blob: f964ac6d385f6b39d8068b6a53ea22652629840d [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 Dukhanc60742b2020-11-23 12:33:27 -0800116 "src/tables/exp2minus-k-over-16.c",
Marat Dukhan1f256fc2020-09-24 21:27:14 -0700117 "src/tables/exp2minus-k-over-64.c",
118 "src/tables/exp2minus-k-over-2048.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800119]
120
Marat Dukhan08c4a432019-10-03 09:29:21 -0700121SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800122 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800123 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800124 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700125 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
126 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700127 "src/f32-clamp/gen/scalar-x1.c",
128 "src/f32-clamp/gen/scalar-x2.c",
129 "src/f32-clamp/gen/scalar-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700130 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700131 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700132 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700133 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700134 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
135 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
136 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700137 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700138 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
139 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
140 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700141 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700142 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
143 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
144 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700145 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700146 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
147 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
148 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700149 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700150 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
151 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
152 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700153 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700154 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
155 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
156 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700157 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c",
158 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c",
159 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700160 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700161 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700162 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c",
163 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c",
164 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c",
165 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c",
166 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700167 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c",
168 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c",
169 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700170 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700171 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700172 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c",
173 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c",
174 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700175 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c",
176 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c",
177 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c",
178 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700179 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700180 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c",
181 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700182 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700183 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700184 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c",
Marat Dukhan29c0c332020-10-28 22:11:00 -0700185 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c",
186 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c",
187 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c",
188 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c",
189 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c",
190 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c",
191 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c",
192 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c",
193 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c",
194 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700195 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700196 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
197 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700198 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
199 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
200 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700201 "src/f32-gemm/gen/1x4-minmax-scalar.c",
202 "src/f32-gemm/gen/1x4-relu-scalar.c",
203 "src/f32-gemm/gen/1x4-scalar.c",
204 "src/f32-gemm/gen/2x4-minmax-scalar.c",
205 "src/f32-gemm/gen/2x4-relu-scalar.c",
206 "src/f32-gemm/gen/2x4-scalar.c",
207 "src/f32-gemm/gen/4x2-minmax-scalar.c",
208 "src/f32-gemm/gen/4x2-relu-scalar.c",
209 "src/f32-gemm/gen/4x2-scalar.c",
210 "src/f32-gemm/gen/4x4-minmax-scalar.c",
211 "src/f32-gemm/gen/4x4-relu-scalar.c",
212 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700213 "src/f32-hswish/gen/hswish-scalar-x1.c",
214 "src/f32-hswish/gen/hswish-scalar-x2.c",
215 "src/f32-hswish/gen/hswish-scalar-x4.c",
XNNPACK Team21432672020-10-19 19:58:48 -0700216 "src/f32-ibilinear-chw/gen/scalar-p1.c",
217 "src/f32-ibilinear-chw/gen/scalar-p2.c",
218 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700219 "src/f32-ibilinear/gen/scalar-c1.c",
220 "src/f32-ibilinear/gen/scalar-c2.c",
221 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700222 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700223 "src/f32-igemm/gen/1x4-relu-scalar.c",
224 "src/f32-igemm/gen/1x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700225 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700226 "src/f32-igemm/gen/2x4-relu-scalar.c",
227 "src/f32-igemm/gen/2x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700228 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700229 "src/f32-igemm/gen/4x2-relu-scalar.c",
230 "src/f32-igemm/gen/4x2-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700231 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700232 "src/f32-igemm/gen/4x4-relu-scalar.c",
233 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700234 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
235 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
236 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700237 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
238 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
239 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
240 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800241 "src/f32-prelu/gen/scalar-2x1.c",
242 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800243 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800244 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700245 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800246 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
247 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700248 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800249 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800250 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700251 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800252 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
253 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700254 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700255 "src/f32-relu/gen/scalar-x1.c",
256 "src/f32-relu/gen/scalar-x2.c",
257 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700258 "src/f32-relu/gen/scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700259 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800260 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
261 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
262 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700263 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
264 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
265 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800266 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
267 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
268 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700269 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
270 "src/f32-spmm/gen/1x1-minmax-scalar.c",
271 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
272 "src/f32-spmm/gen/2x1-minmax-scalar.c",
273 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
274 "src/f32-spmm/gen/4x1-minmax-scalar.c",
275 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
276 "src/f32-spmm/gen/8x1-minmax-scalar.c",
277 "src/f32-spmm/gen/8x2-minmax-scalar.c",
278 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700279 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
280 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
281 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700282 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700283 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
284 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
285 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700286 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700287 "src/f32-vbinary/gen/vadd-scalar-x1.c",
288 "src/f32-vbinary/gen/vadd-scalar-x2.c",
289 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700290 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700291 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
292 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
293 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700294 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700295 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
296 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
297 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700298 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700299 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
300 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
301 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700302 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700303 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
304 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
305 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700306 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700307 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
308 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
309 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700310 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700311 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
312 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
313 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700314 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700315 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
316 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
317 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700318 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700319 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
320 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
321 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700322 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700323 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
324 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
325 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700326 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800327 "src/f32-vbinary/gen/vmax-scalar-x1.c",
328 "src/f32-vbinary/gen/vmax-scalar-x2.c",
329 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700330 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800331 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
332 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
333 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700334 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800335 "src/f32-vbinary/gen/vmin-scalar-x1.c",
336 "src/f32-vbinary/gen/vmin-scalar-x2.c",
337 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700338 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800339 "src/f32-vbinary/gen/vminc-scalar-x1.c",
340 "src/f32-vbinary/gen/vminc-scalar-x2.c",
341 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700342 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700343 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
344 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
345 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700346 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700347 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
348 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
349 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700350 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700351 "src/f32-vbinary/gen/vmul-scalar-x1.c",
352 "src/f32-vbinary/gen/vmul-scalar-x2.c",
353 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700354 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700355 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
356 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
357 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700358 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700359 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
360 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
361 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700362 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700363 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
364 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
365 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700366 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700367 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
368 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
369 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700370 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700371 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
372 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
373 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700374 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700375 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
376 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
377 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700378 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700379 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
380 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
381 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700382 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700383 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
384 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
385 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700386 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700387 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
388 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
389 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700390 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700391 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
392 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
393 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700394 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700395 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
396 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
397 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700398 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700399 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
400 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
401 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700402 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700403 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
404 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
405 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700406 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700407 "src/f32-vbinary/gen/vsub-scalar-x1.c",
408 "src/f32-vbinary/gen/vsub-scalar-x2.c",
409 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700410 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700411 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
412 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
413 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700414 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700415 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
416 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
417 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700418 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700419 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
420 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
421 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700422 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700423 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
424 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
425 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700426 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
427 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
428 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700429 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
430 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
431 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700432 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
433 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
434 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
435 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
436 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
437 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
438 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
439 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
440 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700441 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
442 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
443 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700444 "src/f32-vunary/gen/vabs-scalar-x1.c",
445 "src/f32-vunary/gen/vabs-scalar-x2.c",
446 "src/f32-vunary/gen/vabs-scalar-x4.c",
447 "src/f32-vunary/gen/vneg-scalar-x1.c",
448 "src/f32-vunary/gen/vneg-scalar-x2.c",
449 "src/f32-vunary/gen/vneg-scalar-x4.c",
450 "src/f32-vunary/gen/vsqr-scalar-x1.c",
451 "src/f32-vunary/gen/vsqr-scalar-x2.c",
452 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -0800453 "src/math/expminus-scalar-rr2-lut64-p2.c",
454 "src/math/expminus-scalar-rr2-lut2048-p1.c",
455 "src/math/expminus-scalar-rr2-p5.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800456 "src/math/expm1minus-scalar-rr2-lut16-p3.c",
457 "src/math/expm1minus-scalar-rr2-lut16-p4.c",
458 "src/math/expm1minus-scalar-rr2-p5.c",
459 "src/math/expm1minus-scalar-rr2-p6.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700460 "src/math/roundd-scalar-addsub.c",
461 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700462 "src/math/roundd-scalar-floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700463 "src/math/roundne-scalar-addsub.c",
464 "src/math/roundne-scalar-nearbyint.c",
465 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700466 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700467 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700468 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700469 "src/math/roundz-scalar-addsub.c",
470 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700471 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700472 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700473 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700474 "src/math/sigmoid-scalar-rr2-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700475 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
476 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700477 "src/qs8-requantization/fp32-scalar-lrintf.c",
478 "src/qs8-requantization/fp32-scalar-magic.c",
479 "src/qs8-requantization/precise-scalar-signed64.c",
480 "src/qs8-requantization/precise-scalar-unsigned32.c",
481 "src/qs8-requantization/precise-scalar-unsigned64.c",
482 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700483 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
484 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
485 "src/qu8-dwconv/up1x9-minmax-scalar.c",
486 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
487 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700488 "src/qu8-gemm/2x2-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700489 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
490 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700491 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700492 "src/qu8-requantization/fp32-scalar-lrintf.c",
493 "src/qu8-requantization/fp32-scalar-magic.c",
494 "src/qu8-requantization/precise-scalar-signed64.c",
495 "src/qu8-requantization/precise-scalar-unsigned32.c",
496 "src/qu8-requantization/precise-scalar-unsigned64.c",
497 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700498 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700499 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700500 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700501 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700502 "src/u8-rmax/scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700503 "src/x8-lut/scalar.c",
504 "src/x8-zip/x2-scalar.c",
505 "src/x8-zip/x3-scalar.c",
506 "src/x8-zip/x4-scalar.c",
507 "src/x8-zip/xm-scalar.c",
Marat Dukhanad71b9a2020-11-20 00:01:51 -0800508 "src/x32-depthtospace2d-chw2hwc/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700509 "src/x32-fill/scalar-float.c",
510 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700511 "src/x32-packx/x2-scalar.c",
512 "src/x32-packx/x3-scalar.c",
513 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700514 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700515 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700516 "src/x32-unpool/scalar.c",
517 "src/x32-zip/x2-scalar.c",
518 "src/x32-zip/x3-scalar.c",
519 "src/x32-zip/x4-scalar.c",
520 "src/x32-zip/xm-scalar.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700521]
522
Marat Dukhan436ebe62019-12-04 15:10:12 -0800523WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700524 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
525 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700526 "src/f32-clamp/gen/wasm-x1.c",
527 "src/f32-clamp/gen/wasm-x2.c",
528 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700529 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
530 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700531 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
532 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700533 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
534 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700535 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
536 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700537 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
538 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700539 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
540 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700541 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
542 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700543 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
544 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700545 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
546 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700547 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
548 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700549 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
550 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700551 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
552 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700553 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
554 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700555 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
556 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
557 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
558 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700559 "src/f32-gemm/gen/1x4-relu-wasm.c",
560 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700561 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700562 "src/f32-gemm/gen/2x4-relu-wasm.c",
563 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700564 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700565 "src/f32-gemm/gen/4x2-relu-wasm.c",
566 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700567 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700568 "src/f32-gemm/gen/4x4-relu-wasm.c",
569 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700570 "src/f32-hswish/gen/hswish-wasm-x1.c",
571 "src/f32-hswish/gen/hswish-wasm-x2.c",
572 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700573 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700574 "src/f32-igemm/gen/1x4-relu-wasm.c",
575 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700576 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700577 "src/f32-igemm/gen/2x4-relu-wasm.c",
578 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700579 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700580 "src/f32-igemm/gen/4x2-relu-wasm.c",
581 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700582 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700583 "src/f32-igemm/gen/4x4-relu-wasm.c",
584 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700585 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
586 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
587 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700588 "src/f32-prelu/gen/wasm-2x1.c",
589 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700590 "src/f32-relu/gen/wasm-x1.c",
591 "src/f32-relu/gen/wasm-x2.c",
592 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700593 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700594 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
595 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
596 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700597 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700598 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
599 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
600 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700601 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700602 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
603 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
604 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
605 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700606 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
607 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
608 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700609 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700610 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
611 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
612 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
613 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700614 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
615 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
616 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700617 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700618 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
619 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
620 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
621 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700622 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
623 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
624 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700625 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800626 "src/f32-vbinary/gen/vmax-wasm-x1.c",
627 "src/f32-vbinary/gen/vmax-wasm-x2.c",
628 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700629 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800630 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
631 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
632 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700633 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800634 "src/f32-vbinary/gen/vmin-wasm-x1.c",
635 "src/f32-vbinary/gen/vmin-wasm-x2.c",
636 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700637 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800638 "src/f32-vbinary/gen/vminc-wasm-x1.c",
639 "src/f32-vbinary/gen/vminc-wasm-x2.c",
640 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700641 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700642 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
643 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
644 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700645 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700646 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
647 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
648 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700649 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700650 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
651 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
652 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
653 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700654 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
655 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
656 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700657 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700658 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
659 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
660 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
661 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700662 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
663 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
664 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700665 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700666 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
667 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
668 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
669 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700670 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
671 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
672 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700673 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700674 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
675 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
676 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
677 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700678 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
679 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
680 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700681 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700682 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
683 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
684 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
685 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700686 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
687 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
688 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700689 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700690 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
691 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
692 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700693 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
694 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
695 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800696]
697
Marat Dukhan290055c2020-06-09 12:24:29 -0700698WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700699 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
700 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
701 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700702 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
703 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
704 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
705 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700706 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
707 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
708 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
