538 name = "neondot_ukernels",
3539 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003540 aarch32_copts = [
3541 "-marm",
3542 "-march=armv8.2-a+dotprod",
3543 "-mfpu=neon-fp-armv8",
3544 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003545 aarch32_srcs = NEONDOT_UKERNELS,
3546 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3547 aarch64_srcs = NEONDOT_UKERNELS,
3548 gcc_copts = xnnpack_gcc_std_copts(),
3549 msvc_copts = xnnpack_msvc_std_copts(),
3550 deps = [
3551 ":tables",
3552 "@FP16",
3553 "@pthreadpool",
3554 ],
3555)
3556
3557xnnpack_cc_library(
3558 name = "neondot_ukernels_test_mode",
3559 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003560 aarch32_copts = [
3561 "-marm",
3562 "-march=armv8.2-a+dotprod",
3563 "-mfpu=neon-fp-armv8",
3564 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003565 aarch32_srcs = NEONDOT_UKERNELS,
3566 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3567 aarch64_srcs = NEONDOT_UKERNELS,
3568 copts = [
3569 "-UNDEBUG",
3570 "-DXNN_TEST_MODE=1",
3571 ],
3572 gcc_copts = xnnpack_gcc_std_copts(),
3573 msvc_copts = xnnpack_msvc_std_copts(),
3574 deps = [
3575 ":tables",
3576 "@FP16",
3577 "@pthreadpool",
3578 ],
3579)
3580
3581xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003582 name = "sse2_ukernels",
3583 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003584 gcc_copts = xnnpack_gcc_std_copts(),
3585 gcc_x86_copts = ["-msse2"],
3586 msvc_copts = xnnpack_msvc_std_copts(),
3587 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003588 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003589 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003590 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003591 "@FP16",
3592 "@pthreadpool",
3593 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003594)
3595
3596xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003597 name = "sse2_ukernels_test_mode",
3598 hdrs = INTERNAL_HDRS,
3599 copts = [
3600 "-UNDEBUG",
3601 "-DXNN_TEST_MODE=1",
3602 ],
3603 gcc_copts = xnnpack_gcc_std_copts(),
3604 gcc_x86_copts = ["-msse2"],
3605 msvc_copts = xnnpack_msvc_std_copts(),
3606 msvc_x86_32_copts = ["/arch:SSE2"],
3607 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3608 deps = [
3609 ":tables",
3610 "@FP16",
3611 "@pthreadpool",
3612 ],
3613)
3614
3615xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003616 name = "ssse3_ukernels",
3617 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003618 gcc_copts = xnnpack_gcc_std_copts(),
3619 gcc_x86_copts = ["-mssse3"],
3620 msvc_copts = xnnpack_msvc_std_copts(),
3621 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003622 x86_srcs = SSSE3_UKERNELS,
3623 deps = [
3624 ":tables",
3625 "@FP16",
3626 "@pthreadpool",
3627 ],
3628)
3629
3630xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003631 name = "ssse3_ukernels_test_mode",
3632 hdrs = INTERNAL_HDRS,
3633 copts = [
3634 "-UNDEBUG",
3635 "-DXNN_TEST_MODE=1",
3636 ],
3637 gcc_copts = xnnpack_gcc_std_copts(),
3638 gcc_x86_copts = ["-mssse3"],
3639 msvc_copts = xnnpack_msvc_std_copts(),
3640 msvc_x86_32_copts = ["/arch:SSE2"],
3641 x86_srcs = SSSE3_UKERNELS,
3642 deps = [
3643 ":tables",
3644 "@FP16",
3645 "@pthreadpool",
3646 ],
3647)
3648
3649xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003650 name = "sse41_ukernels",
3651 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003652 gcc_copts = xnnpack_gcc_std_copts(),
3653 gcc_x86_copts = ["-msse4.1"],
3654 msvc_copts = xnnpack_msvc_std_copts(),
3655 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003656 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003657 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003658 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003659 "@FP16",
3660 "@pthreadpool",
3661 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003662)
3663
3664xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003665 name = "sse41_ukernels_test_mode",
3666 hdrs = INTERNAL_HDRS,
3667 copts = [
3668 "-UNDEBUG",
3669 "-DXNN_TEST_MODE=1",
3670 ],
3671 gcc_copts = xnnpack_gcc_std_copts(),
3672 gcc_x86_copts = ["-msse4.1"],
3673 msvc_copts = xnnpack_msvc_std_copts(),
3674 msvc_x86_32_copts = ["/arch:SSE2"],
3675 x86_srcs = SSE41_UKERNELS,
3676 deps = [
3677 ":tables",
3678 "@FP16",
3679 "@pthreadpool",
3680 ],
3681)
3682
3683xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003684 name = "avx_ukernels",
3685 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003686 gcc_copts = xnnpack_gcc_std_copts(),
3687 gcc_x86_copts = ["-mavx"],
3688 msvc_copts = xnnpack_msvc_std_copts(),
3689 msvc_x86_32_copts = ["/arch:AVX"],
3690 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003691 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003692 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003693 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003694 "@FP16",
3695 "@pthreadpool",
3696 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003697)
3698
3699xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003700 name = "avx_ukernels_test_mode",
3701 hdrs = INTERNAL_HDRS,
3702 copts = [
3703 "-UNDEBUG",
3704 "-DXNN_TEST_MODE=1",
3705 ],
3706 gcc_copts = xnnpack_gcc_std_copts(),
3707 gcc_x86_copts = ["-mavx"],
3708 msvc_copts = xnnpack_msvc_std_copts(),
3709 msvc_x86_32_copts = ["/arch:AVX"],
3710 msvc_x86_64_copts = ["/arch:AVX"],
3711 x86_srcs = AVX_UKERNELS,
3712 deps = [
3713 ":tables",
3714 "@FP16",
3715 "@pthreadpool",
3716 ],
3717)
3718
3719xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003720 name = "xop_ukernels",
3721 hdrs = INTERNAL_HDRS,
3722 gcc_copts = xnnpack_gcc_std_copts(),
3723 gcc_x86_copts = ["-mxop"],
3724 msvc_copts = xnnpack_msvc_std_copts(),
3725 msvc_x86_32_copts = ["/arch:AVX"],
3726 msvc_x86_64_copts = ["/arch:AVX"],
3727 x86_srcs = XOP_UKERNELS,
3728 deps = [
3729 ":tables",
3730 "@FP16",
3731 "@pthreadpool",
3732 ],
3733)
3734
3735xnnpack_cc_library(
3736 name = "xop_ukernels_test_mode",
3737 hdrs = INTERNAL_HDRS,
3738 copts = [
3739 "-UNDEBUG",
3740 "-DXNN_TEST_MODE=1",
3741 ],
3742 gcc_copts = xnnpack_gcc_std_copts(),
3743 gcc_x86_copts = ["-mxop"],
3744 msvc_copts = xnnpack_msvc_std_copts(),
3745 msvc_x86_32_copts = ["/arch:AVX"],
3746 msvc_x86_64_copts = ["/arch:AVX"],
3747 x86_srcs = XOP_UKERNELS,
3748 deps = [
3749 ":tables",
3750 "@FP16",
3751 "@pthreadpool",
3752 ],
3753)
3754
3755xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003756 name = "fma3_ukernels",
3757 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003758 gcc_copts = xnnpack_gcc_std_copts(),
3759 gcc_x86_copts = ["-mfma"],
3760 msvc_copts = xnnpack_msvc_std_copts(),
3761 msvc_x86_32_copts = ["/arch:AVX"],
3762 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003763 x86_srcs = FMA3_UKERNELS,
3764 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003765 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003766 "@FP16",
3767 "@pthreadpool",
3768 ],
3769)
3770
3771xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003772 name = "fma3_ukernels_test_mode",
3773 hdrs = INTERNAL_HDRS,
3774 copts = [
3775 "-UNDEBUG",
3776 "-DXNN_TEST_MODE=1",
3777 ],
3778 gcc_copts = xnnpack_gcc_std_copts(),
3779 gcc_x86_copts = ["-mfma"],
3780 msvc_copts = xnnpack_msvc_std_copts(),
3781 msvc_x86_32_copts = ["/arch:AVX"],
3782 msvc_x86_64_copts = ["/arch:AVX"],
3783 x86_srcs = FMA3_UKERNELS,
3784 deps = [
3785 ":tables",
3786 "@FP16",
3787 "@pthreadpool",
3788 ],
3789)
3790
3791xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003792 name = "avx2_ukernels",
3793 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003794 gcc_copts = xnnpack_gcc_std_copts(),
3795 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003796 "-mfma",
3797 "-mavx2",
3798 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003799 msvc_copts = xnnpack_msvc_std_copts(),
3800 msvc_x86_32_copts = ["/arch:AVX2"],
3801 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003802 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003803 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003804 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003805 "@FP16",
3806 "@pthreadpool",
3807 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003808)
3809
3810xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003811 name = "avx2_ukernels_test_mode",
3812 hdrs = INTERNAL_HDRS,
3813 copts = [
3814 "-UNDEBUG",
3815 "-DXNN_TEST_MODE=1",
3816 ],
3817 gcc_copts = xnnpack_gcc_std_copts(),
3818 gcc_x86_copts = [
3819 "-mfma",
3820 "-mavx2",
3821 ],
3822 msvc_copts = xnnpack_msvc_std_copts(),
3823 msvc_x86_32_copts = ["/arch:AVX2"],
3824 msvc_x86_64_copts = ["/arch:AVX2"],
3825 x86_srcs = AVX2_UKERNELS,
3826 deps = [
3827 ":tables",
3828 "@FP16",
3829 "@pthreadpool",
3830 ],
3831)
3832
3833xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003834 name = "avx512f_ukernels",
3835 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003836 gcc_copts = xnnpack_gcc_std_copts(),
3837 gcc_x86_copts = ["-mavx512f"],
3838 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3839 msvc_copts = xnnpack_msvc_std_copts(),
3840 msvc_x86_32_copts = ["/arch:AVX512"],
3841 msvc_x86_64_copts = ["/arch:AVX512"],
3842 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003843 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003844 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003845 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003846 "@FP16",
3847 "@pthreadpool",
3848 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003849)
3850
3851xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003852 name = "avx512f_ukernels_test_mode",
3853 hdrs = INTERNAL_HDRS,
3854 copts = [
3855 "-UNDEBUG",
3856 "-DXNN_TEST_MODE=1",
3857 ],
3858 gcc_copts = xnnpack_gcc_std_copts(),
3859 gcc_x86_copts = ["-mavx512f"],
3860 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3861 msvc_copts = xnnpack_msvc_std_copts(),
3862 msvc_x86_32_copts = ["/arch:AVX512"],
3863 msvc_x86_64_copts = ["/arch:AVX512"],
3864 msys_copts = ["-fno-asynchronous-unwind-tables"],
3865 x86_srcs = AVX512F_UKERNELS,
3866 deps = [
3867 ":tables",
3868 "@FP16",
3869 "@pthreadpool",
3870 ],
3871)
3872
3873xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003874 name = "avx512skx_ukernels",
3875 hdrs = INTERNAL_HDRS,
3876 gcc_copts = xnnpack_gcc_std_copts(),
3877 gcc_x86_copts = [
3878 "-mavx512f",
3879 "-mavx512cd",
3880 "-mavx512bw",
3881 "-mavx512dq",
3882 "-mavx512vl",
3883 ],
3884 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3885 msvc_copts = xnnpack_msvc_std_copts(),
3886 msvc_x86_32_copts = ["/arch:AVX512"],
3887 msvc_x86_64_copts = ["/arch:AVX512"],
3888 msys_copts = ["-fno-asynchronous-unwind-tables"],
3889 x86_srcs = AVX512SKX_UKERNELS,
3890 deps = [
3891 ":tables",
3892 "@FP16",
3893 "@pthreadpool",
3894 ],
3895)
3896
3897xnnpack_cc_library(
3898 name = "avx512skx_ukernels_test_mode",
3899 hdrs = INTERNAL_HDRS,
3900 copts = [
3901 "-UNDEBUG",
3902 "-DXNN_TEST_MODE=1",
3903 ],
3904 gcc_copts = xnnpack_gcc_std_copts(),
3905 gcc_x86_copts = [
3906 "-mavx512f",
3907 "-mavx512cd",
3908 "-mavx512bw",
3909 "-mavx512dq",
3910 "-mavx512vl",
3911 ],
3912 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3913 msvc_copts = xnnpack_msvc_std_copts(),
3914 msvc_x86_32_copts = ["/arch:AVX512"],
3915 msvc_x86_64_copts = ["/arch:AVX512"],
3916 msys_copts = ["-fno-asynchronous-unwind-tables"],
3917 x86_srcs = AVX512SKX_UKERNELS,
3918 deps = [
3919 ":tables",
3920 "@FP16",
3921 "@pthreadpool",
3922 ],
3923)
3924
3925xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003926 name = "asm_ukernels",
3927 hdrs = ["src/xnnpack/assembly.h"],
3928 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07003929 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003930 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003931 wasm_srcs = WASM32_ASM_UKERNELS,
3932 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003933)
3934
Marat Dukhan3b59de22020-06-03 20:15:19 -07003935xnnpack_cc_library(
3936 name = "logging_utils",
3937 srcs = LOGGING_SRCS,
3938 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3939 copts = LOGGING_COPTS + [
3940 "-Isrc",
3941 "-Iinclude",
3942 ] + select({
3943 ":debug_build": [],
3944 "//conditions:default": xnnpack_min_size_copts(),
3945 }),
3946 gcc_copts = xnnpack_gcc_std_copts(),
3947 msvc_copts = xnnpack_msvc_std_copts(),
3948 visibility = xnnpack_visibility(),
3949 deps = [
3950 "@FP16",
3951 "@clog",
3952 "@pthreadpool",
3953 ],
3954)
3955
Marat Dukhan08c4a432019-10-03 09:29:21 -07003956xnnpack_aggregate_library(
3957 name = "ukernels",
3958 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003959 ":neon_ukernels",
3960 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003961 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003962 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003963 ":asm_ukernels",
3964 ],
3965 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003966 ":neon_ukernels",
3967 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003968 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003969 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003970 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003971 ":asm_ukernels",
3972 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003973 generic_deps = [
3974 ":scalar_ukernels",
3975 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003976 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003977 ":psimd_fastmath_ukernels",
3978 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003979 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003980 wasm_deps = [
3981 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003982 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003983 ],
3984 wasmsimd_deps = [
3985 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003986 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003987 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003988 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003989 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003990 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003991 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003992 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003993 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003994 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003995 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003996 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003997 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003998 ],
3999)
4000
Marat Dukhan33fcf782020-05-24 14:27:15 -07004001xnnpack_aggregate_library(
4002 name = "ukernels_test_mode",
4003 aarch32_deps = [
4004 ":neon_ukernels_test_mode",
4005 ":neonfma_ukernels_test_mode",
4006 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004007 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004008 ":asm_ukernels",
4009 ],
4010 aarch64_deps = [
4011 ":neon_ukernels_test_mode",
4012 ":neonfma_ukernels_test_mode",
4013 ":neonv8_ukernels_test_mode",
4014 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004015 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004016 ":asm_ukernels",
4017 ],
4018 generic_deps = [
4019 ":scalar_ukernels_test_mode",
4020 ],
4021 psimd_deps = [
4022 ":psimd_fastmath_ukernels_test_mode",
4023 ":psimd_accmath_ukernels_test_mode",
4024 ],
4025 wasm_deps = [
4026 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004027 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004028 ],
4029 wasmsimd_deps = [
4030 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004031 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004032 ],
4033 x86_deps = [
4034 ":sse2_ukernels_test_mode",
4035 ":ssse3_ukernels_test_mode",
4036 ":sse41_ukernels_test_mode",
4037 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004038 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004039 ":fma3_ukernels_test_mode",
4040 ":avx2_ukernels_test_mode",
4041 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004042 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004043 ],
4044)
4045
Marat Dukhan86df3842020-06-17 14:03:31 -07004046# This target must be compatible with Android NDK r17 and compile with Clang 6.
4047# It is similar to :ukernels target, but excludes microkernels which require
4048# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07004049xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07004050 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004051 aarch32_deps = [
4052 ":neon_ukernels",
4053 ":neonfma_ukernels",
4054 ":neonv8_ukernels",
4055 ":asm_ukernels",
4056 ],
4057 aarch64_deps = [
4058 ":neon_ukernels",
4059 ":neonfma_ukernels",
4060 ":neonv8_ukernels",
4061 ":asm_ukernels",
4062 ],
4063 generic_deps = [
4064 ":scalar_ukernels",
4065 ],
4066 psimd_deps = [
4067 ":psimd_fastmath_ukernels",
4068 ":psimd_accmath_ukernels",
4069 ],
4070 wasm_deps = [
4071 ":wasm_ukernels",
4072 ],
4073 wasmsimd_deps = [
4074 ":wasm_ukernels",
4075 ],
4076 x86_deps = [
4077 ":sse2_ukernels",
4078 ":ssse3_ukernels",
4079 ":sse41_ukernels",
4080 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004081 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004082 ":fma3_ukernels",
4083 ":avx2_ukernels",
4084 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004085 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004086 ],
4087)
4088
Marat Dukhan08c4a432019-10-03 09:29:21 -07004089xnnpack_cc_library(
4090 name = "im2col",
4091 srcs = ["src/im2col.c"],
4092 hdrs = [
4093 "src/xnnpack/common.h",
4094 "src/xnnpack/im2col.h",
4095 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004096 gcc_copts = xnnpack_gcc_std_copts(),
4097 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004098)
4099
4100xnnpack_cc_library(
4101 name = "indirection",
4102 srcs = ["src/indirection.c"],
4103 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004104 gcc_copts = xnnpack_gcc_std_copts(),
4105 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004106 deps = [
4107 "@FP16",
4108 "@FXdiv",
4109 "@pthreadpool",
4110 ],
4111)
4112
4113xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004114 name = "indirection_test_mode",
4115 srcs = ["src/indirection.c"],
4116 hdrs = INTERNAL_HDRS,
4117 copts = [
4118 "-UNDEBUG",
4119 "-DXNN_TEST_MODE=1",
4120 ],
4121 gcc_copts = xnnpack_gcc_std_copts(),
4122 msvc_copts = xnnpack_msvc_std_copts(),
4123 deps = [
4124 "@FP16",
4125 "@FXdiv",
4126 "@pthreadpool",
4127 ],
4128)
4129
4130xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004131 name = "packing",
4132 srcs = ["src/packing.c"],
4133 hdrs = INTERNAL_HDRS,
4134 gcc_copts = xnnpack_gcc_std_copts(),
4135 msvc_copts = xnnpack_msvc_std_copts(),
4136 deps = [
4137 "@FP16",
4138 "@FXdiv",
4139 "@pthreadpool",
4140 ],
4141)
4142
4143xnnpack_cc_library(
4144 name = "packing_test_mode",
4145 srcs = ["src/packing.c"],
4146 hdrs = INTERNAL_HDRS,
4147 copts = [
4148 "-UNDEBUG",
4149 "-DXNN_TEST_MODE=1",
4150 ],
4151 gcc_copts = xnnpack_gcc_std_copts(),
4152 msvc_copts = xnnpack_msvc_std_copts(),
4153 deps = [
4154 "@FP16",
4155 "@FXdiv",
4156 "@pthreadpool",
4157 ],
4158)
4159
4160xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004161 name = "operator_run",
4162 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004163 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004164 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004165 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4166 "//conditions:default": [],
4167 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004168 gcc_copts = xnnpack_gcc_std_copts(),
4169 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004170 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004171 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004172 "@FP16",
4173 "@FXdiv",
4174 "@clog",
4175 "@pthreadpool",
4176 ],
4177)
4178
Chao Mei6ddfc602020-05-13 22:29:36 -07004179xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004180 name = "operator_run_test_mode",
4181 srcs = ["src/operator-run.c"],
4182 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4183 copts = LOGGING_COPTS + [
4184 "-UNDEBUG",
4185 "-DXNN_TEST_MODE=1",
4186 ] + select({
4187 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4188 "//conditions:default": [],
4189 }),
4190 gcc_copts = xnnpack_gcc_std_copts(),
4191 msvc_copts = xnnpack_msvc_std_copts(),
4192 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004193 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004194 "@FP16",
4195 "@FXdiv",
4196 "@clog",
4197 "@pthreadpool",
4198 ],
4199)
4200
4201xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004202 name = "memory_planner",
4203 srcs = ["src/memory-planner.c"],
4204 hdrs = INTERNAL_HDRS,
4205 defines = select({
4206 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4207 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4208 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4209 }),
4210 gcc_copts = xnnpack_gcc_std_copts(),
4211 msvc_copts = xnnpack_msvc_std_copts(),
4212 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004213 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004214 "@pthreadpool",
4215 ],
4216)
4217
Marat Dukhan33fcf782020-05-24 14:27:15 -07004218xnnpack_cc_library(
4219 name = "memory_planner_test_mode",
4220 srcs = ["src/memory-planner.c"],
4221 hdrs = INTERNAL_HDRS,
4222 copts = [
4223 "-UNDEBUG",
4224 "-DXNN_TEST_MODE=1",
4225 ],
4226 defines = select({
4227 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4228 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4229 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4230 }),
4231 gcc_copts = xnnpack_gcc_std_copts(),
4232 msvc_copts = xnnpack_msvc_std_copts(),
4233 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004234 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004235 "@pthreadpool",
4236 ],
4237)
4238
Marat Dukhan08c4a432019-10-03 09:29:21 -07004239cc_library(
4240 name = "enable_assembly",
4241 defines = select({
4242 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4243 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004244 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004245 }),
4246)
4247
Marat Dukhan9de90e02020-06-18 16:04:12 -07004248cc_library(
4249 name = "enable_sparse",
4250 defines = select({
4251 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4252 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07004253 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004254 }),
4255)
4256
Marat Dukhancf056b22019-10-07 10:26:29 -07004257xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004258 name = "operators",
4259 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004260 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004261 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004262 ],
4263 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004264 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004265 "-Isrc",
4266 "-Iinclude",
4267 ] + select({
4268 ":debug_build": [],
4269 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004270 }) + select({
4271 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4272 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004273 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004274 gcc_copts = xnnpack_gcc_std_copts(),
4275 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004276 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004277 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004278 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004279 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004280 "@FP16",
4281 "@FXdiv",
4282 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004283 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004284 ],
4285)
4286
Marat Dukhan10a38082020-04-17 03:58:35 -07004287xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004288 name = "operators_test_mode",
4289 srcs = OPERATOR_SRCS + [
4290 "src/memory.c",
4291 "src/operator-delete.c",
4292 ],
4293 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4294 copts = LOGGING_COPTS + [
4295 "-Isrc",
4296 "-Iinclude",
4297 "-UNDEBUG",
4298 "-DXNN_TEST_MODE=1",
4299 ] + select({
4300 ":debug_build": [],
4301 "//conditions:default": xnnpack_min_size_copts(),
4302 }) + select({
4303 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4304 "//conditions:default": [],
4305 }),
4306 gcc_copts = xnnpack_gcc_std_copts(),
4307 msvc_copts = xnnpack_msvc_std_copts(),
4308 deps = [
4309 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004310 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004311 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004312 "@FP16",
4313 "@FXdiv",
4314 "@clog",
4315 "@pthreadpool",
4316 ],
4317)
4318
4319xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004320 name = "XNNPACK",
4321 srcs = [
4322 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004323 "src/runtime.c",
4324 "src/subgraph.c",
4325 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004326 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004327 hdrs = ["include/xnnpack.h"],
4328 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004329 "-Isrc",
4330 "-Iinclude",
4331 ] + select({
4332 ":debug_build": [],
4333 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004334 }) + select({
4335 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4336 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004337 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004338 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004339 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004340 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004341 visibility = xnnpack_visibility(),
4342 deps = [
4343 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004344 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004345 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004346 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004347 ":operator_run",
4348 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004349 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004350 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004351 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004352 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004353 ] + select({
4354 ":emscripten": [],
4355 "//conditions:default": ["@cpuinfo"],
4356 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004357)
4358
Marat Dukhan10a38082020-04-17 03:58:35 -07004359xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004360 name = "XNNPACK_test_mode",
4361 srcs = [
4362 "src/init.c",
4363 "src/runtime.c",
4364 "src/subgraph.c",
4365 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004366 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004367 hdrs = ["include/xnnpack.h"],
4368 copts = LOGGING_COPTS + [
4369 "-Isrc",
4370 "-Iinclude",
4371 "-UNDEBUG",
4372 "-DXNN_TEST_MODE=1",
4373 ] + select({
4374 ":debug_build": [],
4375 "//conditions:default": xnnpack_min_size_copts(),
4376 }) + select({
4377 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4378 "//conditions:default": [],
4379 }),
4380 gcc_copts = xnnpack_gcc_std_copts(),
4381 includes = ["include"],
4382 msvc_copts = xnnpack_msvc_std_copts(),
4383 visibility = xnnpack_visibility(),
4384 deps = [
4385 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004386 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004387 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004388 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004389 ":operator_run_test_mode",
4390 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004391 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004392 "@clog",
4393 "@FP16",
4394 "@pthreadpool",
4395 ] + select({
4396 ":emscripten": [],
4397 "//conditions:default": ["@cpuinfo"],
4398 }),
4399)
4400
4401xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004402 name = "xnnpack_f32",
4403 srcs = [
4404 "src/init.c",
4405 "src/runtime.c",
4406 "src/subgraph.c",
4407 "src/tensor.c",
4408 ] + SUBGRAPH_SRCS,
4409 hdrs = ["include/xnnpack.h"],
4410 copts = LOGGING_COPTS + [
4411 "-Isrc",
4412 "-Iinclude",
4413 ] + select({
4414 ":debug_build": [],
4415 "//conditions:default": xnnpack_min_size_copts(),
4416 }) + select({
4417 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4418 "//conditions:default": [],
4419 }),
4420 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004421 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004422 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004423 "XNN_NO_U8_OPERATORS",
4424 "XNN_NO_X8_OPERATORS",
4425 "XNN_NO_F16_OPERATORS",
4426 "XNN_NO_X16_OPERATORS",
4427 ],
4428 gcc_copts = xnnpack_gcc_std_copts(),
4429 includes = ["include"],
4430 msvc_copts = xnnpack_msvc_std_copts(),
4431 visibility = xnnpack_visibility(),
4432 deps = [
4433 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004434 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004435 ":logging_utils",
4436 ":memory_planner",
4437 ":operator_run",
4438 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004439 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004440 "@clog",
4441 "@FP16",
4442 "@pthreadpool",
4443 ] + select({
4444 ":emscripten": [],
4445 "//conditions:default": ["@cpuinfo"],
4446 }),
4447)
4448
4449xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004450 name = "xnnpack_operators_nhwc_f32",
4451 srcs = [
4452 "src/init.c",
4453 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004454 hdrs = ["include/xnnpack.h"],
4455 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004456 "-Isrc",
4457 "-Iinclude",
4458 ] + select({
4459 ":debug_build": [],
4460 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004461 }) + select({
4462 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4463 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004464 }),
4465 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004466 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004467 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004468 "XNN_NO_U8_OPERATORS",