709 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700710 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700711 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700712 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700713 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700714 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
715 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700716 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700717 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700718 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700719 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
720 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
721 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
722 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
723 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
724 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
725 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700726 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700727 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700728 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700729 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
730 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700731 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700732 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700733 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700734 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
735 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
736 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
737 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
738 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
739 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
Frank Barchard3b800452020-11-22 12:12:35 -0800740 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc2.c",
741 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc3.c",
742 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc4.c",
743 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4.c",
744 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-2x4-acc2.c",
745 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-2x4.c",
746 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-3x4.c",
747 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-4x4.c",
748 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-5x4.c",
749 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-6x4.c",
750 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc2.c",
751 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc3.c",
752 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc4.c",
753 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4.c",
754 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-2x4-acc2.c",
755 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-2x4.c",
756 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-3x4.c",
757 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-4x4.c",
758 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-5x4.c",
759 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-6x4.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -0800760 "src/f32-dwconv2d-chw/gen/3x3s2p1-wasmsimd-1x4-acc3-arm.c",
Frank Barcharddb5c32d2020-11-16 23:58:42 -0800761 "src/f32-dwconv2d-chw/gen/3x3s2p1-wasmsimd-1x4-acc3-x86.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -0800762 "src/f32-dwconv2d-chw/gen/5x5p2-wasmsimd-3x4-arm.c",
Frank Barcharddb5c32d2020-11-16 23:58:42 -0800763 "src/f32-dwconv2d-chw/gen/5x5p2-wasmsimd-3x4-x86.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -0800764 "src/f32-dwconv2d-chw/gen/5x5s2p2-wasmsimd-1x4-acc2-arm.c",
Frank Barcharddb5c32d2020-11-16 23:58:42 -0800765 "src/f32-dwconv2d-chw/gen/5x5s2p2-wasmsimd-1x4-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700766 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
767 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700768 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
769 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
770 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
771 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700772 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
773 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
774 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
775 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
776 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
777 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
778 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
779 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
780 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
781 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
782 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
783 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
784 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
785 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
786 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
787 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
788 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
789 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
790 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
791 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
792 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
793 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
794 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
795 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
796 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
797 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
798 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
799 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
800 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
801 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
802 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
803 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
804 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
805 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
806 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
807 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
808 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
809 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
810 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
811 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
812 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
813 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
814 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
815 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
816 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
817 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
818 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
819 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
820 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
821 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
822 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
823 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
824 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
825 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
826 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
827 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
828 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
829 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
830 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
831 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
832 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
833 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
834 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
835 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
836 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
837 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
838 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
839 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
840 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
841 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700842 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
843 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
844 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
XNNPACK Team965272b2020-10-23 21:10:15 -0700845 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
846 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700847 "src/f32-ibilinear/gen/wasmsimd-c4.c",
848 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700849 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
850 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
851 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
852 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
853 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
854 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
855 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
856 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
857 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
858 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
859 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
860 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
861 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
862 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
863 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
864 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
865 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
866 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
867 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
868 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
869 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
870 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
871 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
872 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
873 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
874 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
875 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
876 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
877 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
878 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
879 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
880 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
881 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
882 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
883 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
884 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
885 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
886 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
887 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
888 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700889 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
890 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700891 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
892 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
893 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
894 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700895 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
896 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700897 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
898 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
899 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700900 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
901 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700902 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
903 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
904 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
905 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
906 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
907 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
908 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700909 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
910 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700911 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
912 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
913 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
914 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700915 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700916 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700917 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700918 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
919 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700920 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700921 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
922 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700923 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700924 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
925 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700926 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700927 "src/f32-relu/gen/wasmsimd-x4.c",
928 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700929 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700930 "src/f32-rmax/wasmsimd-arm.c",
931 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700932 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
933 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
934 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
935 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
936 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
937 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700938 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
939 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
940 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
941 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
942 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
943 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800944 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined-x2.c",
945 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700946 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c",
947 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700948 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800949 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined-x2.c",
950 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700951 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c",
952 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700953 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800954 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined-x2.c",
955 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700956 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c",
957 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700958 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800959 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined-x2.c",
960 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700961 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c",
962 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700963 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800964 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined-x2.c",
965 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700966 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c",
967 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700968 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800969 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined-x2.c",
970 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700971 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c",
972 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700973 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800974 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined-x2.c",
975 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700976 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c",
977 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -0700978 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800979 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined-x2.c",
980 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700981 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c",
982 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -0700983 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700984 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
985 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700986 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700987 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
988 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700989 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700990 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
991 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700992 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700993 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
994 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700995 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700996 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
997 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700998 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700999 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
1000 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001001 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001002 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
1003 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001004 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001005 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
1006 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001007 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001008 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
1009 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001010 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001011 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
1012 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001013 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001014 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
1015 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001016 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001017 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
1018 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001019 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001020 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
1021 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001022 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001023 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
1024 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001025 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001026 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
1027 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001028 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001029 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
1030 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001031 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001032 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
1033 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001034 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001035 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
1036 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001037 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001038 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
1039 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001040 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001041 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
1042 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001043 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001044 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
1045 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001046 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001047 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
1048 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001049 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001050 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
1051 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001052 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001053 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
1054 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001055 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001056 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
1057 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001058 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001059 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
1060 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001061 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001062 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
1063 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001064 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001065 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
1066 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001067 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001068 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
1069 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001070 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001071 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
1072 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001073 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001074 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
1075 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001076 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001077 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
1078 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001079 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001080 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
1081 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001082 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001083 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
1084 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001085 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001086 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
1087 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001088 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001089 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
1090 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001091 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001092 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1093 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001094 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001095 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1096 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001097 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001098 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1099 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001100 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001101 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1102 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001103 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001104 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1105 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001106 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001107 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1108 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001109 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001110 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1111 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001112 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001113 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1114 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001115 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001116 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1117 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001118 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001119 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1120 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001121 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001122 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1123 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001124 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001125 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1126 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001127 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001128 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1129 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001130 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001131 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1132 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001133 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001134 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1135 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1136 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1137 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001138 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001139 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001140 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001141 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001142 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1143 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1144 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1145 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001146 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1147 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1148 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1149 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1150 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1151 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1152 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1153 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1154 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1155 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001156 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1157 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001158 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1159 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1160 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1161 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1162 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1163 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001164 "src/math/roundd-wasmsimd-addsub.c",
1165 "src/math/roundd-wasmsimd-cvt.c",
1166 "src/math/roundne-wasmsimd-addsub.c",
1167 "src/math/roundu-wasmsimd-addsub.c",
1168 "src/math/roundu-wasmsimd-cvt.c",
1169 "src/math/roundz-wasmsimd-addsub.c",
1170 "src/math/roundz-wasmsimd-cvt.c",
1171 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1172 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001173 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1174 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1175 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001176 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1177 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1178 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001179 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1180 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1181 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001182 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001183 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001184 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001185 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1186 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001187 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001188 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1189 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001190 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1191 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001192 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001193 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001194 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001195 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001196 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001197 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001198 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001199 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1200 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1201 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1202 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1203 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1204 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1205 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1206 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001207 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001208 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001209 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001210 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001211 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001212 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001213 "src/x32-zip/x2-wasmsimd.c",
1214 "src/x32-zip/x3-wasmsimd.c",
1215 "src/x32-zip/x4-wasmsimd.c",
1216 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001217]
1218
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001219PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -07001220 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001221]
1222
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001223PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001224 "src/qs8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001225 "src/qs8-requantization/precise-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001226 "src/qu8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001227 "src/qu8-requantization/precise-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001228]
1229
Marat Dukhan08c4a432019-10-03 09:29:21 -07001230# ISA-specific micro-kernels
1231NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001232 "src/f32-argmaxpool/4x-neon-c4.c",
1233 "src/f32-argmaxpool/9p8x-neon-c4.c",
1234 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001235 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1236 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001237 "src/f32-clamp/gen/neon-x4.c",
1238 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001239 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001240 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001241 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001242 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001243 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001244 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001245 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001246 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001247 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001248 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001249 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001250 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001251 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001252 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1253 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1254 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1255 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1256 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001257 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001258 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001259 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1260 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1261 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001262 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001263 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001264 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1265 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1266 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1267 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1268 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001269 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1270 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1271 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001272 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001273 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001274 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1275 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1276 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001277 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1278 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1279 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1280 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001281 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001282 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1283 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001284 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001285 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001286 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001287 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001288 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1289 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001290 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1291 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1292 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1293 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1294 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1295 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1296 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1297 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001298 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001299 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001300 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001301 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1302 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001303 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001304 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1305 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001306 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001307 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1308 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1309 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1310 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1311 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001312 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1313 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001314 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1315 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001316 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1317 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001318 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1319 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1320 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1321 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1322 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1323 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1324 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1325 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1326 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1327 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1328 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1329 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1330 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1331 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1332 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1333 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001334 "src/f32-hswish/gen/hswish-neon-x4.c",
1335 "src/f32-hswish/gen/hswish-neon-x8.c",
1336 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001337 "src/f32-ibilinear/gen/neon-c4.c",
1338 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001339 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001340 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001341 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001342 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1343 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001344 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001345 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1346 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1347 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1348 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001349 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1350 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001351 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1352 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001353 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1354 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001355 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1356 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1357 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001358 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1359 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001360 "src/f32-prelu/gen/neon-1x4.c",
1361 "src/f32-prelu/gen/neon-1x8.c",
1362 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001363 "src/f32-prelu/gen/neon-2x4.c",
1364 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001365 "src/f32-prelu/gen/neon-2x16.c",
1366 "src/f32-prelu/gen/neon-4x4.c",
1367 "src/f32-prelu/gen/neon-4x8.c",
1368 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001369 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001370 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001371 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001372 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1373 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001374 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001375 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1376 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001377 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001378 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1379 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001380 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1381 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1382 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1383 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1384 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1385 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1386 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1387 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1388 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1389 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1390 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1391 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1392 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001393 "src/f32-relu/gen/neon-x4.c",
1394 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001395 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001396 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001397 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1398 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1399 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1400 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1401 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1402 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1403 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1404 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1405 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1406 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1407 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1408 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001409 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1410 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1411 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1412 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1413 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1414 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001415 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1416 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1417 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1418 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001419 "src/f32-vbinary/gen/vmax-neon-x4.c",
1420 "src/f32-vbinary/gen/vmax-neon-x8.c",
1421 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1422 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1423 "src/f32-vbinary/gen/vmin-neon-x4.c",
1424 "src/f32-vbinary/gen/vmin-neon-x8.c",
1425 "src/f32-vbinary/gen/vminc-neon-x4.c",
1426 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001427 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1428 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1429 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1430 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1431 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1432 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001433 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1434 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1435 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1436 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001437 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1438 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1439 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1440 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001441 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1442 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001443 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1444 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001445 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1446 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001447 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1448 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1449 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1450 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1451 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1452 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001453 "src/f32-vunary/gen/vabs-neon-x4.c",
1454 "src/f32-vunary/gen/vabs-neon-x8.c",
1455 "src/f32-vunary/gen/vneg-neon-x4.c",
1456 "src/f32-vunary/gen/vneg-neon-x8.c",
1457 "src/f32-vunary/gen/vsqr-neon-x4.c",
1458 "src/f32-vunary/gen/vsqr-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001459 "src/math/roundd-neon-addsub.c",
1460 "src/math/roundd-neon-cvt.c",
1461 "src/math/roundne-neon-addsub.c",
1462 "src/math/roundu-neon-addsub.c",
1463 "src/math/roundu-neon-cvt.c",
1464 "src/math/roundz-neon-addsub.c",
1465 "src/math/roundz-neon-cvt.c",
1466 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
1467 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1468 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1469 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1470 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1471 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1472 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1473 "src/math/sqrt-neon-nr1rsqrts.c",
1474 "src/math/sqrt-neon-nr2rsqrts.c",
1475 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001476 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1477 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1478 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1479 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001480 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1481 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1482 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1483 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001484 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1485 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1486 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1487 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001488 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001489 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001490 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001491 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1492 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001493 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001494 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001495 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001496 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001497 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001498 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001499 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1500 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1501 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1502 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1503 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1504 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1505 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1506 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001507 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1508 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1509 "src/qu8-dwconv/up8x9-minmax-neon.c",
1510 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1511 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1512 "src/qu8-gemm/4x8-minmax-neon.c",
1513 "src/qu8-gemm/8x8-minmax-neon.c",
1514 "src/qu8-igemm/4x8-minmax-neon.c",
1515 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001516 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001517 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001518 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001519 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001520 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001521 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001522 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001523 "src/x8-zip/x2-neon.c",
1524 "src/x8-zip/x3-neon.c",
1525 "src/x8-zip/x4-neon.c",
1526 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001527 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001528 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001529 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001530 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001531 "src/x32-zip/x2-neon.c",
1532 "src/x32-zip/x3-neon.c",
1533 "src/x32-zip/x4-neon.c",
1534 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001535]
1536
1537NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001538 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1539 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1540 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1541 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1542 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1543 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1544 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1545 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1546 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1547 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1548 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1549 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1550 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1551 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1552 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1553 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1554 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1555 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1556 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1557 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1558 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1559 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1560 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1561 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1562 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1563 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1564 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1565 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1566 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1567 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001568 "src/f32-ibilinear/gen/neonfma-c4.c",
1569 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001570 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001571 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001572 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001573 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1574 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001575 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1576 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001577 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1578 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001579 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1580 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001581 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001582 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001583 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001584 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1585 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001586 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001587 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1588 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001589 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001590 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1591 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001592 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1593 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1594 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1595 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1596 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1597 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1598 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1599 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1600 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1601 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1602 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1603 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1604 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001605 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1606 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1607 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1608 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1609 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1610 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001611 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1612 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1613 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1614 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1615 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1616 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001617 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1618 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1619 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1620 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1621 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1622 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001623 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1624 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1625 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1626 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1627 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1628 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001629 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1630 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1631 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1632 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1633 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1634 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001635 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1636 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1637 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1638 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1639 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1640 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001641 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1642 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1643 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1644 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1645 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1646 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1647 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1648 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1649 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1650 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1651 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1652 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1653 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1654 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1655 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1656 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1657 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1658 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001659 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1660 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001661 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1662 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1663 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1664 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1665 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1666 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1667 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1668 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1669 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1670 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1671 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1672 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1673 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1674 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1675 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1676 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1677 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1678 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1679 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1680 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08001681 "src/math/exp-neonfma-rr2-lut64-p2.c",
1682 "src/math/exp-neonfma-rr2-p5.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08001683 "src/math/expminus-neonfma-rr2-lut64-p2.c",
1684 "src/math/expminus-neonfma-rr2-lut2048-p1.c",
1685 "src/math/expminus-neonfma-rr2-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001686 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1687 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1688 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001689 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1690 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1691 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001692 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1693 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1694 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001695 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1696 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1697 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001698 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1699 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1700 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001701 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1702 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1703 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001704 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001705 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001706 "src/math/sqrt-neonfma-nr2fma.c",
1707 "src/math/sqrt-neonfma-nr2fma1adj.c",
1708 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001709]
1710
1711AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001712 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001713 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001714 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001715 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001716 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001717 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001718 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001719 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001720 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001721 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
1722 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
1723 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001724 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001725 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001726 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
1727 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
1728 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
1729 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
1730 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001731 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
1732 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
1733 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001734 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001735 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001736 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
1737 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
1738 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001739 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
1740 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
1741 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
1742 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001743 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001744 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
1745 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001746 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001747 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001748 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001749 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001750 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
1751 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001752 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
1753 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
1754 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
1755 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
1756 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
1757 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
1758 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
1759 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001760 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001761 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001762 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1763 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1764 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1765 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1766 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1767 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1768 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1769 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1770 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1771 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1772 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1773 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1774 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1775 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1776 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1777 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1778 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1779 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1780 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1781 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001782 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1783 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1784 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1785 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1786 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1787 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1788 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1789 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1790 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1791 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1792 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1793 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001794 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1795 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1796 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1797 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1798 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1799 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001800 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001801 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001802 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1803 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1804 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1805 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001806 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001807 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1808 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1809 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001810 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1811 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1812 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1813 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001814 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001815 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1816 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1817 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001818 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001819 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001820 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
1821 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
1822 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001823 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1824 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1825 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1826 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1827 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1828 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001829 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1830 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001831 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001832 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001833 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001834 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001835 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001836 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001837]
1838
Marat Dukhan8853b822020-05-07 12:19:01 -07001839NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001840 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1841 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001842 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1843 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1844 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1845 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1846 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1847 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001848 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001849 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001850 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001851 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001852]
1853
Marat Dukhan08c4a432019-10-03 09:29:21 -07001854AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001855 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001856 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001857 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1858 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1859 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1860 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001861 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1862 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1863 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1864 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1865 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1866 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1867 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1868 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001869 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1870 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001871 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1872 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1873 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1874 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1875 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1876 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1877 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