4469 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004470 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004471 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004472 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004473 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004474 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004475 visibility = xnnpack_visibility(),
4476 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004477 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004478 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004479 ":operator_run",
4480 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004481 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004482 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004483 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004484 ] + select({
4485 ":emscripten": [],
4486 "//conditions:default": ["@cpuinfo"],
4487 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004488)
4489
Marat Dukhancf056b22019-10-07 10:26:29 -07004490xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004491 name = "bench_utils",
4492 srcs = ["bench/utils.cc"],
4493 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004494 deps = [
4495 "@com_google_benchmark//:benchmark",
4496 "@cpuinfo",
4497 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004498)
4499
Frank Barchard7e955972019-10-11 10:34:25 -07004500######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004501
4502xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004503 name = "qs8_gemm_bench",
4504 srcs = [
4505 "bench/gemm.h",
4506 "bench/qs8-gemm.cc",
4507 "src/xnnpack/AlignedAllocator.h",
4508 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004509 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4510 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004511)
4512
4513xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004514 name = "qs8_requantization_bench",
4515 srcs = [
4516 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004517 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004518 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004519 ] + MICROKERNEL_BENCHMARK_HDRS,
4520 deps = MICROKERNEL_BENCHMARK_DEPS,
4521)
4522
4523xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004524 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004525 srcs = [
4526 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004527 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004528 "src/xnnpack/AlignedAllocator.h",
4529 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004530 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004531 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004532)
4533
4534xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004535 name = "qu8_requantization_bench",
4536 srcs = [
4537 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004538 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004539 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004540 ] + MICROKERNEL_BENCHMARK_HDRS,
4541 deps = MICROKERNEL_BENCHMARK_DEPS,
4542)
4543
4544xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004545 name = "f16_igemm_bench",
4546 srcs = [
4547 "bench/f16-igemm.cc",
4548 "bench/conv.h",
4549 "bench/google/conv.h",
4550 "src/xnnpack/AlignedAllocator.h",
4551 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004552 deps = MICROKERNEL_BENCHMARK_DEPS + [
4553 ":indirection",
4554 ":packing",
4555 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004556)
4557
4558xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004559 name = "f16_gemm_bench",
4560 srcs = [
4561 "bench/f16-gemm.cc",
4562 "bench/gemm.h",
4563 "src/xnnpack/AlignedAllocator.h",
4564 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004565 deps = MICROKERNEL_BENCHMARK_DEPS + [
4566 ":packing",
4567 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004568)
4569
4570xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004571 name = "f16_spmm_bench",
4572 srcs = [
4573 "bench/f16-spmm.cc",
4574 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004575 "src/xnnpack/AlignedAllocator.h",
4576 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004577 deps = MICROKERNEL_BENCHMARK_DEPS,
4578)
4579
4580xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004581 name = "f32_igemm_bench",
4582 srcs = [
4583 "bench/f32-igemm.cc",
4584 "bench/conv.h",
4585 "src/xnnpack/AlignedAllocator.h",
4586 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004587 deps = MICROKERNEL_BENCHMARK_DEPS + [
4588 ":indirection",
4589 ":packing",
4590 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004591)
4592
4593xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004594 name = "f16_relu_bench",
4595 srcs = [
4596 "bench/f16-relu.cc",
4597 "src/xnnpack/AlignedAllocator.h",
4598 ] + MICROKERNEL_BENCHMARK_HDRS,
4599 deps = MICROKERNEL_BENCHMARK_DEPS,
4600)
4601
4602xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004603 name = "f32_conv_hwc_bench",
4604 srcs = [
4605 "bench/f32-conv-hwc.cc",
4606 "bench/dconv.h",
4607 "src/xnnpack/AlignedAllocator.h",
4608 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004609 deps = MICROKERNEL_BENCHMARK_DEPS + [
4610 ":packing",
4611 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004612)
4613
4614xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004615 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004616 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004617 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004618 "bench/dconv.h",
4619 "src/xnnpack/AlignedAllocator.h",
4620 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004621 deps = MICROKERNEL_BENCHMARK_DEPS + [
4622 ":packing",
4623 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004624)
4625
4626xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004627 name = "f16_dwconv_bench",
4628 srcs = [
4629 "bench/f16-dwconv.cc",
4630 "bench/dwconv.h",
4631 "bench/google/dwconv.h",
4632 "src/xnnpack/AlignedAllocator.h",
4633 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004634 deps = MICROKERNEL_BENCHMARK_DEPS + [
4635 ":indirection",
4636 ":packing",
4637 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004638)
4639
4640xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004641 name = "f32_dwconv_bench",
4642 srcs = [
4643 "bench/f32-dwconv.cc",
4644 "bench/dwconv.h",
4645 "src/xnnpack/AlignedAllocator.h",
4646 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004647 deps = MICROKERNEL_BENCHMARK_DEPS + [
4648 ":indirection",
4649 ":packing",
4650 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004651)
4652
4653xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07004654 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004655 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07004656 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004657 "bench/dwconv.h",
4658 "src/xnnpack/AlignedAllocator.h",
4659 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004660 deps = MICROKERNEL_BENCHMARK_DEPS + [
4661 ":indirection",
4662 ":packing",
4663 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004664)
4665
4666xnnpack_benchmark(
4667 name = "f32_gemm_bench",
4668 srcs = [
4669 "bench/f32-gemm.cc",
4670 "bench/gemm.h",
4671 "src/xnnpack/AlignedAllocator.h",
4672 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004673 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004674 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004675)
4676
4677xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004678 name = "f32_hswish_bench",
4679 srcs = [
4680 "bench/f32-hswish.cc",
4681 "src/xnnpack/AlignedAllocator.h",
4682 ] + MICROKERNEL_BENCHMARK_HDRS,
4683 deps = MICROKERNEL_BENCHMARK_DEPS,
4684)
4685
4686xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004687 name = "f32_raddexpminusmax_bench",
4688 srcs = [
4689 "bench/f32-raddexpminusmax.cc",
4690 "src/xnnpack/AlignedAllocator.h",
4691 ] + MICROKERNEL_BENCHMARK_HDRS,
4692 deps = MICROKERNEL_BENCHMARK_DEPS,
4693)
4694
4695xnnpack_benchmark(
4696 name = "f32_raddextexp_bench",
4697 srcs = [
4698 "bench/f32-raddextexp.cc",
4699 "src/xnnpack/AlignedAllocator.h",
4700 ] + MICROKERNEL_BENCHMARK_HDRS,
4701 deps = MICROKERNEL_BENCHMARK_DEPS,
4702)
4703
4704xnnpack_benchmark(
4705 name = "f32_raddstoreexpminusmax_bench",
4706 srcs = [
4707 "bench/f32-raddstoreexpminusmax.cc",
4708 "src/xnnpack/AlignedAllocator.h",
4709 ] + MICROKERNEL_BENCHMARK_HDRS,
4710 deps = MICROKERNEL_BENCHMARK_DEPS,
4711)
4712
4713xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004714 name = "f32_relu_bench",
4715 srcs = [
4716 "bench/f32-relu.cc",
4717 "src/xnnpack/AlignedAllocator.h",
4718 ] + MICROKERNEL_BENCHMARK_HDRS,
4719 deps = MICROKERNEL_BENCHMARK_DEPS,
4720)
4721
4722xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004723 name = "f32_rmax_bench",
4724 srcs = [
4725 "bench/f32-rmax.cc",
4726 "src/xnnpack/AlignedAllocator.h",
4727 ] + MICROKERNEL_BENCHMARK_HDRS,
4728 deps = MICROKERNEL_BENCHMARK_DEPS,
4729)
4730
4731xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004732 name = "f32_sigmoid_bench",
4733 srcs = [
4734 "bench/f32-sigmoid.cc",
4735 "src/xnnpack/AlignedAllocator.h",
4736 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004737 deps = MICROKERNEL_BENCHMARK_DEPS,
4738)
4739
4740xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004741 name = "f32_spmm_bench",
4742 srcs = [
4743 "bench/f32-spmm.cc",
4744 "bench/gemm.h",
4745 "src/xnnpack/AlignedAllocator.h",
4746 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004747 deps = MICROKERNEL_BENCHMARK_DEPS,
4748)
4749
4750xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004751 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004752 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004753 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004754 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004755 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004756 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004757)
4758
4759xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004760 name = "f32_vscaleexpminusmax_bench",
4761 srcs = [
4762 "bench/f32-vscaleexpminusmax.cc",
4763 "src/xnnpack/AlignedAllocator.h",
4764 ] + MICROKERNEL_BENCHMARK_HDRS,
4765 deps = MICROKERNEL_BENCHMARK_DEPS,
4766)
4767
4768xnnpack_benchmark(
4769 name = "f32_vscaleextexp_bench",
4770 srcs = [
4771 "bench/f32-vscaleextexp.cc",
4772 "src/xnnpack/AlignedAllocator.h",
4773 ] + MICROKERNEL_BENCHMARK_HDRS,
4774 deps = MICROKERNEL_BENCHMARK_DEPS,
4775)
4776
4777xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004778 name = "f32_vsqrt_bench",
4779 srcs = [
4780 "bench/f32-vsqrt.cc",
4781 "src/xnnpack/AlignedAllocator.h",
4782 ] + MICROKERNEL_BENCHMARK_HDRS,
4783 deps = MICROKERNEL_BENCHMARK_DEPS,
4784)
4785
4786xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004787 name = "f32_im2col_gemm_bench",
4788 srcs = [
4789 "bench/f32-im2col-gemm.cc",
4790 "bench/conv.h",
4791 "src/xnnpack/AlignedAllocator.h",
4792 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004793 deps = MICROKERNEL_BENCHMARK_DEPS + [
4794 ":im2col",
4795 ":packing",
4796 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004797)
4798
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004799xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004800 name = "rounding_bench",
4801 srcs = [
4802 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004803 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004804 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004805 ] + MICROKERNEL_BENCHMARK_HDRS,
4806 deps = MICROKERNEL_BENCHMARK_DEPS,
4807)
4808
Marat Dukhan08c4a432019-10-03 09:29:21 -07004809########################### Benchmarks for operators ###########################
4810
4811xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004812 name = "average_pooling_bench",
4813 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004814 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004815 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004816 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004817)
4818
4819xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004820 name = "bankers_rounding_bench",
4821 srcs = ["bench/bankers-rounding.cc"],
4822 copts = xnnpack_optional_tflite_copts(),
4823 tags = ["nowin32"],
4824 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4825)
4826
4827xnnpack_benchmark(
4828 name = "ceiling_bench",
4829 srcs = ["bench/ceiling.cc"],
4830 copts = xnnpack_optional_tflite_copts(),
4831 tags = ["nowin32"],
4832 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4833)
4834
4835xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004836 name = "channel_shuffle_bench",
4837 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004838 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004839)
4840
4841xnnpack_benchmark(
4842 name = "convolution_bench",
4843 srcs = ["bench/convolution.cc"],
4844 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004845 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004846 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004847)
4848
4849xnnpack_benchmark(
4850 name = "deconvolution_bench",
4851 srcs = ["bench/deconvolution.cc"],
4852 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004853 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004854 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004855)
4856
4857xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004858 name = "floor_bench",
4859 srcs = ["bench/floor.cc"],
4860 copts = xnnpack_optional_tflite_copts(),
4861 tags = ["nowin32"],
4862 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4863)
4864
4865xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004866 name = "global_average_pooling_bench",
4867 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004868 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004869)
4870
4871xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004872 name = "hardswish_bench",
4873 srcs = ["bench/hardswish.cc"],
4874 copts = xnnpack_optional_tflite_copts(),
4875 tags = ["nowin32"],
4876 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4877)