1878 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1879 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1880 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1881 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1882 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1883 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1884 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1885 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1886 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001887 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001888 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1889 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1890 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1891 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1892 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1893 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1894 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1895 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1896 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001897 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001898 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001899 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001900 "src/f16-relu/gen/neonfp16arith-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001901 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001902 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001903 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001904 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001905 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001906 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001907 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001908 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1909 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1910 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1911 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1912 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1913 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1914 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1915 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1916 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1917 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1918 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1919 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1920 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1921 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1922 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1923 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1924 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1925 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1926 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1927 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1928 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1929 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1930 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1931 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1932 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1933 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
1934 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1935 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1936 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1937 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
1938 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001939]
1940
Benoit Jacoba9644732020-08-13 12:48:55 -07001941NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07001942 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001943 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001944 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001945 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001946 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001947 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001948 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001949 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001950 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001951 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001952 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001953 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001954 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001955 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001956 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001957 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001958 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001959 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001960]
1961
Marat Dukhan08c4a432019-10-03 09:29:21 -07001962SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001963 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1964 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001965 "src/f32-clamp/gen/sse-x4.c",
1966 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001967 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1968 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001969 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1970 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1971 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1972 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001973 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1974 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001975 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1976 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1977 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1978 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001979 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1980 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001981 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
1982 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
1983 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07001984 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001985 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07001986 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
1987 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
1988 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
1989 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
1990 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07001991 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
1992 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
1993 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001994 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07001995 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001996 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
1997 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
1998 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07001999 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
2000 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
2001 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
2002 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
2003 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
2004 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
2005 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
2006 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
2007 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
2008 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
2009 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
2010 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
2011 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002012 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
2013 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
2014 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
2015 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
2016 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
2017 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
2018 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
2019 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002020 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08002021 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002022 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002023 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
2024 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002025 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
2026 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
2027 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
2028 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
2029 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
2030 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002031 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
2032 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
2033 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
2034 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
2035 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
2036 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
2037 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002038 "src/f32-hswish/gen/hswish-sse-x4.c",
2039 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07002040 "src/f32-ibilinear/gen/sse-c4.c",
2041 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002042 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
2043 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
2044 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
2045 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2046 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2047 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2048 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002049 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2050 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2051 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002052 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002053 "src/f32-prelu/gen/sse-2x4.c",
2054 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002055 "src/f32-relu/gen/sse-x4.c",
2056 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002057 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002058 "src/f32-spmm/gen/4x1-minmax-sse.c",
2059 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002060 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002061 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002062 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2063 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2064 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2065 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2066 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2067 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2068 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2069 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002070 "src/f32-vbinary/gen/vmax-sse-x4.c",
2071 "src/f32-vbinary/gen/vmax-sse-x8.c",
2072 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2073 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2074 "src/f32-vbinary/gen/vmin-sse-x4.c",
2075 "src/f32-vbinary/gen/vmin-sse-x8.c",
2076 "src/f32-vbinary/gen/vminc-sse-x4.c",
2077 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002078 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2079 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2080 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2081 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2082 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2083 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2084 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2085 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002086 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2087 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2088 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2089 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002090 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2091 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2092 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2093 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002094 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2095 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002096 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2097 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002098 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2099 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002100 "src/f32-vunary/gen/vabs-sse-x4.c",
2101 "src/f32-vunary/gen/vabs-sse-x8.c",
2102 "src/f32-vunary/gen/vneg-sse-x4.c",
2103 "src/f32-vunary/gen/vneg-sse-x8.c",
2104 "src/f32-vunary/gen/vsqr-sse-x4.c",
2105 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002106 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002107 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002108 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002109 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002110 "src/math/sqrt-sse-hh1mac.c",
2111 "src/math/sqrt-sse-nr1mac.c",
2112 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002113 "src/x32-fill/sse.c",
2114 "src/x32-packx/x4-sse.c",
2115 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002116]
2117
2118SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002119 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002120 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002121 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002122 "src/f32-prelu/gen/sse2-2x4.c",
2123 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002124 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002125 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002126 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002127 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2128 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002129 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002130 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2131 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002132 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002133 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2134 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002135 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002136 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
2137 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
2138 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
2139 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
2140 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
2141 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002142 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002143 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002144 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002145 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002146 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
2147 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002148 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2149 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002150 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2151 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002152 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2153 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2154 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2155 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2156 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2157 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002158 "src/math/exp-sse2-rr2-lut64-p2.c",
2159 "src/math/exp-sse2-rr2-p5.c",
Marat Dukhand6d321e2020-11-20 18:35:55 -08002160 "src/math/expm1minus-sse2-rr2-p5.c",
Marat Dukhana438aca2020-11-20 15:45:01 -08002161 "src/math/expm1minus-sse2-rr2-p6.c",
Frank Barchard3b800452020-11-22 12:12:35 -08002162 "src/math/expminus-sse2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002163 "src/math/roundd-sse2-cvt.c",
2164 "src/math/roundne-sse2-cvt.c",
2165 "src/math/roundu-sse2-cvt.c",
2166 "src/math/roundz-sse2-cvt.c",
2167 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2168 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2169 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2170 "src/math/sigmoid-sse2-rr2-p5-div.c",
2171 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2172 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002173 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2174 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2175 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002176 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2177 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2178 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002179 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2180 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2181 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002182 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002183 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002184 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002185 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002186 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002187 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002188 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2189 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002190 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002191 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2192 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002193 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002194 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2195 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2196 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002197 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002198 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002199 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002200 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002201 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002202 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002203 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002204 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002205 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2206 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002207 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002208 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002209 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002210 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2211 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2212 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2213 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002214 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2215 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2216 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2217 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002218 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2219 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002220 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2221 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2222 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2223 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2224 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002225 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002226 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002227 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002228 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002229 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002230 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002231 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002232 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002233 "src/x8-zip/x2-sse2.c",
2234 "src/x8-zip/x3-sse2.c",
2235 "src/x8-zip/x4-sse2.c",
2236 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002237 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002238 "src/x32-zip/x2-sse2.c",
2239 "src/x32-zip/x3-sse2.c",
2240 "src/x32-zip/x4-sse2.c",
2241 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002242]
2243
2244SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002245 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2246 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2247 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002248 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002249 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002250 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2251 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2252 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2253 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2254 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002255 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2256 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2257 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002258 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2259 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2260 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002261 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2262 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2263 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002264 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002265 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002266 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002267 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002268 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002269 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002270 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2271 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002272 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002273 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2274 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002275 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002276 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2277 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2278 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002279 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002280 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002281 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002282 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002283 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002284 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002285 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002286 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002287 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2288 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002289 "src/qs8-requantization/precise-ssse3.c",
2290 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002291 "src/qu8-requantization/precise-ssse3.c",
2292 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002293]
2294
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002295SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002296 "src/f32-prelu/gen/sse41-2x4.c",
2297 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002298 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2299 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2300 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2301 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2302 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2303 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002304 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002305 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002306 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002307 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002308 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2309 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002310 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2311 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002312 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2313 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002314 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2315 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2316 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2317 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2318 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2319 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2320 "src/math/roundd-sse41.c",
2321 "src/math/roundne-sse41.c",
2322 "src/math/roundu-sse41.c",
2323 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002324 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2325 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2326 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002327 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2328 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2329 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002330 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2331 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2332 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002333 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002334 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002335 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002336 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002337 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002338 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002339 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2340 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002341 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002342 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2343 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002344 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002345 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2346 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2347 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002348 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002349 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002350 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002351 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002352 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002353 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002354 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002355 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002356 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2357 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002358 "src/qs8-requantization/fp32-sse4.c",
2359 "src/qs8-requantization/precise-sse4.c",
2360 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002361 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2362 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2363 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2364 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002365 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2366 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2367 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2368 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002369 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2370 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2371 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2372 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002373 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2374 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2375 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2376 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002377 "src/qu8-requantization/precise-sse4.c",
2378 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002379]
2380
Marat Dukhan08c4a432019-10-03 09:29:21 -07002381AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002382 "src/f32-clamp/gen/avx-x8.c",
2383 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002384 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2385 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002386 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2387 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002388 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2389 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002390 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2391 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2392 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2393 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2394 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2395 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002396 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002397 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2398 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002399 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002400 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002401 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002402 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002403 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2404 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2405 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2406 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2407 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2408 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2409 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2410 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2411 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2412 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2413 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002414 "src/f32-hswish/gen/hswish-avx-x8.c",
2415 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002416 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002417 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2418 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002419 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002420 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002421 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002422 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002423 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2424 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002425 "src/f32-prelu/gen/avx-2x8.c",
2426 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002427 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002428 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002429 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002430 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2431 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2432 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2433 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2434 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2435 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2436 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2437 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2438 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2439 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2440 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2441 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2442 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2443 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2444 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2445 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2446 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2447 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2448 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2449 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002450 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2451 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2452 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2453 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2454 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2455 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2456 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2457 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002458 "src/f32-vbinary/gen/vmax-avx-x8.c",
2459 "src/f32-vbinary/gen/vmax-avx-x16.c",
2460 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2461 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2462 "src/f32-vbinary/gen/vmin-avx-x8.c",
2463 "src/f32-vbinary/gen/vmin-avx-x16.c",
2464 "src/f32-vbinary/gen/vminc-avx-x8.c",
2465 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002466 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2467 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2468 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2469 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2470 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2471 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2472 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2473 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002474 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2475 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2476 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2477 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002478 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2479 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2480 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2481 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002482 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2483 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002484 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2485 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002486 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2487 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2488 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2489 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2490 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2491 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002492 "src/f32-vscale/avx-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002493 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2494 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002495 "src/f32-vunary/gen/vabs-avx-x8.c",
2496 "src/f32-vunary/gen/vabs-avx-x16.c",
2497 "src/f32-vunary/gen/vneg-avx-x8.c",
2498 "src/f32-vunary/gen/vneg-avx-x16.c",
2499 "src/f32-vunary/gen/vsqr-avx-x8.c",
2500 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002501 "src/math/exp-avx-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002502 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002503 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002504 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2505 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002506]
2507
Marat Dukhan1566fee2020-08-02 21:55:41 -07002508XOP_UKERNELS = [
2509 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002510 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002511 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002512 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002513 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002514 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002515 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2516 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002517 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002518 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2519 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002520 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002521 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2522 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2523 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002524 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002525 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002526 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002527 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002528 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002529 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002530 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002531 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002532 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2533 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002534 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2535 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2536 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2537 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2538 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2539 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2540 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2541 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002542]
2543
Marat Dukhanfda12b82019-11-21 12:27:59 -08002544FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002545 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2546 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002547 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2548 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002549 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2550 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002551 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2552 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2553 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2554 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2555 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2556 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002557 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002558 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2559 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2560 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2561 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002562 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002563 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2564 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002565 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002566 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2567 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002568 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2569 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2570 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002571 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2572 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2573 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2574 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2575 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2576 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2577 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2578 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2579 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2580 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2581 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2582 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2583 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2584 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002585 "src/f32-hswish/gen/hswish-fma3-x8.c",
2586 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002587 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002588 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2589 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2590 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2591 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002592 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002593 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2594 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002595 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002596 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2597 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002598 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2599 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2600 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002601 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2602 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2603 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2604 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2605 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2606 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2607 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2608 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002609 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002610 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002611 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002612]
2613
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002614AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002615 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2616 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002617 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002618 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002619 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002620 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2621 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002622 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002623 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2624 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2625 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002626 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002627 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2628 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002629 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002630 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002631 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002632 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2633 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002634 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002635 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2636 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2637 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002638 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002639 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2640 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002641 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002642 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002643 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002644 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2645 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002646 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002647 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2648 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2649 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002650 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002651 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2652 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2653 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2654 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2655 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2656 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2657 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2658 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2659 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2660 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2661 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2662 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2663 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2664 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2665 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2666 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2667 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2668 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2669 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2670 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2671 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2672 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2673 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2674 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2675 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2676 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2677 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2678 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2679 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2680 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002681 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2682 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2683 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2684 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2685 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2686 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2687 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2688 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2689 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2690 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2691 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2692 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2693 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2694 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2695 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2696 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2697 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2698 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2699 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2700 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2701 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2702 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2703 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2704 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002705 "src/math/exp-avx2-rr2-lut8-p3-perm.c",
2706 "src/math/exp-avx2-rr2-lut8-p4-perm.c",
2707 "src/math/exp-avx2-rr2-p5.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08002708 "src/math/expminus-avx2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002709 "src/math/extexp-avx2-p5.c",
2710 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2711 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2712 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2713 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2714 "src/math/sigmoid-avx2-rr1-p5-div.c",
2715 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
2716 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2717 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2718 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2719 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2720 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2721 "src/math/sigmoid-avx2-rr2-p5-div.c",
2722 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2723 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002724 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002725 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002726 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2727 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002728 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002729 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002730 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002731 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002732 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002733 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002734 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002735 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002736 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2737 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2738 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002739 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2740 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2741 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2742 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2743 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2744 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2745 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2746 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002747]
2748
Marat Dukhan08c4a432019-10-03 09:29:21 -07002749AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002750 "src/f32-clamp/gen/avx512f-x16.c",
2751 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002752 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2753 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002754 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2755 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002756 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2757 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002758 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2759 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2760 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2761 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2762 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2763 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002764 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2765 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2766 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2767 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2768 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2769 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002770 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2771 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2772 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2773 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2774 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2775 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002776 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2777 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002778 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2779 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2780 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2781 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2782 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2783 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002784 "src/f32-prelu/gen/avx512f-2x16.c",
2785 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002786 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2787 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002788 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002789 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002790 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002791 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2792 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002793 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002794 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2795 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2796 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002797 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002798 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2799 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002800 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002801 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002802 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002803 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2804 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002805 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002806 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2807 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2808 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002809 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002810 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2811 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002812 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002813 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002814 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002815 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2816 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002817 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002818 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2819 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2820 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002821 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002822 "src/f32-relu/gen/avx512f-x16.c",
2823 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002824 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002825 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2826 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2827 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2828 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2829 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2830 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2831 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2832 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2833 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2834 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2835 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2836 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2837 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2838 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2839 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2840 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002841 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2842 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2843 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2844 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2845 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2846 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2847 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2848 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2849 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2850 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2851 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2852 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2853 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2854 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2855 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2856 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002857 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2858 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2859 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2860 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2861 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2862 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2863 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2864 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2865 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2866 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2867 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2868 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2869 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2870 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2871 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2872 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002873 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2874 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2875 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2876 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2877 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2878 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2879 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2880 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002881 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2882 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2883 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2884 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2885 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2886 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2887 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2888 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002889 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2890 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2891 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2892 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2893 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2894 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2895 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2896 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002897 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2898 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2899 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2900 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002901 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2902 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2903 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2904 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002905 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2906 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002907 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2908 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
2909 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2910 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2911 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2912 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2913 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2914 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002915 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002916 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2917 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2918 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2919 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2920 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2921 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2922 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2923 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2924 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2925 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2926 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2927 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2928 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2929 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2930 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2931 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2932 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2933 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2934 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2935 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2936 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2937 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2938 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2939 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002940 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2941 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2942 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2943 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2944 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2945 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2946 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2947 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002948 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2949 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2950 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2951 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2952 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2953 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002954 "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
2955 "src/math/exp-avx512f-rr2-lut16-p3-perm.c",
2956 "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
2957 "src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
2958 "src/math/exp-avx512f-rr2-p5-scalef.c",
2959 "src/math/exp-avx512f-rr2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002960 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002961 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002962 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002963 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002964 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002965 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002966 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002967 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002968 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002969 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
2970 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
2971 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
2972 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
2973 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2974 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2975 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2976 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2977 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
2978 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002979 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002980 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002981 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
2982 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
2983 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
2984 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002985 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002986 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002987 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002988]
2989
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002990AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002991 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2992 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002993 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2994 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2995 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2996 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2997 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2998 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2999 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
3000 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
3001]
3002
Frank Barchardbcedc082020-08-17 18:00:51 -07003003WASM32_ASM_UKERNELS = [
3004 "src/f32-relu/wasm_shr_x1.S",
3005 "src/f32-relu/wasm_shr_x2.S",
3006 "src/f32-relu/wasm_shr_x4.S",
3007]
3008
Marat Dukhan08c4a432019-10-03 09:29:21 -07003009AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07003010 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07003011 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003012 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3013 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003014 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003015 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07003016 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003017 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003018 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3019 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003020 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
3021 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
3022 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
3023 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003024]
3025
3026AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003027 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003028 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003029 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003030 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003031 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003032 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07003033 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003034 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
3035 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
3036 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
3037 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
3038 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
3039 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
3040 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003041 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
3042 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003043 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3044 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3045 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003046 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