4878
4879xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004880 name = "max_pooling_bench",
4881 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004882 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004883)
4884
4885xnnpack_benchmark(
4886 name = "sigmoid_bench",
4887 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004888 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004889 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004890 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004891)
4892
4893xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004894 name = "prelu_bench",
4895 srcs = ["bench/prelu.cc"],
4896 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004897 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004898 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004899)
4900
4901xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004902 name = "softmax_bench",
4903 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004904 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004905 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004906 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004907)
4908
Marat Dukhan87727142020-06-24 15:24:10 -07004909xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004910 name = "square_root_bench",
4911 srcs = ["bench/square-root.cc"],
4912 copts = xnnpack_optional_tflite_copts(),
4913 tags = ["nowin32"],
4914 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4915)
4916
4917xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004918 name = "truncation_bench",
4919 srcs = ["bench/truncation.cc"],
4920 deps = OPERATOR_BENCHMARK_DEPS,
4921)
4922
Marat Dukhanc068bb62019-10-04 13:24:39 -07004923############################# End-to-end benchmarks ############################
4924
4925cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004926 name = "fp32_mobilenet_v1",
4927 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004928 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004929 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004930 linkstatic = True,
4931 deps = [
4932 ":XNNPACK",
4933 "@pthreadpool",
4934 ],
4935)
4936
4937cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004938 name = "fp16_mobilenet_v1",
4939 srcs = ["models/fp16-mobilenet-v1.cc"],
4940 hdrs = ["models/models.h"],
4941 copts = xnnpack_std_cxxopts(),
4942 linkstatic = True,
4943 deps = [
4944 ":XNNPACK",
4945 "@FP16",
4946 "@pthreadpool",
4947 ],
4948)
4949
4950cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004951 name = "qs8_mobilenet_v1",
4952 srcs = ["models/qs8-mobilenet-v1.cc"],
4953 hdrs = ["models/models.h"],
4954 copts = xnnpack_std_cxxopts(),
4955 linkstatic = True,
4956 deps = [
4957 ":XNNPACK",
4958 "@pthreadpool",
4959 ],
4960)
4961
4962cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004963 name = "qs8_mobilenet_v2",
4964 srcs = ["models/qs8-mobilenet-v2.cc"],
4965 hdrs = ["models/models.h"],
4966 copts = xnnpack_std_cxxopts(),
4967 linkstatic = True,
4968 deps = [
4969 ":XNNPACK",
4970 "@pthreadpool",
4971 ],
4972)
4973
4974cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004975 name = "fp32_mobilenet_v2",
4976 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004977 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004978 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004979 linkstatic = True,
4980 deps = [
4981 ":XNNPACK",
4982 "@pthreadpool",
4983 ],
4984)
4985
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004986cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004987 name = "fp16_mobilenet_v2",
4988 srcs = ["models/fp16-mobilenet-v2.cc"],
4989 hdrs = ["models/models.h"],
4990 copts = xnnpack_std_cxxopts(),
4991 linkstatic = True,
4992 deps = [
4993 ":XNNPACK",
4994 "@FP16",
4995 "@pthreadpool",
4996 ],
4997)
4998
4999cc_library(
5000 name = "fp32_mobilenet_v3_large",
5001 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005002 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005003 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005004 linkstatic = True,
5005 deps = [
5006 ":XNNPACK",
5007 "@pthreadpool",
5008 ],
5009)
5010
5011cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005012 name = "fp16_mobilenet_v3_large",
5013 srcs = ["models/fp16-mobilenet-v3-large.cc"],
5014 hdrs = ["models/models.h"],
5015 copts = xnnpack_std_cxxopts(),
5016 linkstatic = True,
5017 deps = [
5018 ":XNNPACK",
5019 "@FP16",
5020 "@pthreadpool",
5021 ],
5022)
5023
5024cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005025 name = "fp32_mobilenet_v3_small",
5026 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005027 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005028 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005029 linkstatic = True,
5030 deps = [
5031 ":XNNPACK",
5032 "@pthreadpool",
5033 ],
5034)
5035
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005036cc_library(
5037 name = "fp16_mobilenet_v3_small",
5038 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5039 hdrs = ["models/models.h"],
5040 copts = xnnpack_std_cxxopts(),
5041 linkstatic = True,
5042 deps = [
5043 ":XNNPACK",
5044 "@FP16",
5045 "@pthreadpool",
5046 ],
5047)
5048
Marat Dukhanc068bb62019-10-04 13:24:39 -07005049xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005050 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005051 srcs = [
5052 "bench/f32-dwconv-e2e.cc",
5053 "bench/end2end.h",
5054 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005055 deps = MICROKERNEL_BENCHMARK_DEPS + [
5056 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005057 ":fp32_mobilenet_v1",
5058 ":fp32_mobilenet_v2",
5059 ":fp32_mobilenet_v3_large",
5060 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005061 ],
5062)
5063
5064xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005065 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005066 srcs = [
5067 "bench/f32-gemm-e2e.cc",
5068 "bench/end2end.h",
5069 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005070 deps = MICROKERNEL_BENCHMARK_DEPS + [
5071 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005072 ":fp32_mobilenet_v1",
5073 ":fp32_mobilenet_v2",
5074 ":fp32_mobilenet_v3_large",
5075 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005076 ],
5077)
5078
5079xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005080 name = "end2end_bench",
5081 srcs = ["bench/end2end.cc"],
5082 deps = [
5083 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005084 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005085 ":fp16_mobilenet_v1",
5086 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005087 ":fp16_mobilenet_v3_large",
5088 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005089 ":fp32_mobilenet_v1",
5090 ":fp32_mobilenet_v2",
5091 ":fp32_mobilenet_v3_large",
5092 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005093 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005094 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005095 "@pthreadpool",
5096 ],
5097)
5098
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005099#################### Accuracy evaluation for math functions ####################
5100
5101xnnpack_benchmark(
5102 name = "f32_exp_eval",
5103 srcs = [
5104 "eval/f32-exp.cc",
5105 "src/xnnpack/AlignedAllocator.h",
5106 ] + ACCURACY_EVAL_HDRS,
5107 deps = ACCURACY_EVAL_DEPS,
5108)
5109
Marat Dukhan515c9772019-10-17 18:07:57 -07005110xnnpack_benchmark(
5111 name = "f32_expminus_eval",
5112 srcs = [
5113 "eval/f32-expminus.cc",
5114 "src/xnnpack/AlignedAllocator.h",
5115 ] + ACCURACY_EVAL_HDRS,
5116 deps = ACCURACY_EVAL_DEPS,
5117)
5118
Marat Dukhan98ba4412019-10-23 02:14:28 -07005119xnnpack_benchmark(
Marat Dukhana438aca2020-11-20 15:45:01 -08005120 name = "f32_expm1minus_eval",
5121 srcs = [
5122 "eval/f32-expm1minus.cc",
5123 "src/xnnpack/AlignedAllocator.h",
5124 ] + ACCURACY_EVAL_HDRS,
5125 deps = ACCURACY_EVAL_DEPS,
5126)
5127
5128xnnpack_benchmark(
Marat Dukhan98ba4412019-10-23 02:14:28 -07005129 name = "f32_extexp_eval",
5130 srcs = [
5131 "eval/f32-extexp.cc",
5132 "src/xnnpack/AlignedAllocator.h",
5133 ] + ACCURACY_EVAL_HDRS,
5134 deps = ACCURACY_EVAL_DEPS,
5135)
5136
Marat Dukhan8853b822020-05-07 12:19:01 -07005137xnnpack_unit_test(
5138 name = "f32_roundne_eval",
5139 srcs = [
5140 "eval/f32-roundne.cc",
5141 "src/xnnpack/AlignedAllocator.h",
5142 "src/xnnpack/math-stubs.h",
5143 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005144 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005145 deps = MICROKERNEL_TEST_DEPS,
5146)
5147
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005148xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005149 name = "f32_roundd_eval",
5150 srcs = [
5151 "eval/f32-roundd.cc",
5152 "src/xnnpack/AlignedAllocator.h",
5153 "src/xnnpack/math-stubs.h",
5154 ] + MICROKERNEL_TEST_HDRS,
5155 automatic = False,
5156 deps = MICROKERNEL_TEST_DEPS,
5157)
5158
5159xnnpack_unit_test(
5160 name = "f32_roundu_eval",
5161 srcs = [
5162 "eval/f32-roundu.cc",
5163 "src/xnnpack/AlignedAllocator.h",
5164 "src/xnnpack/math-stubs.h",
5165 ] + MICROKERNEL_TEST_HDRS,
5166 automatic = False,
5167 deps = MICROKERNEL_TEST_DEPS,
5168)
5169
5170xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005171 name = "f32_roundz_eval",
5172 srcs = [
5173 "eval/f32-roundz.cc",
5174 "src/xnnpack/AlignedAllocator.h",
5175 "src/xnnpack/math-stubs.h",
5176 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005177 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005178 deps = MICROKERNEL_TEST_DEPS,
5179)
5180
Marat Dukhan346a9e52019-11-15 09:06:30 -08005181xnnpack_benchmark(
5182 name = "f32_sigmoid_eval",
5183 srcs = [
5184 "eval/f32-sigmoid.cc",
5185 "src/xnnpack/AlignedAllocator.h",
5186 ] + ACCURACY_EVAL_HDRS,
5187 deps = ACCURACY_EVAL_DEPS,
5188)
5189
Marat Dukhan84000762020-06-29 18:38:43 -07005190xnnpack_benchmark(
5191 name = "f32_sqrt_eval",
5192 srcs = [
5193 "eval/f32-sqrt.cc",
5194 "src/xnnpack/AlignedAllocator.h",
5195 ] + ACCURACY_EVAL_HDRS,
5196 deps = ACCURACY_EVAL_DEPS + [
5197 ":bench_utils",
5198 ],
5199)
5200
Marat Dukhan08c4a432019-10-03 09:29:21 -07005201######################### Unit tests for micro-kernels #########################
5202
5203xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005204 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005205 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005206 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005207 "test/gemm-microkernel-tester.h",
5208 "src/xnnpack/AlignedAllocator.h",
5209 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005210 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005211)
5212
5213xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005214 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005215 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005216 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005217 "test/spmm-microkernel-tester.h",
5218 "src/xnnpack/AlignedAllocator.h",
5219 ] + MICROKERNEL_TEST_HDRS,
5220 deps = MICROKERNEL_TEST_DEPS,
5221)
5222
5223xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005224 name = "f32_argmaxpool_test",
5225 srcs = [
5226 "test/f32-argmaxpool.cc",
5227 "test/argmaxpool-microkernel-tester.h",
5228 "src/xnnpack/AlignedAllocator.h",
5229 ] + MICROKERNEL_TEST_HDRS,
5230 deps = MICROKERNEL_TEST_DEPS,
5231)
5232
5233xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005234 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005235 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005236 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005237 "test/avgpool-microkernel-tester.h",
5238 "src/xnnpack/AlignedAllocator.h",
5239 ] + MICROKERNEL_TEST_HDRS,
5240 deps = MICROKERNEL_TEST_DEPS,
5241)
5242
5243xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005244 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005245 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005246 "test/f32-ibilinear.cc",
5247 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005248 "src/xnnpack/AlignedAllocator.h",
5249 ] + MICROKERNEL_TEST_HDRS,
5250 deps = MICROKERNEL_TEST_DEPS,
5251)
5252
5253xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005254 name = "f32_ibilinear_chw_test",
5255 srcs = [
5256 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005257 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005258 "src/xnnpack/AlignedAllocator.h",
5259 ] + MICROKERNEL_TEST_HDRS,
5260 deps = MICROKERNEL_TEST_DEPS,
5261)
5262
5263xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005264 name = "f16_clamp_test",
5265 srcs = [
5266 "test/f16-clamp.cc",
5267 "test/clamp-microkernel-tester.h",
5268 ] + MICROKERNEL_TEST_HDRS,
5269 deps = MICROKERNEL_TEST_DEPS,
5270)
5271
5272xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005273 name = "f32_clamp_test",
5274 srcs = [
5275 "test/f32-clamp.cc",
5276 "test/clamp-microkernel-tester.h",
5277 ] + MICROKERNEL_TEST_HDRS,
5278 deps = MICROKERNEL_TEST_DEPS,
5279)
5280
5281xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005282 name = "f32_igemm_test",
5283 srcs = [
5284 "test/f32-igemm.cc",
5285 "test/gemm-microkernel-tester.h",
5286 "src/xnnpack/AlignedAllocator.h",
5287 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005288 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005289)
5290
5291xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005292 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005293 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005294 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005295 "test/gemm-microkernel-tester.h",
5296 "src/xnnpack/AlignedAllocator.h",
5297 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005298 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005299)
5300
5301xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005302 name = "f32_igemm_minmax_test",
5303 srcs = [
5304 "test/f32-igemm-minmax.cc",
5305 "test/gemm-microkernel-tester.h",
5306 "src/xnnpack/AlignedAllocator.h",