3047 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003048 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3049 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3050 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3051 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003052 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003053 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3054 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003055 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3056 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3057 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3058 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003059 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003060 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003061 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003062 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003063 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3064 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3065 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3066 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3067 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3068 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3069 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3070 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3071 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3072 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3073 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3074 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3075 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3076 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3077 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3078 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3079 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3080 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3081 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3082 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3083 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3084 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003085 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003086 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003087 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3088 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003089 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3090 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3091 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3092 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3093 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3094 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003095 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3096 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3097 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3098 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003099 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3100 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003101 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07003102 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07003103 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003104]
3105
Marat Dukhan1b354632020-03-23 12:50:22 -07003106INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003107 "src/xnnpack/argmaxpool.h",
3108 "src/xnnpack/avgpool.h",
3109 "src/xnnpack/clamp.h",
3110 "src/xnnpack/common.h",
3111 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003112 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003113 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003114 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003115 "src/xnnpack/gavgpool.h",
3116 "src/xnnpack/gemm.h",
3117 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003118 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003119 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003120 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003121 "src/xnnpack/lut.h",
3122 "src/xnnpack/math.h",
3123 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003124 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003125 "src/xnnpack/packx.h",
3126 "src/xnnpack/pad.h",
3127 "src/xnnpack/params.h",
3128 "src/xnnpack/pavgpool.h",
3129 "src/xnnpack/ppmm.h",
3130 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003131 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003132 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003133 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003134 "src/xnnpack/rmax.h",
3135 "src/xnnpack/scalar-utils.h",
3136 "src/xnnpack/spmm.h",
3137 "src/xnnpack/unpool.h",
3138 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003139 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003140 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003141 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003142 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003143 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003144 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003145 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003146]
3147
3148INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003149 "include/xnnpack.h",
3150 "src/xnnpack/allocator.h",
3151 "src/xnnpack/compute.h",
3152 "src/xnnpack/im2col.h",
3153 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003154 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003155 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003156 "src/xnnpack/operator.h",
3157 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003158 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003159 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003160 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003161 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003162]
3163
Marat Dukhan1b354632020-03-23 12:50:22 -07003164ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003165 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003166]
3167
Marat Dukhan1b354632020-03-23 12:50:22 -07003168MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003169 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003170 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003171]
3172
Marat Dukhan1b354632020-03-23 12:50:22 -07003173MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003174 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003175 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003176 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003177 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003178]
3179
3180OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003181 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003182 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003183]
3184
3185WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003186 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003187 "src/xnnpack/operator.h",
3188 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003189]
3190
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003191LOGGING_COPTS = select({
3192 # No logging in optimized mode
3193 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3194 # Full logging in debug mode
3195 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3196 # Error-only logging in default (fastbuild) mode
3197 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3198})
3199
Marat Dukhan3b59de22020-06-03 20:15:19 -07003200LOGGING_SRCS = select({
3201 # No logging in optimized mode
3202 ":optimized_build": [],
3203 "//conditions:default": [
3204 "src/operator-strings.c",
3205 "src/subgraph-strings.c",
3206 ],
3207})
3208
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003209LOGGING_HDRS = [
3210 "src/xnnpack/log.h",
3211]
3212
Marat Dukhan08c4a432019-10-03 09:29:21 -07003213xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003214 name = "tables",
3215 srcs = TABLE_SRCS,
3216 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003217 gcc_copts = xnnpack_gcc_std_copts(),
3218 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003219)
3220
3221xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003222 name = "scalar_ukernels",
3223 srcs = SCALAR_UKERNELS,
3224 hdrs = INTERNAL_HDRS,
3225 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003226 gcc_copts = xnnpack_gcc_std_copts(),
3227 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003228 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003229 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003230 "@FP16",
3231 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003232 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003233 ],
3234)
3235
3236xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003237 name = "scalar_ukernels_test_mode",
3238 srcs = SCALAR_UKERNELS,
3239 hdrs = INTERNAL_HDRS,
3240 aarch32_copts = ["-marm"],
3241 copts = [
3242 "-UNDEBUG",
3243 "-DXNN_TEST_MODE=1",
3244 ],
3245 gcc_copts = xnnpack_gcc_std_copts(),
3246 msvc_copts = xnnpack_msvc_std_copts(),
3247 deps = [
3248 ":tables",
3249 "@FP16",
3250 "@FXdiv",
3251 "@pthreadpool",
3252 ],
3253)
3254
3255xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003256 name = "wasm_ukernels",
3257 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003258 gcc_copts = xnnpack_gcc_std_copts(),
3259 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003260 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003261 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003262 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003263 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003264 "@FP16",
3265 "@FXdiv",
3266 "@pthreadpool",
3267 ],
3268)
3269
3270xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003271 name = "wasm_ukernels_test_mode",
3272 hdrs = INTERNAL_HDRS,
3273 copts = [
3274 "-UNDEBUG",
3275 "-DXNN_TEST_MODE=1",
3276 ],
3277 gcc_copts = xnnpack_gcc_std_copts(),
3278 msvc_copts = xnnpack_msvc_std_copts(),
3279 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003280 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003281 deps = [
3282 ":tables",
3283 "@FP16",
3284 "@FXdiv",
3285 "@pthreadpool",
3286 ],
3287)
3288
3289xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003290 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003291 hdrs = INTERNAL_HDRS,
3292 aarch32_copts = [
3293 "-marm",
3294 "-mfpu=neon",
3295 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003296 gcc_copts = xnnpack_gcc_std_copts(),
3297 gcc_x86_copts = ["-msse2"],
3298 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003299 optimized_copts = [
3300 "-O3",
3301 "-ffast-math",
3302 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003303 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003304 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003305 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003306 "@FP16",
3307 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003308 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003309 ],
3310)
3311
3312xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003313 name = "psimd_fastmath_ukernels_test_mode",
3314 hdrs = INTERNAL_HDRS,
3315 aarch32_copts = [
3316 "-marm",
3317 "-mfpu=neon",
3318 ],
3319 copts = [
3320 "-UNDEBUG",
3321 "-DXNN_TEST_MODE=1",
3322 ],
3323 gcc_copts = xnnpack_gcc_std_copts(),
3324 gcc_x86_copts = ["-msse2"],
3325 msvc_copts = xnnpack_msvc_std_copts(),
3326 optimized_copts = [
3327 "-O3",
3328 "-ffast-math",
3329 ],
3330 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
3331 deps = [
3332 ":tables",
3333 "@FP16",
3334 "@psimd",
3335 "@pthreadpool",
3336 ],
3337)
3338
3339xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003340 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003341 hdrs = INTERNAL_HDRS,
3342 aarch32_copts = [
3343 "-marm",
3344 "-mfpu=neon",
3345 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003346 gcc_copts = xnnpack_gcc_std_copts(),
3347 gcc_x86_copts = ["-msse2"],
3348 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003349 optimized_copts = [
3350 "-O3",
3351 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003352 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003353 deps = [
3354 ":tables",
3355 "@FP16",
3356 "@psimd",
3357 "@pthreadpool",
3358 ],
3359)
3360
3361xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003362 name = "psimd_accmath_ukernels_test_mode",
3363 hdrs = INTERNAL_HDRS,
3364 aarch32_copts = [
3365 "-marm",
3366 "-mfpu=neon",
3367 ],
3368 copts = [
3369 "-UNDEBUG",
3370 "-DXNN_TEST_MODE=1",
3371 ],
3372 gcc_copts = xnnpack_gcc_std_copts(),
3373 gcc_x86_copts = ["-msse2"],
3374 msvc_copts = xnnpack_msvc_std_copts(),
3375 optimized_copts = [
3376 "-O3",
3377 ],
3378 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3379 deps = [
3380 ":tables",
3381 "@FP16",
3382 "@psimd",
3383 "@pthreadpool",
3384 ],
3385)
3386
3387xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003388 name = "neon_ukernels",
3389 hdrs = INTERNAL_HDRS,
3390 aarch32_copts = [
3391 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003392 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003393 "-mfpu=neon",
3394 ],
3395 aarch32_srcs = NEON_UKERNELS,
3396 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003397 gcc_copts = xnnpack_gcc_std_copts(),
3398 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003399 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003400 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003401 "@FP16",
3402 "@pthreadpool",
3403 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003404)
3405
3406xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003407 name = "neon_ukernels_test_mode",
3408 hdrs = INTERNAL_HDRS,
3409 aarch32_copts = [
3410 "-marm",
3411 "-march=armv7-a",
3412 "-mfpu=neon",
3413 ],
3414 aarch32_srcs = NEON_UKERNELS,
3415 aarch64_srcs = NEON_UKERNELS,
3416 copts = [
3417 "-UNDEBUG",
3418 "-DXNN_TEST_MODE=1",
3419 ],
3420 gcc_copts = xnnpack_gcc_std_copts(),
3421 msvc_copts = xnnpack_msvc_std_copts(),
3422 deps = [
3423 ":tables",
3424 "@FP16",
3425 "@pthreadpool",
3426 ],
3427)
3428
3429xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003430 name = "neonfma_ukernels",
3431 hdrs = INTERNAL_HDRS,
3432 aarch32_copts = [
3433 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003434 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003435 "-mfpu=neon-vfpv4",
3436 ],
3437 aarch32_srcs = NEONFMA_UKERNELS,
3438 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003439 apple_aarch32_copts = [
3440 "-mcpu=swift",
3441 "-mtune=generic",
3442 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003443 gcc_copts = xnnpack_gcc_std_copts(),
3444 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003445 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003446 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003447 "@FP16",
3448 "@pthreadpool",
3449 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003450)
3451
3452xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003453 name = "neonfma_ukernels_test_mode",
3454 hdrs = INTERNAL_HDRS,
3455 aarch32_copts = [
3456 "-marm",
3457 "-march=armv7-a",
3458 "-mfpu=neon-vfpv4",
3459 ],
3460 aarch32_srcs = NEONFMA_UKERNELS,
3461 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003462 apple_aarch32_copts = [
3463 "-mcpu=swift",
3464 "-mtune=generic",
3465 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003466 copts = [
3467 "-UNDEBUG",
3468 "-DXNN_TEST_MODE=1",
3469 ],
3470 gcc_copts = xnnpack_gcc_std_copts(),
3471 msvc_copts = xnnpack_msvc_std_copts(),
3472 deps = [
3473 ":tables",
3474 "@FP16",
3475 "@pthreadpool",
3476 ],
3477)
3478
3479xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003480 name = "neonv8_ukernels",
3481 hdrs = INTERNAL_HDRS,
3482 aarch32_copts = [
3483 "-marm",
3484 "-march=armv8-a",
3485 "-mfpu=neon-fp-armv8",
3486 ],
3487 aarch32_srcs = NEONV8_UKERNELS,
3488 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003489 apple_aarch32_copts = [
3490 "-mcpu=cyclone",
3491 "-mtune=generic",
3492 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003493 gcc_copts = xnnpack_gcc_std_copts(),
3494 msvc_copts = xnnpack_msvc_std_copts(),
3495 deps = [
3496 ":tables",
3497 "@FP16",
3498 "@pthreadpool",
3499 ],
3500)
3501
3502xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003503 name = "neonv8_ukernels_test_mode",
3504 hdrs = INTERNAL_HDRS,
3505 aarch32_copts = [
3506 "-marm",
3507 "-march=armv8-a",
3508 "-mfpu=neon-fp-armv8",
3509 ],
3510 aarch32_srcs = NEONV8_UKERNELS,
3511 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003512 apple_aarch32_copts = [
3513 "-mcpu=cyclone",
3514 "-mtune=generic",
3515 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003516 copts = [
3517 "-UNDEBUG",
3518 "-DXNN_TEST_MODE=1",
3519 ],
3520 gcc_copts = xnnpack_gcc_std_copts(),
3521 msvc_copts = xnnpack_msvc_std_copts(),
3522 deps = [
3523 ":tables",
3524 "@FP16",
3525 "@pthreadpool",
3526 ],
3527)
3528
3529xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003530 name = "neonfp16arith_ukernels",
3531 hdrs = INTERNAL_HDRS,
3532 aarch64_copts = ["-march=armv8.2-a+fp16"],
3533 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003534 gcc_copts = xnnpack_gcc_std_copts(),
3535 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003536 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003537 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003538 "@FP16",
3539 "@pthreadpool",
3540 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003541)
3542
3543xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003544 name = "neonfp16arith_ukernels_test_mode",
3545 hdrs = INTERNAL_HDRS,
3546 aarch64_copts = ["-march=armv8.2-a+fp16"],
3547 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3548 copts = [
3549 "-UNDEBUG",
3550 "-DXNN_TEST_MODE=1",
3551 ],
3552 gcc_copts = xnnpack_gcc_std_copts(),
3553 msvc_copts = xnnpack_msvc_std_copts(),
3554 deps = [
3555 ":tables",
3556 "@FP16",
3557 "@pthreadpool",
3558 ],
3559)
3560
3561xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003562 name = "neondot_ukernels",
3563 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003564 aarch32_copts = [
3565 "-marm",
3566 "-march=armv8.2-a+dotprod",
3567 "-mfpu=neon-fp-armv8",
3568 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003569 aarch32_srcs = NEONDOT_UKERNELS,
3570 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3571 aarch64_srcs = NEONDOT_UKERNELS,
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(
3582 name = "neondot_ukernels_test_mode",
3583 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003584 aarch32_copts = [
3585 "-marm",
3586 "-march=armv8.2-a+dotprod",
3587 "-mfpu=neon-fp-armv8",
3588 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003589 aarch32_srcs = NEONDOT_UKERNELS,
3590 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3591 aarch64_srcs = NEONDOT_UKERNELS,
3592 copts = [
3593 "-UNDEBUG",
3594 "-DXNN_TEST_MODE=1",
3595 ],
3596 gcc_copts = xnnpack_gcc_std_copts(),
3597 msvc_copts = xnnpack_msvc_std_copts(),
3598 deps = [
3599 ":tables",
3600 "@FP16",
3601 "@pthreadpool",
3602 ],
3603)
3604
3605xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003606 name = "sse2_ukernels",
3607 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003608 gcc_copts = xnnpack_gcc_std_copts(),
3609 gcc_x86_copts = ["-msse2"],
3610 msvc_copts = xnnpack_msvc_std_copts(),
3611 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003612 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003613 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003614 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003615 "@FP16",
3616 "@pthreadpool",
3617 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003618)
3619
3620xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003621 name = "sse2_ukernels_test_mode",
3622 hdrs = INTERNAL_HDRS,
3623 copts = [
3624 "-UNDEBUG",
3625 "-DXNN_TEST_MODE=1",
3626 ],
3627 gcc_copts = xnnpack_gcc_std_copts(),
3628 gcc_x86_copts = ["-msse2"],
3629 msvc_copts = xnnpack_msvc_std_copts(),
3630 msvc_x86_32_copts = ["/arch:SSE2"],
3631 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3632 deps = [
3633 ":tables",
3634 "@FP16",
3635 "@pthreadpool",
3636 ],
3637)
3638
3639xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003640 name = "ssse3_ukernels",
3641 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003642 gcc_copts = xnnpack_gcc_std_copts(),
3643 gcc_x86_copts = ["-mssse3"],
3644 msvc_copts = xnnpack_msvc_std_copts(),
3645 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003646 x86_srcs = SSSE3_UKERNELS,
3647 deps = [
3648 ":tables",
3649 "@FP16",
3650 "@pthreadpool",
3651 ],
3652)
3653
3654xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003655 name = "ssse3_ukernels_test_mode",
3656 hdrs = INTERNAL_HDRS,
3657 copts = [
3658 "-UNDEBUG",
3659 "-DXNN_TEST_MODE=1",
3660 ],
3661 gcc_copts = xnnpack_gcc_std_copts(),
3662 gcc_x86_copts = ["-mssse3"],
3663 msvc_copts = xnnpack_msvc_std_copts(),
3664 msvc_x86_32_copts = ["/arch:SSE2"],
3665 x86_srcs = SSSE3_UKERNELS,
3666 deps = [
3667 ":tables",
3668 "@FP16",
3669 "@pthreadpool",
3670 ],
3671)
3672
3673xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003674 name = "sse41_ukernels",
3675 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003676 gcc_copts = xnnpack_gcc_std_copts(),
3677 gcc_x86_copts = ["-msse4.1"],
3678 msvc_copts = xnnpack_msvc_std_copts(),
3679 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003680 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003681 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003682 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003683 "@FP16",
3684 "@pthreadpool",
3685 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003686)
3687
3688xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003689 name = "sse41_ukernels_test_mode",
3690 hdrs = INTERNAL_HDRS,
3691 copts = [
3692 "-UNDEBUG",
3693 "-DXNN_TEST_MODE=1",
3694 ],
3695 gcc_copts = xnnpack_gcc_std_copts(),
3696 gcc_x86_copts = ["-msse4.1"],
3697 msvc_copts = xnnpack_msvc_std_copts(),
3698 msvc_x86_32_copts = ["/arch:SSE2"],
3699 x86_srcs = SSE41_UKERNELS,
3700 deps = [
3701 ":tables",
3702 "@FP16",
3703 "@pthreadpool",
3704 ],
3705)
3706
3707xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003708 name = "avx_ukernels",
3709 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003710 gcc_copts = xnnpack_gcc_std_copts(),
3711 gcc_x86_copts = ["-mavx"],
3712 msvc_copts = xnnpack_msvc_std_copts(),
3713 msvc_x86_32_copts = ["/arch:AVX"],
3714 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003715 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003716 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003717 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003718 "@FP16",
3719 "@pthreadpool",
3720 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003721)
3722
3723xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003724 name = "avx_ukernels_test_mode",
3725 hdrs = INTERNAL_HDRS,
3726 copts = [
3727 "-UNDEBUG",
3728 "-DXNN_TEST_MODE=1",
3729 ],
3730 gcc_copts = xnnpack_gcc_std_copts(),
3731 gcc_x86_copts = ["-mavx"],
3732 msvc_copts = xnnpack_msvc_std_copts(),
3733 msvc_x86_32_copts = ["/arch:AVX"],
3734 msvc_x86_64_copts = ["/arch:AVX"],
3735 x86_srcs = AVX_UKERNELS,
3736 deps = [
3737 ":tables",
3738 "@FP16",
3739 "@pthreadpool",
3740 ],
3741)
3742
3743xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003744 name = "xop_ukernels",
3745 hdrs = INTERNAL_HDRS,
3746 gcc_copts = xnnpack_gcc_std_copts(),
3747 gcc_x86_copts = ["-mxop"],
3748 msvc_copts = xnnpack_msvc_std_copts(),
3749 msvc_x86_32_copts = ["/arch:AVX"],
3750 msvc_x86_64_copts = ["/arch:AVX"],
3751 x86_srcs = XOP_UKERNELS,
3752 deps = [
3753 ":tables",
3754 "@FP16",
3755 "@pthreadpool",
3756 ],
3757)
3758
3759xnnpack_cc_library(
3760 name = "xop_ukernels_test_mode",
3761 hdrs = INTERNAL_HDRS,
3762 copts = [
3763 "-UNDEBUG",
3764 "-DXNN_TEST_MODE=1",
3765 ],
3766 gcc_copts = xnnpack_gcc_std_copts(),
3767 gcc_x86_copts = ["-mxop"],
3768 msvc_copts = xnnpack_msvc_std_copts(),
3769 msvc_x86_32_copts = ["/arch:AVX"],
3770 msvc_x86_64_copts = ["/arch:AVX"],
3771 x86_srcs = XOP_UKERNELS,
3772 deps = [
3773 ":tables",
3774 "@FP16",
3775 "@pthreadpool",
3776 ],
3777)
3778
3779xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003780 name = "fma3_ukernels",
3781 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003782 gcc_copts = xnnpack_gcc_std_copts(),
3783 gcc_x86_copts = ["-mfma"],
3784 msvc_copts = xnnpack_msvc_std_copts(),
3785 msvc_x86_32_copts = ["/arch:AVX"],
3786 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003787 x86_srcs = FMA3_UKERNELS,
3788 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003789 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003790 "@FP16",
3791 "@pthreadpool",
3792 ],
3793)
3794
3795xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003796 name = "fma3_ukernels_test_mode",
3797 hdrs = INTERNAL_HDRS,
3798 copts = [
3799 "-UNDEBUG",
3800 "-DXNN_TEST_MODE=1",
3801 ],
3802 gcc_copts = xnnpack_gcc_std_copts(),
3803 gcc_x86_copts = ["-mfma"],
3804 msvc_copts = xnnpack_msvc_std_copts(),
3805 msvc_x86_32_copts = ["/arch:AVX"],
3806 msvc_x86_64_copts = ["/arch:AVX"],
3807 x86_srcs = FMA3_UKERNELS,
3808 deps = [
3809 ":tables",
3810 "@FP16",
3811 "@pthreadpool",
3812 ],
3813)
3814
3815xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003816 name = "avx2_ukernels",
3817 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003818 gcc_copts = xnnpack_gcc_std_copts(),
3819 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003820 "-mfma",
3821 "-mavx2",
3822 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003823 msvc_copts = xnnpack_msvc_std_copts(),
3824 msvc_x86_32_copts = ["/arch:AVX2"],
3825 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003826 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003827 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003828 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003829 "@FP16",
3830 "@pthreadpool",
3831 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003832)
3833
3834xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003835 name = "avx2_ukernels_test_mode",
3836 hdrs = INTERNAL_HDRS,
3837 copts = [
3838 "-UNDEBUG",
3839 "-DXNN_TEST_MODE=1",
3840 ],
3841 gcc_copts = xnnpack_gcc_std_copts(),
3842 gcc_x86_copts = [
3843 "-mfma",
3844 "-mavx2",
3845 ],
3846 msvc_copts = xnnpack_msvc_std_copts(),
3847 msvc_x86_32_copts = ["/arch:AVX2"],
3848 msvc_x86_64_copts = ["/arch:AVX2"],
3849 x86_srcs = AVX2_UKERNELS,
3850 deps = [
3851 ":tables",
3852 "@FP16",
3853 "@pthreadpool",
3854 ],
3855)
3856
3857xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003858 name = "avx512f_ukernels",
3859 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003860 gcc_copts = xnnpack_gcc_std_copts(),
3861 gcc_x86_copts = ["-mavx512f"],
3862 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3863 msvc_copts = xnnpack_msvc_std_copts(),
3864 msvc_x86_32_copts = ["/arch:AVX512"],
3865 msvc_x86_64_copts = ["/arch:AVX512"],
3866 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003867 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003868 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003869 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003870 "@FP16",
3871 "@pthreadpool",
3872 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003873)
3874
3875xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003876 name = "avx512f_ukernels_test_mode",
3877 hdrs = INTERNAL_HDRS,
3878 copts = [
3879 "-UNDEBUG",
3880 "-DXNN_TEST_MODE=1",
3881 ],
3882 gcc_copts = xnnpack_gcc_std_copts(),
3883 gcc_x86_copts = ["-mavx512f"],
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 = AVX512F_UKERNELS,
3890 deps = [
3891 ":tables",
3892 "@FP16",
3893 "@pthreadpool",
3894 ],
3895)
3896
3897xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003898 name = "avx512skx_ukernels",
3899 hdrs = INTERNAL_HDRS,
3900 gcc_copts = xnnpack_gcc_std_copts(),
3901 gcc_x86_copts = [
3902 "-mavx512f",
3903 "-mavx512cd",
3904 "-mavx512bw",
3905 "-mavx512dq",
3906 "-mavx512vl",
3907 ],
3908 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3909 msvc_copts = xnnpack_msvc_std_copts(),
3910 msvc_x86_32_copts = ["/arch:AVX512"],
3911 msvc_x86_64_copts = ["/arch:AVX512"],
3912 msys_copts = ["-fno-asynchronous-unwind-tables"],
3913 x86_srcs = AVX512SKX_UKERNELS,
3914 deps = [
3915 ":tables",
3916 "@FP16",
3917 "@pthreadpool",
3918 ],
3919)
3920
3921xnnpack_cc_library(
3922 name = "avx512skx_ukernels_test_mode",
3923 hdrs = INTERNAL_HDRS,
3924 copts = [
3925 "-UNDEBUG",
3926 "-DXNN_TEST_MODE=1",
3927 ],
3928 gcc_copts = xnnpack_gcc_std_copts(),
3929 gcc_x86_copts = [
3930 "-mavx512f",
3931 "-mavx512cd",
3932 "-mavx512bw",
3933 "-mavx512dq",
3934 "-mavx512vl",
3935 ],
3936 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3937 msvc_copts = xnnpack_msvc_std_copts(),
3938 msvc_x86_32_copts = ["/arch:AVX512"],
3939 msvc_x86_64_copts = ["/arch:AVX512"],
3940 msys_copts = ["-fno-asynchronous-unwind-tables"],
3941 x86_srcs = AVX512SKX_UKERNELS,
3942 deps = [
3943 ":tables",
3944 "@FP16",
3945 "@pthreadpool",
3946 ],
3947)
3948
3949xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003950 name = "asm_ukernels",
3951 hdrs = ["src/xnnpack/assembly.h"],
3952 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07003953 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003954 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003955 wasm_srcs = WASM32_ASM_UKERNELS,
3956 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003957)
3958
Marat Dukhan3b59de22020-06-03 20:15:19 -07003959xnnpack_cc_library(
3960 name = "logging_utils",
3961 srcs = LOGGING_SRCS,
3962 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3963 copts = LOGGING_COPTS + [
3964 "-Isrc",
3965 "-Iinclude",
3966 ] + select({
3967 ":debug_build": [],
3968 "//conditions:default": xnnpack_min_size_copts(),
3969 }),
3970 gcc_copts = xnnpack_gcc_std_copts(),
3971 msvc_copts = xnnpack_msvc_std_copts(),
3972 visibility = xnnpack_visibility(),
3973 deps = [
3974 "@FP16",
3975 "@clog",
3976 "@pthreadpool",
3977 ],
3978)
3979
Marat Dukhan08c4a432019-10-03 09:29:21 -07003980xnnpack_aggregate_library(
3981 name = "ukernels",
3982 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003983 ":neon_ukernels",
3984 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003985 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003986 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003987 ":asm_ukernels",
3988 ],
3989 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003990 ":neon_ukernels",
3991 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003992 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003993 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003994 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003995 ":asm_ukernels",
3996 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003997 generic_deps = [
3998 ":scalar_ukernels",
3999 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07004000 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08004001 ":psimd_fastmath_ukernels",
4002 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004003 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004004 wasm_deps = [
4005 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004006 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004007 ],
4008 wasmsimd_deps = [
4009 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004010 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004011 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004012 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004013 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004014 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004015 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004016 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004017 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004018 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004019 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004020 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004021 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004022 ],
4023)
4024
Marat Dukhan33fcf782020-05-24 14:27:15 -07004025xnnpack_aggregate_library(
4026 name = "ukernels_test_mode",
4027 aarch32_deps = [
4028 ":neon_ukernels_test_mode",
4029 ":neonfma_ukernels_test_mode",
4030 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004031 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004032 ":asm_ukernels",
4033 ],
4034 aarch64_deps = [
4035 ":neon_ukernels_test_mode",
4036 ":neonfma_ukernels_test_mode",
4037 ":neonv8_ukernels_test_mode",
4038 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004039 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004040 ":asm_ukernels",
4041 ],
4042 generic_deps = [
4043 ":scalar_ukernels_test_mode",
4044 ],
4045 psimd_deps = [
4046 ":psimd_fastmath_ukernels_test_mode",
4047 ":psimd_accmath_ukernels_test_mode",
4048 ],
4049 wasm_deps = [
4050 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004051 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004052 ],
4053 wasmsimd_deps = [
4054 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004055 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004056 ],
4057 x86_deps = [
4058 ":sse2_ukernels_test_mode",
4059 ":ssse3_ukernels_test_mode",
4060 ":sse41_ukernels_test_mode",
4061 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004062 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004063 ":fma3_ukernels_test_mode",
4064 ":avx2_ukernels_test_mode",
4065 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004066 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004067 ],
4068)
4069
Marat Dukhan86df3842020-06-17 14:03:31 -07004070# This target must be compatible with Android NDK r17 and compile with Clang 6.
4071# It is similar to :ukernels target, but excludes microkernels which require
4072# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07004073xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07004074 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004075 aarch32_deps = [
4076 ":neon_ukernels",
4077 ":neonfma_ukernels",
4078 ":neonv8_ukernels",
4079 ":asm_ukernels",
4080 ],
4081 aarch64_deps = [
4082 ":neon_ukernels",
4083 ":neonfma_ukernels",
4084 ":neonv8_ukernels",
4085 ":asm_ukernels",
4086 ],
4087 generic_deps = [
4088 ":scalar_ukernels",
4089 ],
4090 psimd_deps = [
4091 ":psimd_fastmath_ukernels",
4092 ":psimd_accmath_ukernels",
4093 ],
4094 wasm_deps = [
4095 ":wasm_ukernels",
4096 ],
4097 wasmsimd_deps = [
4098 ":wasm_ukernels",
4099 ],
4100 x86_deps = [
4101 ":sse2_ukernels",
4102 ":ssse3_ukernels",
4103 ":sse41_ukernels",
4104 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004105 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004106 ":fma3_ukernels",
4107 ":avx2_ukernels",
4108 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004109 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004110 ],
4111)
4112
Marat Dukhan08c4a432019-10-03 09:29:21 -07004113xnnpack_cc_library(
4114 name = "im2col",
4115 srcs = ["src/im2col.c"],
4116 hdrs = [
4117 "src/xnnpack/common.h",
4118 "src/xnnpack/im2col.h",
4119 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004120 gcc_copts = xnnpack_gcc_std_copts(),
4121 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004122)
4123
4124xnnpack_cc_library(
4125 name = "indirection",
4126 srcs = ["src/indirection.c"],
4127 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004128 gcc_copts = xnnpack_gcc_std_copts(),
4129 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004130 deps = [
4131 "@FP16",
4132 "@FXdiv",
4133 "@pthreadpool",
4134 ],
4135)
4136
4137xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004138 name = "indirection_test_mode",
4139 srcs = ["src/indirection.c"],
4140 hdrs = INTERNAL_HDRS,
4141 copts = [
4142 "-UNDEBUG",
4143 "-DXNN_TEST_MODE=1",
4144 ],
4145 gcc_copts = xnnpack_gcc_std_copts(),
4146 msvc_copts = xnnpack_msvc_std_copts(),
4147 deps = [
4148 "@FP16",
4149 "@FXdiv",
4150 "@pthreadpool",
4151 ],
4152)
4153
4154xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004155 name = "packing",
4156 srcs = ["src/packing.c"],
4157 hdrs = INTERNAL_HDRS,
4158 gcc_copts = xnnpack_gcc_std_copts(),
4159 msvc_copts = xnnpack_msvc_std_copts(),
4160 deps = [
4161 "@FP16",
4162 "@FXdiv",
4163 "@pthreadpool",
4164 ],
4165)
4166
4167xnnpack_cc_library(
4168 name = "packing_test_mode",
4169 srcs = ["src/packing.c"],
4170 hdrs = INTERNAL_HDRS,
4171 copts = [
4172 "-UNDEBUG",
4173 "-DXNN_TEST_MODE=1",
4174 ],
4175 gcc_copts = xnnpack_gcc_std_copts(),
4176 msvc_copts = xnnpack_msvc_std_copts(),
4177 deps = [
4178 "@FP16",
4179 "@FXdiv",
4180 "@pthreadpool",
4181 ],
4182)
4183
4184xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004185 name = "operator_run",
4186 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004187 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004188 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004189 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4190 "//conditions:default": [],
4191 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004192 gcc_copts = xnnpack_gcc_std_copts(),
4193 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004194 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004195 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004196 "@FP16",
4197 "@FXdiv",
4198 "@clog",
4199 "@pthreadpool",
4200 ],
4201)
4202
Chao Mei6ddfc602020-05-13 22:29:36 -07004203xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004204 name = "operator_run_test_mode",
4205 srcs = ["src/operator-run.c"],
4206 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4207 copts = LOGGING_COPTS + [
4208 "-UNDEBUG",
4209 "-DXNN_TEST_MODE=1",
4210 ] + select({
4211 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4212 "//conditions:default": [],
4213 }),
4214 gcc_copts = xnnpack_gcc_std_copts(),
4215 msvc_copts = xnnpack_msvc_std_copts(),
4216 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004217 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004218 "@FP16",
4219 "@FXdiv",
4220 "@clog",
4221 "@pthreadpool",
4222 ],
4223)
4224
4225xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004226 name = "memory_planner",
4227 srcs = ["src/memory-planner.c"],
4228 hdrs = INTERNAL_HDRS,
4229 defines = select({
4230 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4231 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4232 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4233 }),
4234 gcc_copts = xnnpack_gcc_std_copts(),
4235 msvc_copts = xnnpack_msvc_std_copts(),
4236 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004237 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004238 "@pthreadpool",
4239 ],
4240)
4241
Marat Dukhan33fcf782020-05-24 14:27:15 -07004242xnnpack_cc_library(
4243 name = "memory_planner_test_mode",
4244 srcs = ["src/memory-planner.c"],
4245 hdrs = INTERNAL_HDRS,
4246 copts = [
4247 "-UNDEBUG",
4248 "-DXNN_TEST_MODE=1",
4249 ],
4250 defines = select({
4251 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4252 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4253 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4254 }),
4255 gcc_copts = xnnpack_gcc_std_copts(),
4256 msvc_copts = xnnpack_msvc_std_copts(),
4257 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004258 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004259 "@pthreadpool",
4260 ],
4261)
4262
Marat Dukhan08c4a432019-10-03 09:29:21 -07004263cc_library(
4264 name = "enable_assembly",
4265 defines = select({
4266 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4267 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004268 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004269 }),
4270)
4271
Marat Dukhan9de90e02020-06-18 16:04:12 -07004272cc_library(
4273 name = "enable_sparse",
4274 defines = select({
4275 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4276 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07004277 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004278 }),
4279)
4280
Marat Dukhancf056b22019-10-07 10:26:29 -07004281xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004282 name = "operators",
4283 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004284 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004285 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004286 ],
4287 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004288 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004289 "-Isrc",
4290 "-Iinclude",
4291 ] + select({
4292 ":debug_build": [],
4293 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004294 }) + select({
4295 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4296 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004297 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004298 gcc_copts = xnnpack_gcc_std_copts(),
4299 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004300 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004301 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004302 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004303 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004304 "@FP16",
4305 "@FXdiv",
4306 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004307 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004308 ],
4309)
4310
Marat Dukhan10a38082020-04-17 03:58:35 -07004311xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004312 name = "operators_test_mode",
4313 srcs = OPERATOR_SRCS + [
4314 "src/memory.c",
4315 "src/operator-delete.c",
4316 ],
4317 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4318 copts = LOGGING_COPTS + [
4319 "-Isrc",
4320 "-Iinclude",
4321 "-UNDEBUG",
4322 "-DXNN_TEST_MODE=1",
4323 ] + select({
4324 ":debug_build": [],
4325 "//conditions:default": xnnpack_min_size_copts(),
4326 }) + select({
4327 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4328 "//conditions:default": [],
4329 }),
4330 gcc_copts = xnnpack_gcc_std_copts(),
4331 msvc_copts = xnnpack_msvc_std_copts(),
4332 deps = [
4333 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004334 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004335 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004336 "@FP16",
4337 "@FXdiv",
4338 "@clog",
4339 "@pthreadpool",
4340 ],
4341)
4342
4343xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004344 name = "XNNPACK",
4345 srcs = [
4346 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004347 "src/runtime.c",
4348 "src/subgraph.c",
4349 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004350 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004351 hdrs = ["include/xnnpack.h"],
4352 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004353 "-Isrc",
4354 "-Iinclude",
4355 ] + select({
4356 ":debug_build": [],
4357 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004358 }) + select({
4359 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4360 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004361 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004362 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004363 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004364 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004365 visibility = xnnpack_visibility(),
4366 deps = [
4367 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004368 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004369 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004370 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004371 ":operator_run",
4372 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004373 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004374 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004375 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004376 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004377 ] + select({
4378 ":emscripten": [],
4379 "//conditions:default": ["@cpuinfo"],
4380 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004381)
4382
Marat Dukhan10a38082020-04-17 03:58:35 -07004383xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004384 name = "XNNPACK_test_mode",
4385 srcs = [
4386 "src/init.c",
4387 "src/runtime.c",
4388 "src/subgraph.c",
4389 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004390 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004391 hdrs = ["include/xnnpack.h"],
4392 copts = LOGGING_COPTS + [
4393 "-Isrc",
4394 "-Iinclude",
4395 "-UNDEBUG",
4396 "-DXNN_TEST_MODE=1",
4397 ] + select({
4398 ":debug_build": [],
4399 "//conditions:default": xnnpack_min_size_copts(),
4400 }) + select({
4401 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4402 "//conditions:default": [],
4403 }),
4404 gcc_copts = xnnpack_gcc_std_copts(),
4405 includes = ["include"],
4406 msvc_copts = xnnpack_msvc_std_copts(),
4407 visibility = xnnpack_visibility(),
4408 deps = [
4409 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004410 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004411 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004412 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004413 ":operator_run_test_mode",
4414 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004415 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004416 "@clog",
4417 "@FP16",
4418 "@pthreadpool",
4419 ] + select({
4420 ":emscripten": [],
4421 "//conditions:default": ["@cpuinfo"],
4422 }),
4423)
4424
4425xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004426 name = "xnnpack_f32",
4427 srcs = [
4428 "src/init.c",
4429 "src/runtime.c",
4430 "src/subgraph.c",
4431 "src/tensor.c",
4432 ] + SUBGRAPH_SRCS,
4433 hdrs = ["include/xnnpack.h"],
4434 copts = LOGGING_COPTS + [
4435 "-Isrc",
4436 "-Iinclude",
4437 ] + select({
4438 ":debug_build": [],
4439 "//conditions:default": xnnpack_min_size_copts(),
4440 }) + select({
4441 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4442 "//conditions:default": [],
4443 }),
4444 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004445 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004446 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004447 "XNN_NO_U8_OPERATORS",
4448 "XNN_NO_X8_OPERATORS",
4449 "XNN_NO_F16_OPERATORS",
4450 "XNN_NO_X16_OPERATORS",
4451 ],
4452 gcc_copts = xnnpack_gcc_std_copts(),
4453 includes = ["include"],
4454 msvc_copts = xnnpack_msvc_std_copts(),
4455 visibility = xnnpack_visibility(),
4456 deps = [
4457 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004458 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004459 ":logging_utils",
4460 ":memory_planner",
4461 ":operator_run",
4462 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004463 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004464 "@clog",
4465 "@FP16",
4466 "@pthreadpool",
4467 ] + select({
4468 ":emscripten": [],
4469 "//conditions:default": ["@cpuinfo"],
4470 }),
4471)
4472
4473xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004474 name = "xnnpack_operators_nhwc_f32",
4475 srcs = [
4476 "src/init.c",
4477 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004478 hdrs = ["include/xnnpack.h"],
4479 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004480 "-Isrc",
4481 "-Iinclude",
4482 ] + select({
4483 ":debug_build": [],
4484 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004485 }) + select({
4486 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4487 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004488 }),
4489 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004490 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004491 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004492 "XNN_NO_U8_OPERATORS",
4493 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004494 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004495 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004496 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004497 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004498 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004499 visibility = xnnpack_visibility(),
4500 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004501 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004502 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004503 ":operator_run",
4504 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004505 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004506 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004507 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004508 ] + select({
4509 ":emscripten": [],
4510 "//conditions:default": ["@cpuinfo"],
4511 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004512)
4513
Marat Dukhancf056b22019-10-07 10:26:29 -07004514xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004515 name = "bench_utils",
4516 srcs = ["bench/utils.cc"],
4517 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004518 deps = [
4519 "@com_google_benchmark//:benchmark",
4520 "@cpuinfo",
4521 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004522)
4523
Frank Barchard7e955972019-10-11 10:34:25 -07004524######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004525
4526xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004527 name = "qs8_gemm_bench",
4528 srcs = [
4529 "bench/gemm.h",
4530 "bench/qs8-gemm.cc",
4531 "src/xnnpack/AlignedAllocator.h",
4532 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004533 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4534 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004535)
4536
4537xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004538 name = "qs8_requantization_bench",
4539 srcs = [
4540 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004541 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004542 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004543 ] + MICROKERNEL_BENCHMARK_HDRS,
4544 deps = MICROKERNEL_BENCHMARK_DEPS,
4545)
4546
4547xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004548 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004549 srcs = [
4550 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004551 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004552 "src/xnnpack/AlignedAllocator.h",
4553 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004554 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004555 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004556)
4557
4558xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004559 name = "qu8_requantization_bench",