5307 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005308 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005309)
5310
5311xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005312 name = "f16_igemm_minmax_test",
5313 srcs = [
5314 "test/f16-igemm-minmax.cc",
5315 "test/gemm-microkernel-tester.h",
5316 "src/xnnpack/AlignedAllocator.h",
5317 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005318 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005319)
5320
5321xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005322 name = "f32_conv_hwc_test",
5323 srcs = [
5324 "test/f32-conv-hwc.cc",
5325 "test/conv-hwc-microkernel-tester.h",
5326 "src/xnnpack/AlignedAllocator.h",
5327 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005328 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005329)
5330
5331xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005332 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005333 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005334 "test/f32-conv-hwc2chw.cc",
5335 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005336 "src/xnnpack/AlignedAllocator.h",
5337 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005338 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005339)
5340
5341xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005342 name = "f32_dwconv_test",
5343 srcs = [
5344 "test/f32-dwconv.cc",
5345 "test/dwconv-microkernel-tester.h",
5346 "src/xnnpack/AlignedAllocator.h",
5347 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005348 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005349)
5350
5351xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005352 name = "f16_dwconv_minmax_test",
5353 srcs = [
5354 "test/f16-dwconv-minmax.cc",
5355 "test/dwconv-microkernel-tester.h",
5356 "src/xnnpack/AlignedAllocator.h",
5357 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005358 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005359)
5360
5361xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005362 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005363 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005364 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005365 "test/dwconv-microkernel-tester.h",
5366 "src/xnnpack/AlignedAllocator.h",
5367 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005368 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005369)
5370
5371xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005372 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005373 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005374 "test/f32-dwconv2d-chw.cc",
5375 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005376 "src/xnnpack/AlignedAllocator.h",
5377 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005378 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005379)
5380
5381xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005382 name = "f16_gavgpool_minmax_test",
5383 srcs = [
5384 "test/f16-gavgpool-minmax.cc",
5385 "test/gavgpool-microkernel-tester.h",
5386 "src/xnnpack/AlignedAllocator.h",
5387 ] + MICROKERNEL_TEST_HDRS,
5388 deps = MICROKERNEL_TEST_DEPS,
5389)
5390
5391xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005392 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005393 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005394 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005395 "test/gavgpool-microkernel-tester.h",
5396 "src/xnnpack/AlignedAllocator.h",
5397 ] + MICROKERNEL_TEST_HDRS,
5398 deps = MICROKERNEL_TEST_DEPS,
5399)
5400
5401xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005402 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005403 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005404 "test/f32-gavgpool-cw.cc",
5405 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005406 "src/xnnpack/AlignedAllocator.h",
5407 ] + MICROKERNEL_TEST_HDRS,
5408 deps = MICROKERNEL_TEST_DEPS,
5409)
5410
5411xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005412 name = "f32_gemm_test",
5413 srcs = [
5414 "test/f32-gemm.cc",
5415 "test/gemm-microkernel-tester.h",
5416 "src/xnnpack/AlignedAllocator.h",
5417 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005418 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005419)
5420
5421xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005422 name = "f32_gemm_relu_test",
5423 srcs = [
5424 "test/f32-gemm-relu.cc",
5425 "test/gemm-microkernel-tester.h",
5426 "src/xnnpack/AlignedAllocator.h",
5427 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005428 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005429)
5430
5431xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005432 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005433 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005434 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005435 "test/gemm-microkernel-tester.h",
5436 "src/xnnpack/AlignedAllocator.h",
5437 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005438 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005439)
5440
5441xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005442 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005443 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005444 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005445 "test/gemm-microkernel-tester.h",
5446 "src/xnnpack/AlignedAllocator.h",
5447 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005448 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005449)
5450
5451xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005452 name = "f16_hswish_test",
5453 srcs = [
5454 "test/f16-hswish.cc",
5455 "test/hswish-microkernel-tester.h",
5456 ] + MICROKERNEL_TEST_HDRS,
5457 deps = MICROKERNEL_TEST_DEPS,
5458)
5459
5460xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005461 name = "f32_hswish_test",
5462 srcs = [
5463 "test/f32-hswish.cc",
5464 "test/hswish-microkernel-tester.h",
5465 ] + MICROKERNEL_TEST_HDRS,
5466 deps = MICROKERNEL_TEST_DEPS,
5467)
5468
5469xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005470 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005471 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005472 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005473 "test/maxpool-microkernel-tester.h",
5474 ] + MICROKERNEL_TEST_HDRS,
5475 deps = MICROKERNEL_TEST_DEPS,
5476)
5477
5478xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005479 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005480 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005481 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005482 "test/avgpool-microkernel-tester.h",
5483 "src/xnnpack/AlignedAllocator.h",
5484 ] + MICROKERNEL_TEST_HDRS,
5485 deps = MICROKERNEL_TEST_DEPS,
5486)
5487
5488xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005489 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005490 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005491 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005492 "test/gemm-microkernel-tester.h",
5493 "src/xnnpack/AlignedAllocator.h",
5494 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005495 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005496)
5497
5498xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005499 name = "f16_prelu_test",
5500 srcs = [
5501 "test/f16-prelu.cc",
5502 "test/prelu-microkernel-tester.h",
5503 "src/xnnpack/AlignedAllocator.h",
5504 ] + MICROKERNEL_TEST_HDRS,
5505 deps = MICROKERNEL_TEST_DEPS,
5506)
5507
5508xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005509 name = "f32_prelu_test",
5510 srcs = [
5511 "test/f32-prelu.cc",
5512 "test/prelu-microkernel-tester.h",
5513 "src/xnnpack/AlignedAllocator.h",
5514 ] + MICROKERNEL_TEST_HDRS,
5515 deps = MICROKERNEL_TEST_DEPS,
5516)
5517
5518xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005519 name = "f32_raddexpminusmax_test",
5520 srcs = [
5521 "test/f32-raddexpminusmax.cc",
5522 "test/raddexpminusmax-microkernel-tester.h",
5523 ] + MICROKERNEL_TEST_HDRS,
5524 deps = MICROKERNEL_TEST_DEPS,
5525)
5526
5527xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005528 name = "f32_raddextexp_test",
5529 srcs = [
5530 "test/f32-raddextexp.cc",
5531 "test/raddextexp-microkernel-tester.h",
5532 ] + MICROKERNEL_TEST_HDRS,
5533 deps = MICROKERNEL_TEST_DEPS,
5534)
5535
5536xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005537 name = "f32_raddstoreexpminusmax_test",
5538 srcs = [
5539 "test/f32-raddstoreexpminusmax.cc",
5540 "test/raddstoreexpminusmax-microkernel-tester.h",
5541 ] + MICROKERNEL_TEST_HDRS,
5542 deps = MICROKERNEL_TEST_DEPS,
5543)
5544
5545xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005546 name = "f32_relu_test",
5547 srcs = [
5548 "test/f32-relu.cc",
5549 "test/vunary-microkernel-tester.h",
5550 ] + MICROKERNEL_TEST_HDRS,
5551 deps = MICROKERNEL_TEST_DEPS,
5552)
5553
5554xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005555 name = "f32_rmax_test",
5556 srcs = [
5557 "test/f32-rmax.cc",
5558 "test/rmax-microkernel-tester.h",
5559 ] + MICROKERNEL_TEST_HDRS,
5560 deps = MICROKERNEL_TEST_DEPS,
5561)
5562
5563xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005564 name = "f32_sigmoid_test",
5565 srcs = [
5566 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005567 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005568 ] + MICROKERNEL_TEST_HDRS,
5569 deps = MICROKERNEL_TEST_DEPS,
5570)
5571
5572xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005573 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005574 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005575 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005576 "test/spmm-microkernel-tester.h",
5577 "src/xnnpack/AlignedAllocator.h",
5578 ] + MICROKERNEL_TEST_HDRS,
5579 deps = MICROKERNEL_TEST_DEPS,
5580)
5581
5582xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005583 name = "f16_vadd_minmax_test",
5584 srcs = [
5585 "test/f16-vadd-minmax.cc",
5586 "test/vbinary-microkernel-tester.h",
5587 ] + MICROKERNEL_TEST_HDRS,
5588 deps = MICROKERNEL_TEST_DEPS,
5589)
5590
5591xnnpack_unit_test(
5592 name = "f16_vaddc_minmax_test",
5593 srcs = [
5594 "test/f16-vaddc-minmax.cc",
5595 "test/vbinaryc-microkernel-tester.h",
5596 ] + MICROKERNEL_TEST_HDRS,
5597 deps = MICROKERNEL_TEST_DEPS,
5598)
5599
5600xnnpack_unit_test(
5601 name = "f16_vdiv_minmax_test",
5602 srcs = [
5603 "test/f16-vdiv-minmax.cc",
5604 "test/vbinary-microkernel-tester.h",
5605 ] + MICROKERNEL_TEST_HDRS,
5606 deps = MICROKERNEL_TEST_DEPS,
5607)
5608
5609xnnpack_unit_test(
5610 name = "f16_vdivc_minmax_test",
5611 srcs = [
5612 "test/f16-vdivc-minmax.cc",
5613 "test/vbinaryc-microkernel-tester.h",
5614 ] + MICROKERNEL_TEST_HDRS,
5615 deps = MICROKERNEL_TEST_DEPS,
5616)
5617
5618xnnpack_unit_test(
5619 name = "f16_vrdivc_minmax_test",
5620 srcs = [
5621 "test/f16-vrdivc-minmax.cc",
5622 "test/vbinaryc-microkernel-tester.h",
5623 ] + MICROKERNEL_TEST_HDRS,
5624 deps = MICROKERNEL_TEST_DEPS,
5625)
5626
5627xnnpack_unit_test(
5628 name = "f16_vmax_test",
5629 srcs = [
5630 "test/f16-vmax.cc",
5631 "test/vbinary-microkernel-tester.h",
5632 ] + MICROKERNEL_TEST_HDRS,
5633 deps = MICROKERNEL_TEST_DEPS,
5634)
5635
5636xnnpack_unit_test(
5637 name = "f16_vmaxc_test",
5638 srcs = [
5639 "test/f16-vmaxc.cc",
5640 "test/vbinaryc-microkernel-tester.h",
5641 ] + MICROKERNEL_TEST_HDRS,
5642 deps = MICROKERNEL_TEST_DEPS,
5643)
5644
5645xnnpack_unit_test(
5646 name = "f16_vmin_test",
5647 srcs = [
5648 "test/f16-vmin.cc",
5649 "test/vbinary-microkernel-tester.h",
5650 ] + MICROKERNEL_TEST_HDRS,
5651 deps = MICROKERNEL_TEST_DEPS,
5652)
5653
5654xnnpack_unit_test(
5655 name = "f16_vminc_test",
5656 srcs = [
5657 "test/f16-vminc.cc",
5658 "test/vbinaryc-microkernel-tester.h",
5659 ] + MICROKERNEL_TEST_HDRS,
5660 deps = MICROKERNEL_TEST_DEPS,
5661)
5662
5663xnnpack_unit_test(
5664 name = "f16_vmul_minmax_test",
5665 srcs = [
5666 "test/f16-vmul-minmax.cc",
5667 "test/vbinary-microkernel-tester.h",
5668 ] + MICROKERNEL_TEST_HDRS,
5669 deps = MICROKERNEL_TEST_DEPS,
5670)
5671
5672xnnpack_unit_test(
5673 name = "f16_vmulc_minmax_test",
5674 srcs = [
5675 "test/f16-vmulc-minmax.cc",
5676 "test/vbinaryc-microkernel-tester.h",
5677 ] + MICROKERNEL_TEST_HDRS,
5678 deps = MICROKERNEL_TEST_DEPS,
5679)
5680
5681xnnpack_unit_test(
5682 name = "f16_vsub_minmax_test",
5683 srcs = [
5684 "test/f16-vsub-minmax.cc",
5685 "test/vbinary-microkernel-tester.h",
5686 ] + MICROKERNEL_TEST_HDRS,
5687 deps = MICROKERNEL_TEST_DEPS,
5688)
5689
5690xnnpack_unit_test(
5691 name = "f16_vsubc_minmax_test",
5692 srcs = [
5693 "test/f16-vsubc-minmax.cc",
5694 "test/vbinaryc-microkernel-tester.h",
5695 ] + MICROKERNEL_TEST_HDRS,
5696 deps = MICROKERNEL_TEST_DEPS,
5697)
5698
5699xnnpack_unit_test(
5700 name = "f16_vrsubc_minmax_test",
5701 srcs = [
5702 "test/f16-vrsubc-minmax.cc",
5703 "test/vbinaryc-microkernel-tester.h",
5704 ] + MICROKERNEL_TEST_HDRS,
5705 deps = MICROKERNEL_TEST_DEPS,
5706)
5707
5708xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005709 name = "f32_vabs_test",
5710 srcs = [
5711 "test/f32-vabs.cc",
5712 "test/vunary-microkernel-tester.h",
5713 ] + MICROKERNEL_TEST_HDRS,
5714 deps = MICROKERNEL_TEST_DEPS,
5715)
5716
5717xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005718 name = "f32_vadd_test",
5719 srcs = [
5720 "test/f32-vadd.cc",
5721 "test/vbinary-microkernel-tester.h",
5722 ] + MICROKERNEL_TEST_HDRS,
5723 deps = MICROKERNEL_TEST_DEPS,
5724)
5725
5726xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005727 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005728 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005729 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005730 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005731 ] + MICROKERNEL_TEST_HDRS,
5732 deps = MICROKERNEL_TEST_DEPS,
5733)
5734
5735xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005736 name = "f32_vadd_relu_test",
5737 srcs = [
5738 "test/f32-vadd-relu.cc",
5739 "test/vbinary-microkernel-tester.h",
5740 ] + MICROKERNEL_TEST_HDRS,
5741 deps = MICROKERNEL_TEST_DEPS,
5742)
5743
5744xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005745 name = "f32_vaddc_test",
5746 srcs = [
5747 "test/f32-vaddc.cc",
5748 "test/vbinaryc-microkernel-tester.h",
5749 ] + MICROKERNEL_TEST_HDRS,
5750 deps = MICROKERNEL_TEST_DEPS,
5751)
5752
5753xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005754 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005755 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005756 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005757 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005758 ] + MICROKERNEL_TEST_HDRS,
5759 deps = MICROKERNEL_TEST_DEPS,
5760)
5761
5762xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005763 name = "f32_vaddc_relu_test",
5764 srcs = [
5765 "test/f32-vaddc-relu.cc",
5766 "test/vbinaryc-microkernel-tester.h",
5767 ] + MICROKERNEL_TEST_HDRS,
5768 deps = MICROKERNEL_TEST_DEPS,
5769)
5770
5771xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005772 name = "f32_vdiv_test",
5773 srcs = [
5774 "test/f32-vdiv.cc",
5775 "test/vbinary-microkernel-tester.h",
5776 ] + MICROKERNEL_TEST_HDRS,
5777 deps = MICROKERNEL_TEST_DEPS,
5778)
5779
5780xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005781 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005782 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005783 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005784 "test/vbinary-microkernel-tester.h",
5785 ] + MICROKERNEL_TEST_HDRS,
5786 deps = MICROKERNEL_TEST_DEPS,
5787)
5788
5789xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005790 name = "f32_vdiv_relu_test",
5791 srcs = [
5792 "test/f32-vdiv-relu.cc",
5793 "test/vbinary-microkernel-tester.h",
5794 ] + MICROKERNEL_TEST_HDRS,
5795 deps = MICROKERNEL_TEST_DEPS,
5796)
5797
5798xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005799 name = "f32_vdivc_test",
5800 srcs = [
5801 "test/f32-vdivc.cc",
5802 "test/vbinaryc-microkernel-tester.h",
5803 ] + MICROKERNEL_TEST_HDRS,
5804 deps = MICROKERNEL_TEST_DEPS,
5805)
5806
5807xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005808 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005809 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005810 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005811 "test/vbinaryc-microkernel-tester.h",
5812 ] + MICROKERNEL_TEST_HDRS,
5813 deps = MICROKERNEL_TEST_DEPS,
5814)
5815
5816xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005817 name = "f32_vdivc_relu_test",
5818 srcs = [
5819 "test/f32-vdivc-relu.cc",
5820 "test/vbinaryc-microkernel-tester.h",
5821 ] + MICROKERNEL_TEST_HDRS,
5822 deps = MICROKERNEL_TEST_DEPS,
5823)
5824
5825xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005826 name = "f32_vrdivc_test",
5827 srcs = [
5828 "test/f32-vrdivc.cc",
5829 "test/vbinaryc-microkernel-tester.h",
5830 ] + MICROKERNEL_TEST_HDRS,
5831 deps = MICROKERNEL_TEST_DEPS,
5832)
5833
5834xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005835 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005836 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005837 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005838 "test/vbinaryc-microkernel-tester.h",
5839 ] + MICROKERNEL_TEST_HDRS,
5840 deps = MICROKERNEL_TEST_DEPS,
5841)
5842
5843xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005844 name = "f32_vrdivc_relu_test",
5845 srcs = [
5846 "test/f32-vrdivc-relu.cc",
5847 "test/vbinaryc-microkernel-tester.h",
5848 ] + MICROKERNEL_TEST_HDRS,
5849 deps = MICROKERNEL_TEST_DEPS,
5850)
5851
5852xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005853 name = "f32_vmax_test",
5854 srcs = [
5855 "test/f32-vmax.cc",
5856 "test/vbinary-microkernel-tester.h",
5857 ] + MICROKERNEL_TEST_HDRS,
5858 deps = MICROKERNEL_TEST_DEPS,
5859)
5860
5861xnnpack_unit_test(
5862 name = "f32_vmaxc_test",
5863 srcs = [
5864 "test/f32-vmaxc.cc",
5865 "test/vbinaryc-microkernel-tester.h",
5866 ] + MICROKERNEL_TEST_HDRS,
5867 deps = MICROKERNEL_TEST_DEPS,
5868)
5869
5870xnnpack_unit_test(
5871 name = "f32_vmin_test",
5872 srcs = [
5873 "test/f32-vmin.cc",
5874 "test/vbinary-microkernel-tester.h",
5875 ] + MICROKERNEL_TEST_HDRS,
5876 deps = MICROKERNEL_TEST_DEPS,
5877)
5878
5879xnnpack_unit_test(
5880 name = "f32_vminc_test",
5881 srcs = [
5882 "test/f32-vminc.cc",
5883 "test/vbinaryc-microkernel-tester.h",
5884 ] + MICROKERNEL_TEST_HDRS,
5885 deps = MICROKERNEL_TEST_DEPS,
5886)
5887
5888xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005889 name = "f32_vmul_test",
5890 srcs = [
5891 "test/f32-vmul.cc",
5892 "test/vbinary-microkernel-tester.h",
5893 ] + MICROKERNEL_TEST_HDRS,
5894 deps = MICROKERNEL_TEST_DEPS,
5895)
5896
5897xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005898 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005899 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005900 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005901 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005902 ] + MICROKERNEL_TEST_HDRS,
5903 deps = MICROKERNEL_TEST_DEPS,
5904)
5905
5906xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005907 name = "f32_vmul_relu_test",
5908 srcs = [
5909 "test/f32-vmul-relu.cc",
5910 "test/vbinary-microkernel-tester.h",
5911 ] + MICROKERNEL_TEST_HDRS,
5912 deps = MICROKERNEL_TEST_DEPS,
5913)
5914
5915xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005916 name = "f32_vmulc_test",
5917 srcs = [
5918 "test/f32-vmulc.cc",
5919 "test/vbinaryc-microkernel-tester.h",
5920 ] + MICROKERNEL_TEST_HDRS,
5921 deps = MICROKERNEL_TEST_DEPS,
5922)
5923
5924xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005925 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005926 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005927 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005928 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005929 ] + MICROKERNEL_TEST_HDRS,
5930 deps = MICROKERNEL_TEST_DEPS,
5931)
5932
5933xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005934 name = "f32_vmulc_relu_test",
5935 srcs = [
5936 "test/f32-vmulc-relu.cc",
5937 "test/vbinaryc-microkernel-tester.h",
5938 ] + MICROKERNEL_TEST_HDRS,
5939 deps = MICROKERNEL_TEST_DEPS,
5940)
5941
5942xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005943 name = "f16_vmulcaddc_minmax_test",
5944 srcs = [
5945 "test/f16-vmulcaddc-minmax.cc",
5946 "test/vmulcaddc-microkernel-tester.h",
5947 "src/xnnpack/AlignedAllocator.h",
5948 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005949 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005950)
5951
5952xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005953 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005954 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005955 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005956 "test/vmulcaddc-microkernel-tester.h",
5957 "src/xnnpack/AlignedAllocator.h",
5958 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005959 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005960)
5961
5962xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005963 name = "f32_vlrelu_test",
5964 srcs = [
5965 "test/f32-vlrelu.cc",
5966 "test/vunary-microkernel-tester.h",
5967 ] + MICROKERNEL_TEST_HDRS,
5968 deps = MICROKERNEL_TEST_DEPS,
5969)
5970
5971xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005972 name = "f32_vneg_test",
5973 srcs = [
5974 "test/f32-vneg.cc",
5975 "test/vunary-microkernel-tester.h",
5976 ] + MICROKERNEL_TEST_HDRS,
5977 deps = MICROKERNEL_TEST_DEPS,
5978)
5979
5980xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005981 name = "f32_vrndne_test",
5982 srcs = [
5983 "test/f32-vrndne.cc",
5984 "test/vunary-microkernel-tester.h",
5985 ] + MICROKERNEL_TEST_HDRS,
5986 deps = MICROKERNEL_TEST_DEPS,
5987)
5988
5989xnnpack_unit_test(
5990 name = "f32_vrndz_test",
5991 srcs = [
5992 "test/f32-vrndz.cc",
5993 "test/vunary-microkernel-tester.h",
5994 ] + MICROKERNEL_TEST_HDRS,
5995 deps = MICROKERNEL_TEST_DEPS,
5996)
5997
5998xnnpack_unit_test(
5999 name = "f32_vrndu_test",
6000 srcs = [
6001 "test/f32-vrndu.cc",
6002 "test/vunary-microkernel-tester.h",
6003 ] + MICROKERNEL_TEST_HDRS,
6004 deps = MICROKERNEL_TEST_DEPS,
6005)
6006
6007xnnpack_unit_test(
6008 name = "f32_vrndd_test",
6009 srcs = [
6010 "test/f32-vrndd.cc",
6011 "test/vunary-microkernel-tester.h",
6012 ] + MICROKERNEL_TEST_HDRS,
6013 deps = MICROKERNEL_TEST_DEPS,
6014)
6015
6016xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07006017 name = "f32_vscale_test",
6018 srcs = [
6019 "test/f32-vscale.cc",
6020 "test/vscale-microkernel-tester.h",
6021 ] + MICROKERNEL_TEST_HDRS,
6022 deps = MICROKERNEL_TEST_DEPS,
6023)
6024
6025xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006026 name = "f32_vscaleexpminusmax_test",
6027 srcs = [
6028 "test/f32-vscaleexpminusmax.cc",
6029 "test/vscaleexpminusmax-microkernel-tester.h",
6030 ] + MICROKERNEL_TEST_HDRS,
6031 deps = MICROKERNEL_TEST_DEPS,
6032)
6033
6034xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006035 name = "f32_vscaleextexp_test",
6036 srcs = [
6037 "test/f32-vscaleextexp.cc",
6038 "test/vscaleextexp-microkernel-tester.h",
6039 ] + MICROKERNEL_TEST_HDRS,
6040 deps = MICROKERNEL_TEST_DEPS,
6041)
6042
6043xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006044 name = "f32_vsqr_test",
6045 srcs = [
6046 "test/f32-vsqr.cc",
6047 "test/vunary-microkernel-tester.h",
6048 ] + MICROKERNEL_TEST_HDRS,
6049 deps = MICROKERNEL_TEST_DEPS,
6050)
6051
6052xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006053 name = "f32_vsqrdiff_test",
6054 srcs = [
6055 "test/f32-vsqrdiff.cc",
6056 "test/vbinary-microkernel-tester.h",
6057 ] + MICROKERNEL_TEST_HDRS,
6058 deps = MICROKERNEL_TEST_DEPS,
6059)
6060
6061xnnpack_unit_test(
6062 name = "f32_vsqrdiffc_test",
6063 srcs = [
6064 "test/f32-vsqrdiffc.cc",
6065 "test/vbinaryc-microkernel-tester.h",
6066 ] + MICROKERNEL_TEST_HDRS,
6067 deps = MICROKERNEL_TEST_DEPS,
6068)
6069
6070xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006071 name = "f32_vsqrt_test",
6072 srcs = [
6073 "test/f32-vsqrt.cc",
6074 "test/vunary-microkernel-tester.h",
6075 ] + MICROKERNEL_TEST_HDRS,
6076 deps = MICROKERNEL_TEST_DEPS,
6077)
6078
6079xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006080 name = "f32_vsub_test",
6081 srcs = [
6082 "test/f32-vsub.cc",
6083 "test/vbinary-microkernel-tester.h",
6084 ] + MICROKERNEL_TEST_HDRS,
6085 deps = MICROKERNEL_TEST_DEPS,
6086)
6087
6088xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006089 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006090 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006091 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006092 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006093 ] + MICROKERNEL_TEST_HDRS,
6094 deps = MICROKERNEL_TEST_DEPS,
6095)
6096
6097xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006098 name = "f32_vsub_relu_test",
6099 srcs = [
6100 "test/f32-vsub-relu.cc",
6101 "test/vbinary-microkernel-tester.h",
6102 ] + MICROKERNEL_TEST_HDRS,
6103 deps = MICROKERNEL_TEST_DEPS,
6104)
6105
6106xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006107 name = "f32_vsubc_test",
6108 srcs = [
6109 "test/f32-vsubc.cc",
6110 "test/vbinaryc-microkernel-tester.h",
6111 ] + MICROKERNEL_TEST_HDRS,
6112 deps = MICROKERNEL_TEST_DEPS,
6113)
6114
6115xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006116 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006117 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006118 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006119 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006120 ] + MICROKERNEL_TEST_HDRS,
6121 deps = MICROKERNEL_TEST_DEPS,
6122)
6123
6124xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006125 name = "f32_vsubc_relu_test",
6126 srcs = [
6127 "test/f32-vsubc-relu.cc",
6128 "test/vbinaryc-microkernel-tester.h",
6129 ] + MICROKERNEL_TEST_HDRS,
6130 deps = MICROKERNEL_TEST_DEPS,
6131)
6132
6133xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006134 name = "f32_vrsubc_test",
6135 srcs = [
6136 "test/f32-vrsubc.cc",
6137 "test/vbinaryc-microkernel-tester.h",
6138 ] + MICROKERNEL_TEST_HDRS,
6139 deps = MICROKERNEL_TEST_DEPS,
6140)
6141
6142xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006143 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006144 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006145 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006146 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006147 ] + MICROKERNEL_TEST_HDRS,
6148 deps = MICROKERNEL_TEST_DEPS,
6149)
6150
6151xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006152 name = "f32_vrsubc_relu_test",
6153 srcs = [
6154 "test/f32-vrsubc-relu.cc",
6155 "test/vbinaryc-microkernel-tester.h",
6156 ] + MICROKERNEL_TEST_HDRS,
6157 deps = MICROKERNEL_TEST_DEPS,
6158)
6159
6160xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006161 name = "qs8_dwconv_minmax_test",
6162 srcs = [
6163 "test/qs8-dwconv-minmax.cc",
6164 "test/dwconv-microkernel-tester.h",
6165 "src/xnnpack/AlignedAllocator.h",
6166 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6167 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6168)
6169
6170xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006171 name = "qs8_gavgpool_minmax_test",
6172 srcs = [
6173 "test/qs8-gavgpool-minmax.cc",
6174 "test/gavgpool-microkernel-tester.h",
6175 "src/xnnpack/AlignedAllocator.h",
6176 ] + MICROKERNEL_TEST_HDRS,
6177 deps = MICROKERNEL_TEST_DEPS,
6178)
6179
6180xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006181 name = "qs8_gemm_minmax_test",
6182 srcs = [
6183 "test/qs8-gemm-minmax.cc",
6184 "test/gemm-microkernel-tester.h",
6185 "src/xnnpack/AlignedAllocator.h",
6186 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6187 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6188)
6189
6190xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006191 name = "qs8_igemm_minmax_test",
6192 srcs = [
6193 "test/qs8-igemm-minmax.cc",
6194 "test/gemm-microkernel-tester.h",
6195 "src/xnnpack/AlignedAllocator.h",
6196 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6197 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6198)
6199
6200xnnpack_unit_test(
6201 name = "qs8_requantization_test",
6202 srcs = [
6203 "src/xnnpack/requantization-stubs.h",
6204 "test/qs8-requantization.cc",
6205 "test/requantization-tester.h",
6206 ] + MICROKERNEL_TEST_HDRS,
6207 deps = MICROKERNEL_TEST_DEPS,
6208)
6209
6210xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006211 name = "qs8_vadd_minmax_test",
6212 srcs = [
6213 "test/qs8-vadd-minmax.cc",
6214 "test/vadd-microkernel-tester.h",
6215 ] + MICROKERNEL_TEST_HDRS,
6216 deps = MICROKERNEL_TEST_DEPS,
6217)
6218
6219xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006220 name = "qs8_vaddc_minmax_test",
6221 srcs = [
6222 "test/qs8-vaddc-minmax.cc",
6223 "test/vaddc-microkernel-tester.h",
6224 ] + MICROKERNEL_TEST_HDRS,
6225 deps = MICROKERNEL_TEST_DEPS,
6226)
6227
6228xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006229 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006230 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006231 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006232 "test/avgpool-microkernel-tester.h",
6233 "src/xnnpack/AlignedAllocator.h",
6234 ] + MICROKERNEL_TEST_HDRS,
6235 deps = MICROKERNEL_TEST_DEPS,
6236)
6237
6238xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006239 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006240 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006241 "test/qu8-dwconv-minmax.cc",
6242 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006243 "src/xnnpack/AlignedAllocator.h",
6244 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006245 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006246)
6247
6248xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006249 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006250 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006251 "test/qu8-igemm-minmax.cc",
6252 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006253 "src/xnnpack/AlignedAllocator.h",
6254 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006255 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006256)
6257
6258xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006259 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006260 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006261 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006262 "test/gavgpool-microkernel-tester.h",
6263 "src/xnnpack/AlignedAllocator.h",
6264 ] + MICROKERNEL_TEST_HDRS,
6265 deps = MICROKERNEL_TEST_DEPS,
6266)
6267
6268xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006269 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006270 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006271 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006272 "test/gemm-microkernel-tester.h",
6273 "src/xnnpack/AlignedAllocator.h",
6274 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006275 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006276)
6277
6278xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006279 name = "qu8_requantization_test",
6280 srcs = [
6281 "src/xnnpack/requantization-stubs.h",
6282 "test/qu8-requantization.cc",
6283 "test/requantization-tester.h",
6284 ] + MICROKERNEL_TEST_HDRS,
6285 deps = MICROKERNEL_TEST_DEPS,
6286)
6287
6288xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006289 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006290 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006291 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006292 "test/vadd-microkernel-tester.h",
6293 ] + MICROKERNEL_TEST_HDRS,
6294 deps = MICROKERNEL_TEST_DEPS,
6295)
6296
6297xnnpack_unit_test(
6298 name = "u8_clamp_test",
6299 srcs = [
6300 "test/u8-clamp.cc",
6301 "test/clamp-microkernel-tester.h",
6302 ] + MICROKERNEL_TEST_HDRS,
6303 deps = MICROKERNEL_TEST_DEPS,
6304)
6305
6306xnnpack_unit_test(
6307 name = "u8_lut32norm_test",
6308 srcs = [
6309 "test/u8-lut32norm.cc",
6310 "test/lut-norm-microkernel-tester.h",
6311 ] + MICROKERNEL_TEST_HDRS,
6312 deps = MICROKERNEL_TEST_DEPS,
6313)
6314
6315xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006316 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006317 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006318 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006319 "test/maxpool-microkernel-tester.h",
6320 ] + MICROKERNEL_TEST_HDRS,
6321 deps = MICROKERNEL_TEST_DEPS,
6322)
6323
6324xnnpack_unit_test(
6325 name = "u8_rmax_test",
6326 srcs = [
6327 "test/u8-rmax.cc",
6328 "test/rmax-microkernel-tester.h",
6329 ] + MICROKERNEL_TEST_HDRS,
6330 deps = MICROKERNEL_TEST_DEPS,
6331)
6332
6333xnnpack_unit_test(
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006334 name = "x32_depthtospace2d_chw2hwc_test",
Yury Kartynnike7841862020-11-04 18:22:18 -08006335 srcs = [
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006336 "test/x32-depthtospace2d-chw2hwc.cc",
6337 "test/depthtospace-microkernel-tester.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08006338 ] + MICROKERNEL_TEST_HDRS,
6339 deps = MICROKERNEL_TEST_DEPS,
6340)
6341
6342xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006343 name = "x32_fill_test",
6344 srcs = [
6345 "test/x32-fill.cc",
6346 "test/fill-microkernel-tester.h",
6347 ] + MICROKERNEL_TEST_HDRS,
6348 deps = MICROKERNEL_TEST_DEPS,
6349)
6350
6351xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006352 name = "x32_packx_test",
6353 srcs = [
6354 "test/x32-packx.cc",
6355 "test/pack-microkernel-tester.h",
6356 "src/xnnpack/AlignedAllocator.h",
6357 ] + MICROKERNEL_TEST_HDRS,
6358 deps = MICROKERNEL_TEST_DEPS,
6359)
6360
6361xnnpack_unit_test(
6362 name = "x32_pad_test",
6363 srcs = [
6364 "test/x32-pad.cc",
6365 "test/pad-microkernel-tester.h",
6366 ] + MICROKERNEL_TEST_HDRS,
6367 deps = MICROKERNEL_TEST_DEPS,
6368)
6369
6370xnnpack_unit_test(
6371 name = "x32_unpool_test",
6372 srcs = [
6373 "test/x32-unpool.cc",
6374 "test/unpool-microkernel-tester.h",
6375 ] + MICROKERNEL_TEST_HDRS,
6376 deps = MICROKERNEL_TEST_DEPS,
6377)
6378
6379xnnpack_unit_test(
6380 name = "x32_zip_test",
6381 srcs = [
6382 "test/x32-zip.cc",
6383 "test/zip-microkernel-tester.h",
6384 ] + MICROKERNEL_TEST_HDRS,
6385 deps = MICROKERNEL_TEST_DEPS,
6386)
6387
6388xnnpack_unit_test(
6389 name = "x8_lut_test",
6390 srcs = [
6391 "test/x8-lut.cc",
6392 "test/lut-microkernel-tester.h",
6393 ] + MICROKERNEL_TEST_HDRS,
6394 deps = MICROKERNEL_TEST_DEPS,
6395)
6396
6397xnnpack_unit_test(
6398 name = "x8_zip_test",
6399 srcs = [
6400 "test/x8-zip.cc",
6401 "test/zip-microkernel-tester.h",
6402 ] + MICROKERNEL_TEST_HDRS,
6403 deps = MICROKERNEL_TEST_DEPS,
6404)
6405
Marat Dukhan20c3b922020-03-10 03:45:06 -07006406########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006407
6408xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006409 name = "operator_size_test",
6410 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006411 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006412)
6413
Marat Dukhan20c3b922020-03-10 03:45:06 -07006414xnnpack_binary(
6415 name = "subgraph_size_test",
6416 srcs = ["test/subgraph-size.c"],
6417 deps = [":XNNPACK"],
6418)
6419
6420########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006421
6422xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006423 name = "abs_nc_test",
6424 srcs = [
6425 "test/abs-nc.cc",
6426 "test/abs-operator-tester.h",
6427 ],
6428 deps = OPERATOR_TEST_DEPS,
6429)
6430
6431xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006432 name = "add_nd_test",
6433 srcs = [
6434 "test/add-nd.cc",
6435 "test/binary-elementwise-operator-tester.h",
6436 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006437 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006438)
6439
6440xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006441 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006442 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006443 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006444 "test/argmax-pooling-operator-tester.h",
6445 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006446 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006447)
6448
6449xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006450 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006451 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006452 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006453 "test/average-pooling-operator-tester.h",
6454 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006455 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006456)
6457
6458xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006459 name = "bankers_rounding_nc_test",
6460 srcs = [
6461 "test/bankers-rounding-nc.cc",
6462 "test/bankers-rounding-operator-tester.h",
6463 ],
6464 deps = OPERATOR_TEST_DEPS,
6465)
6466
6467xnnpack_unit_test(
6468 name = "ceiling_nc_test",
6469 srcs = [
6470 "test/ceiling-nc.cc",
6471 "test/ceiling-operator-tester.h",
6472 ],
6473 deps = OPERATOR_TEST_DEPS,
6474)
6475
6476xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006477 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006478 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006479 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006480 "test/channel-shuffle-operator-tester.h",
6481 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006482 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006483)
6484
6485xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006486 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006487 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006488 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006489 "test/clamp-operator-tester.h",
6490 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006491 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006492)
6493
6494xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006495 name = "constant_pad_nd_test",
6496 srcs = [
6497 "test/constant-pad-nd.cc",
6498 "test/constant-pad-operator-tester.h",
6499 ],
6500 deps = OPERATOR_TEST_DEPS,
6501)
6502
6503xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006504 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006505 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006506 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006507 "test/convolution-operator-tester.h",
6508 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006509 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006510)
6511
6512xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006513 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006514 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006515 "test/convolution-nchw.cc",
6516 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006517 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006518 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006519)
6520
6521xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006522 name = "copy_nc_test",
6523 srcs = [
6524 "test/copy-nc.cc",
6525 "test/copy-operator-tester.h",
6526 ],
6527 deps = OPERATOR_TEST_DEPS,
6528)
6529
6530xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006531 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006532 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006533 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006534 "test/deconvolution-operator-tester.h",
6535 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006536 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006537)
6538
6539xnnpack_unit_test(
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08006540 name = "depth_to_space_test",
6541 srcs = [
6542 "test/depth-to-space.cc",
6543 "test/depth-to-space-operator-tester.h",
6544 ] + OPERATOR_TEST_PARAMS_HDRS,
6545 deps = OPERATOR_TEST_DEPS,
6546)
6547
6548xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006549 name = "divide_nd_test",
6550 srcs = [
6551 "test/binary-elementwise-operator-tester.h",
6552 "test/divide-nd.cc",
6553 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006554 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006555)
6556
6557xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006558 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006559 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006560 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006561 "test/fully-connected-operator-tester.h",
6562 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006563 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006564)
6565
6566xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006567 name = "floor_nc_test",
6568 srcs = [
6569 "test/floor-nc.cc",
6570 "test/floor-operator-tester.h",
6571 ],
6572 deps = OPERATOR_TEST_DEPS,
6573)
6574
6575xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006576 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006577 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006578 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006579 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006580 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006581 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006582)
6583
6584xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006585 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006586 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006587 "test/global-average-pooling-ncw.cc",
6588 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006589 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006590 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006591)
6592
6593xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006594 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006595 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006596 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006597 "test/hardswish-operator-tester.h",
6598 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006599 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006600)
6601
6602xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006603 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006604 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006605 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006606 "test/leaky-relu-operator-tester.h",
6607 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006608 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006609)
6610
6611xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006612 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006613 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006614 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006615 "test/max-pooling-operator-tester.h",
6616 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006617 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006618)
6619
6620xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006621 name = "maximum_nd_test",
6622 srcs = [
6623 "test/binary-elementwise-operator-tester.h",
6624 "test/maximum-nd.cc",
6625 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006626 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006627)
6628
6629xnnpack_unit_test(
6630 name = "minimum_nd_test",
6631 srcs = [
6632 "test/binary-elementwise-operator-tester.h",
6633 "test/minimum-nd.cc",
6634 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006635 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006636)
6637
6638xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006639 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006640 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006641 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006642 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006643 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006644 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006645)
6646
6647xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006648 name = "negate_nc_test",
6649 srcs = [
6650 "test/negate-nc.cc",
6651 "test/negate-operator-tester.h",
6652 ],
6653 deps = OPERATOR_TEST_DEPS,
6654)
6655
6656xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006657 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006658 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006659 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006660 "test/prelu-operator-tester.h",
6661 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006662 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006663)
6664
6665xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006666 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006667 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006668 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006669 "test/resize-bilinear-operator-tester.h",
6670 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006671 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006672)
6673
6674xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07006675 name = "resize_bilinear_nchw_test",
6676 srcs = [
6677 "test/resize-bilinear-nchw.cc",
6678 "test/resize-bilinear-operator-tester.h",
6679 ] + OPERATOR_TEST_PARAMS_HDRS,
6680 deps = OPERATOR_TEST_DEPS,
6681)
6682
6683xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006684 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006685 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006686 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006687 "test/sigmoid-operator-tester.h",
6688 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006689 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006690)
6691
6692xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006693 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006694 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006695 "test/softmax-nc.cc",
6696 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006697 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006698 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006699)
6700
6701xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006702 name = "square_nc_test",
6703 srcs = [
6704 "test/square-nc.cc",
6705 "test/square-operator-tester.h",
6706 ],
6707 deps = OPERATOR_TEST_DEPS,
6708)
6709
6710xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006711 name = "square_root_nc_test",
6712 srcs = [
6713 "test/square-root-nc.cc",
6714 "test/square-root-operator-tester.h",
6715 ],
6716 deps = OPERATOR_TEST_DEPS,
6717)
6718
6719xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006720 name = "squared_difference_nd_test",
6721 srcs = [
6722 "test/binary-elementwise-operator-tester.h",
6723 "test/squared-difference-nd.cc",
6724 ],
6725 deps = OPERATOR_TEST_DEPS,
6726)
6727
6728xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006729 name = "subtract_nd_test",
6730 srcs = [
6731 "test/binary-elementwise-operator-tester.h",
6732 "test/subtract-nd.cc",
6733 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006734 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006735)
6736
6737xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006738 name = "truncation_nc_test",
6739 srcs = [
6740 "test/truncation-nc.cc",
6741 "test/truncation-operator-tester.h",
6742 ],
6743 deps = OPERATOR_TEST_DEPS,
6744)
6745
6746xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006747 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006748 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006749 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006750 "test/unpooling-operator-tester.h",
6751 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006752 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006753)
6754
Chao Mei6ddfc602020-05-13 22:29:36 -07006755############################### Misc unit tests ###############################
6756
6757xnnpack_unit_test(
6758 name = "memory_planner_test",
6759 srcs = [
6760 "test/memory-planner-test.cc",
6761 ],
6762 deps = [
6763 ":XNNPACK",
6764 ":memory_planner",
6765 ],
6766)
6767
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07006768xnnpack_unit_test(
6769 name = "subgraph_nchw_test",
6770 srcs = [
6771 "src/xnnpack/subgraph.h",
6772 "test/subgraph-nchw.cc",
6773 "test/subgraph-tester.h",
6774 ],
6775 deps = [
6776 ":XNNPACK",
6777 ],
6778)
6779
Marat Dukhan08c4a432019-10-03 09:29:21 -07006780############################# Build configurations #############################
6781
Marat Dukhanb8642352019-10-30 15:43:02 -07006782# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006783config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006784 name = "xnn_enable_assembly_explicit_true",
6785 define_values = {"xnn_enable_assembly": "true"},
6786)
6787
6788# Disables usage of assembly kernels.
6789config_setting(
6790 name = "xnn_enable_assembly_explicit_false",
6791 define_values = {"xnn_enable_assembly": "false"},
6792)
6793
Marat Dukhan9de90e02020-06-18 16:04:12 -07006794# Enables usage of sparse inference.
6795config_setting(
6796 name = "xnn_enable_sparse_explicit_true",
6797 define_values = {"xnn_enable_sparse": "true"},
6798)
6799
6800# Disables usage of sparse inference.
6801config_setting(
6802 name = "xnn_enable_sparse_explicit_false",
6803 define_values = {"xnn_enable_sparse": "false"},
6804)
6805
Marat Dukhan05702cf2020-03-26 15:41:33 -07006806# Disables usage of HMP-aware optimizations.
6807config_setting(
6808 name = "xnn_enable_hmp_explicit_false",
6809 define_values = {"xnn_enable_hmp": "false"},
6810)
6811
Chao Mei6ddfc602020-05-13 22:29:36 -07006812# Enable usage of optimized memory allocation
6813config_setting(
6814 name = "xnn_enable_memopt_explicit_true",
6815 define_values = {"xnnpack_enable_memopt": "true"},
6816)
6817
6818# Disable usage of optimized memory allocation
6819config_setting(
6820 name = "xnn_enable_memopt_explicit_false",
6821 define_values = {"xnnpack_enable_memopt": "false"},
6822)
6823
Marat Dukhanb8642352019-10-30 15:43:02 -07006824# Builds with -c dbg
6825config_setting(
6826 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006827 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006828 "compilation_mode": "dbg",
6829 },
6830)
6831
6832# Builds with -c opt
6833config_setting(
6834 name = "optimized_build",
6835 values = {
6836 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006837 },
6838)
6839
6840config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006841 name = "linux_k8",
6842 values = {"cpu": "k8"},
6843)
6844
6845config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006846 name = "linux_arm",
6847 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006848)
6849
6850config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006851 name = "linux_armeabi",
6852 values = {"cpu": "armeabi"},
6853)
6854
6855config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006856 name = "linux_armhf",
6857 values = {"cpu": "armhf"},
6858)
6859
6860config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006861 name = "linux_armv7a",
6862 values = {"cpu": "armv7a"},
6863)
6864
6865config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006866 name = "linux_aarch64",
6867 values = {"cpu": "aarch64"},
6868)
6869
6870config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006871 name = "android",
6872 values = {"crosstool_top": "//external:android/crosstool"},
6873)
6874
6875config_setting(
6876 name = "android_armv7",
6877 values = {
6878 "crosstool_top": "//external:android/crosstool",
6879 "cpu": "armeabi-v7a",
6880 },
6881)
6882
6883config_setting(
6884 name = "android_arm64",
6885 values = {
6886 "crosstool_top": "//external:android/crosstool",
6887 "cpu": "arm64-v8a",
6888 },
6889)
6890
6891config_setting(
6892 name = "android_x86",
6893 values = {
6894 "crosstool_top": "//external:android/crosstool",
6895 "cpu": "x86",
6896 },
6897)
6898
6899config_setting(
6900 name = "android_x86_64",
6901 values = {
6902 "crosstool_top": "//external:android/crosstool",
6903 "cpu": "x86_64",
6904 },
6905)
6906
6907config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006908 name = "windows_x86_64",
6909 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006910)
6911
6912config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006913 name = "windows_x86_64_clang",
6914 values = {
6915 "compiler": "clang-cl",
6916 "cpu": "x64_windows",
6917 },
6918)
6919
6920config_setting(
6921 name = "windows_x86_64_mingw",
6922 values = {
6923 "compiler": "mingw-gcc",
6924 "cpu": "x64_windows",
6925 },
6926)
6927
6928config_setting(
6929 name = "windows_x86_64_msys",
6930 values = {
6931 "compiler": "msys-gcc",
6932 "cpu": "x64_windows",
6933 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006934)
6935
6936config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006937 name = "macos_x86_64",
6938 values = {
6939 "apple_platform_type": "macos",
6940 "cpu": "darwin",
6941 },
6942)
6943
6944config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006945 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006946 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006947)
6948
6949config_setting(
6950 name = "emscripten_wasm",
6951 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006952 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006953 "cpu": "wasm",
6954 },
6955)
6956
6957config_setting(
6958 name = "emscripten_wasmsimd",
6959 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006960 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006961 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006962 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006963 },
6964)
6965
6966config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006967 name = "ios_armv7",
6968 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006969 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006970 "cpu": "ios_armv7",
6971 },
6972)
6973
6974config_setting(
6975 name = "ios_arm64",
6976 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006977 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006978 "cpu": "ios_arm64",
6979 },
6980)
6981
6982config_setting(
6983 name = "ios_arm64e",
6984 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006985 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006986 "cpu": "ios_arm64e",
6987 },
6988)
6989
6990config_setting(
6991 name = "ios_x86",
6992 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006993 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006994 "cpu": "ios_i386",
6995 },
6996)
6997
6998config_setting(
6999 name = "ios_x86_64",
7000 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007001 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007002 "cpu": "ios_x86_64",
7003 },
7004)
7005
7006config_setting(
7007 name = "watchos_armv7k",
7008 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007009 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007010 "cpu": "watchos_armv7k",
7011 },
7012)
7013
7014config_setting(
7015 name = "watchos_arm64_32",
7016 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007017 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007018 "cpu": "watchos_arm64_32",
7019 },
7020)
7021
7022config_setting(
7023 name = "watchos_x86",
7024 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007025 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007026 "cpu": "watchos_i386",
7027 },
7028)
7029
7030config_setting(
7031 name = "watchos_x86_64",
7032 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007033 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007034 "cpu": "watchos_x86_64",
7035 },
7036)
7037
7038config_setting(
7039 name = "tvos_arm64",
7040 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007041 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007042 "cpu": "tvos_arm64",
7043 },
7044)
7045
7046config_setting(
7047 name = "tvos_x86_64",
7048 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007049 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007050 "cpu": "tvos_x86_64",
7051 },
7052)