4560 srcs = [
4561 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004562 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004563 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004564 ] + MICROKERNEL_BENCHMARK_HDRS,
4565 deps = MICROKERNEL_BENCHMARK_DEPS,
4566)
4567
4568xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004569 name = "f16_igemm_bench",
4570 srcs = [
4571 "bench/f16-igemm.cc",
4572 "bench/conv.h",
4573 "bench/google/conv.h",
4574 "src/xnnpack/AlignedAllocator.h",
4575 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004576 deps = MICROKERNEL_BENCHMARK_DEPS + [
4577 ":indirection",
4578 ":packing",
4579 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004580)
4581
4582xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004583 name = "f16_gemm_bench",
4584 srcs = [
4585 "bench/f16-gemm.cc",
4586 "bench/gemm.h",
4587 "src/xnnpack/AlignedAllocator.h",
4588 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004589 deps = MICROKERNEL_BENCHMARK_DEPS + [
4590 ":packing",
4591 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004592)
4593
4594xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004595 name = "f16_spmm_bench",
4596 srcs = [
4597 "bench/f16-spmm.cc",
4598 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004599 "src/xnnpack/AlignedAllocator.h",
4600 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004601 deps = MICROKERNEL_BENCHMARK_DEPS,
4602)
4603
4604xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004605 name = "f32_igemm_bench",
4606 srcs = [
4607 "bench/f32-igemm.cc",
4608 "bench/conv.h",
4609 "src/xnnpack/AlignedAllocator.h",
4610 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004611 deps = MICROKERNEL_BENCHMARK_DEPS + [
4612 ":indirection",
4613 ":packing",
4614 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004615)
4616
4617xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004618 name = "f16_relu_bench",
4619 srcs = [
4620 "bench/f16-relu.cc",
4621 "src/xnnpack/AlignedAllocator.h",
4622 ] + MICROKERNEL_BENCHMARK_HDRS,
4623 deps = MICROKERNEL_BENCHMARK_DEPS,
4624)
4625
4626xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004627 name = "f32_conv_hwc_bench",
4628 srcs = [
4629 "bench/f32-conv-hwc.cc",
4630 "bench/dconv.h",
4631 "src/xnnpack/AlignedAllocator.h",
4632 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004633 deps = MICROKERNEL_BENCHMARK_DEPS + [
4634 ":packing",
4635 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004636)
4637
4638xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004639 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004640 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004641 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004642 "bench/dconv.h",
4643 "src/xnnpack/AlignedAllocator.h",
4644 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004645 deps = MICROKERNEL_BENCHMARK_DEPS + [
4646 ":packing",
4647 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004648)
4649
4650xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004651 name = "f16_dwconv_bench",
4652 srcs = [
4653 "bench/f16-dwconv.cc",
4654 "bench/dwconv.h",
4655 "bench/google/dwconv.h",
4656 "src/xnnpack/AlignedAllocator.h",
4657 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004658 deps = MICROKERNEL_BENCHMARK_DEPS + [
4659 ":indirection",
4660 ":packing",
4661 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004662)
4663
4664xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004665 name = "f32_dwconv_bench",
4666 srcs = [
4667 "bench/f32-dwconv.cc",
4668 "bench/dwconv.h",
4669 "src/xnnpack/AlignedAllocator.h",
4670 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004671 deps = MICROKERNEL_BENCHMARK_DEPS + [
4672 ":indirection",
4673 ":packing",
4674 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004675)
4676
4677xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07004678 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004679 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07004680 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004681 "bench/dwconv.h",
4682 "src/xnnpack/AlignedAllocator.h",
4683 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004684 deps = MICROKERNEL_BENCHMARK_DEPS + [
4685 ":indirection",
4686 ":packing",
4687 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004688)
4689
4690xnnpack_benchmark(
4691 name = "f32_gemm_bench",
4692 srcs = [
4693 "bench/f32-gemm.cc",
4694 "bench/gemm.h",
4695 "src/xnnpack/AlignedAllocator.h",
4696 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004697 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004698 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004699)
4700
4701xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004702 name = "f32_hswish_bench",
4703 srcs = [
4704 "bench/f32-hswish.cc",
4705 "src/xnnpack/AlignedAllocator.h",
4706 ] + MICROKERNEL_BENCHMARK_HDRS,
4707 deps = MICROKERNEL_BENCHMARK_DEPS,
4708)
4709
4710xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004711 name = "f32_raddexpminusmax_bench",
4712 srcs = [
4713 "bench/f32-raddexpminusmax.cc",
4714 "src/xnnpack/AlignedAllocator.h",
4715 ] + MICROKERNEL_BENCHMARK_HDRS,
4716 deps = MICROKERNEL_BENCHMARK_DEPS,
4717)
4718
4719xnnpack_benchmark(
4720 name = "f32_raddextexp_bench",
4721 srcs = [
4722 "bench/f32-raddextexp.cc",
4723 "src/xnnpack/AlignedAllocator.h",
4724 ] + MICROKERNEL_BENCHMARK_HDRS,
4725 deps = MICROKERNEL_BENCHMARK_DEPS,
4726)
4727
4728xnnpack_benchmark(
4729 name = "f32_raddstoreexpminusmax_bench",
4730 srcs = [
4731 "bench/f32-raddstoreexpminusmax.cc",
4732 "src/xnnpack/AlignedAllocator.h",
4733 ] + MICROKERNEL_BENCHMARK_HDRS,
4734 deps = MICROKERNEL_BENCHMARK_DEPS,
4735)
4736
4737xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004738 name = "f32_relu_bench",
4739 srcs = [
4740 "bench/f32-relu.cc",
4741 "src/xnnpack/AlignedAllocator.h",
4742 ] + MICROKERNEL_BENCHMARK_HDRS,
4743 deps = MICROKERNEL_BENCHMARK_DEPS,
4744)
4745
4746xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004747 name = "f32_rmax_bench",
4748 srcs = [
4749 "bench/f32-rmax.cc",
4750 "src/xnnpack/AlignedAllocator.h",
4751 ] + MICROKERNEL_BENCHMARK_HDRS,
4752 deps = MICROKERNEL_BENCHMARK_DEPS,
4753)
4754
4755xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004756 name = "f32_sigmoid_bench",
4757 srcs = [
4758 "bench/f32-sigmoid.cc",
4759 "src/xnnpack/AlignedAllocator.h",
4760 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004761 deps = MICROKERNEL_BENCHMARK_DEPS,
4762)
4763
4764xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004765 name = "f32_spmm_bench",
4766 srcs = [
4767 "bench/f32-spmm.cc",
4768 "bench/gemm.h",
4769 "src/xnnpack/AlignedAllocator.h",
4770 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004771 deps = MICROKERNEL_BENCHMARK_DEPS,
4772)
4773
4774xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004775 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004776 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004777 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004778 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004779 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004780 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004781)
4782
4783xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004784 name = "f32_vscaleexpminusmax_bench",
4785 srcs = [
4786 "bench/f32-vscaleexpminusmax.cc",
4787 "src/xnnpack/AlignedAllocator.h",
4788 ] + MICROKERNEL_BENCHMARK_HDRS,
4789 deps = MICROKERNEL_BENCHMARK_DEPS,
4790)
4791
4792xnnpack_benchmark(
4793 name = "f32_vscaleextexp_bench",
4794 srcs = [
4795 "bench/f32-vscaleextexp.cc",
4796 "src/xnnpack/AlignedAllocator.h",
4797 ] + MICROKERNEL_BENCHMARK_HDRS,
4798 deps = MICROKERNEL_BENCHMARK_DEPS,
4799)
4800
4801xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004802 name = "f32_vsqrt_bench",
4803 srcs = [
4804 "bench/f32-vsqrt.cc",
4805 "src/xnnpack/AlignedAllocator.h",
4806 ] + MICROKERNEL_BENCHMARK_HDRS,
4807 deps = MICROKERNEL_BENCHMARK_DEPS,
4808)
4809
4810xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004811 name = "f32_im2col_gemm_bench",
4812 srcs = [
4813 "bench/f32-im2col-gemm.cc",
4814 "bench/conv.h",
4815 "src/xnnpack/AlignedAllocator.h",
4816 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004817 deps = MICROKERNEL_BENCHMARK_DEPS + [
4818 ":im2col",
4819 ":packing",
4820 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004821)
4822
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004823xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004824 name = "rounding_bench",
4825 srcs = [
4826 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004827 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004828 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004829 ] + MICROKERNEL_BENCHMARK_HDRS,
4830 deps = MICROKERNEL_BENCHMARK_DEPS,
4831)
4832
Marat Dukhan08c4a432019-10-03 09:29:21 -07004833########################### Benchmarks for operators ###########################
4834
4835xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004836 name = "average_pooling_bench",
4837 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004838 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004839 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004840 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004841)
4842
4843xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004844 name = "bankers_rounding_bench",
4845 srcs = ["bench/bankers-rounding.cc"],
4846 copts = xnnpack_optional_tflite_copts(),
4847 tags = ["nowin32"],
4848 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4849)
4850
4851xnnpack_benchmark(
4852 name = "ceiling_bench",
4853 srcs = ["bench/ceiling.cc"],
4854 copts = xnnpack_optional_tflite_copts(),
4855 tags = ["nowin32"],
4856 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4857)
4858
4859xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004860 name = "channel_shuffle_bench",
4861 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004862 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004863)
4864
4865xnnpack_benchmark(
4866 name = "convolution_bench",
4867 srcs = ["bench/convolution.cc"],
4868 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004869 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004870 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004871)
4872
4873xnnpack_benchmark(
4874 name = "deconvolution_bench",
4875 srcs = ["bench/deconvolution.cc"],
4876 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004877 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004878 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004879)
4880
4881xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004882 name = "floor_bench",
4883 srcs = ["bench/floor.cc"],
4884 copts = xnnpack_optional_tflite_copts(),
4885 tags = ["nowin32"],
4886 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4887)
4888
4889xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004890 name = "global_average_pooling_bench",
4891 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004892 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004893)
4894
4895xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004896 name = "hardswish_bench",
4897 srcs = ["bench/hardswish.cc"],
4898 copts = xnnpack_optional_tflite_copts(),
4899 tags = ["nowin32"],
4900 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4901)
4902
4903xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004904 name = "max_pooling_bench",
4905 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004906 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004907)
4908
4909xnnpack_benchmark(
4910 name = "sigmoid_bench",
4911 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004912 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004913 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004914 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004915)
4916
4917xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004918 name = "prelu_bench",
4919 srcs = ["bench/prelu.cc"],
4920 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004921 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004922 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004923)
4924
4925xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004926 name = "softmax_bench",
4927 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004928 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004929 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004930 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004931)
4932
Marat Dukhan87727142020-06-24 15:24:10 -07004933xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004934 name = "square_root_bench",
4935 srcs = ["bench/square-root.cc"],
4936 copts = xnnpack_optional_tflite_copts(),
4937 tags = ["nowin32"],
4938 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4939)
4940
4941xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004942 name = "truncation_bench",
4943 srcs = ["bench/truncation.cc"],
4944 deps = OPERATOR_BENCHMARK_DEPS,
4945)
4946
Marat Dukhanc068bb62019-10-04 13:24:39 -07004947############################# End-to-end benchmarks ############################
4948
4949cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004950 name = "fp32_mobilenet_v1",
4951 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004952 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004953 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004954 linkstatic = True,
4955 deps = [
4956 ":XNNPACK",
4957 "@pthreadpool",
4958 ],
4959)
4960
4961cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004962 name = "fp16_mobilenet_v1",
4963 srcs = ["models/fp16-mobilenet-v1.cc"],
4964 hdrs = ["models/models.h"],
4965 copts = xnnpack_std_cxxopts(),
4966 linkstatic = True,
4967 deps = [
4968 ":XNNPACK",
4969 "@FP16",
4970 "@pthreadpool",
4971 ],
4972)
4973
4974cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004975 name = "qs8_mobilenet_v1",
4976 srcs = ["models/qs8-mobilenet-v1.cc"],
4977 hdrs = ["models/models.h"],
4978 copts = xnnpack_std_cxxopts(),
4979 linkstatic = True,
4980 deps = [
4981 ":XNNPACK",
4982 "@pthreadpool",
4983 ],
4984)
4985
4986cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004987 name = "qs8_mobilenet_v2",
4988 srcs = ["models/qs8-mobilenet-v2.cc"],
4989 hdrs = ["models/models.h"],
4990 copts = xnnpack_std_cxxopts(),
4991 linkstatic = True,
4992 deps = [
4993 ":XNNPACK",
4994 "@pthreadpool",
4995 ],
4996)
4997
4998cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004999 name = "fp32_mobilenet_v2",
5000 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005001 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005002 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005003 linkstatic = True,
5004 deps = [
5005 ":XNNPACK",
5006 "@pthreadpool",
5007 ],
5008)
5009
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005010cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005011 name = "fp16_mobilenet_v2",
5012 srcs = ["models/fp16-mobilenet-v2.cc"],
5013 hdrs = ["models/models.h"],
5014 copts = xnnpack_std_cxxopts(),
5015 linkstatic = True,
5016 deps = [
5017 ":XNNPACK",
5018 "@FP16",
5019 "@pthreadpool",
5020 ],
5021)
5022
5023cc_library(
5024 name = "fp32_mobilenet_v3_large",
5025 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005026 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005027 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005028 linkstatic = True,
5029 deps = [
5030 ":XNNPACK",
5031 "@pthreadpool",
5032 ],
5033)
5034
5035cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005036 name = "fp16_mobilenet_v3_large",
5037 srcs = ["models/fp16-mobilenet-v3-large.cc"],
5038 hdrs = ["models/models.h"],
5039 copts = xnnpack_std_cxxopts(),
5040 linkstatic = True,
5041 deps = [
5042 ":XNNPACK",
5043 "@FP16",
5044 "@pthreadpool",
5045 ],
5046)
5047
5048cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005049 name = "fp32_mobilenet_v3_small",
5050 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005051 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005052 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005053 linkstatic = True,
5054 deps = [
5055 ":XNNPACK",
5056 "@pthreadpool",
5057 ],
5058)
5059
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005060cc_library(
5061 name = "fp16_mobilenet_v3_small",
5062 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5063 hdrs = ["models/models.h"],
5064 copts = xnnpack_std_cxxopts(),
5065 linkstatic = True,
5066 deps = [
5067 ":XNNPACK",
5068 "@FP16",
5069 "@pthreadpool",
5070 ],
5071)
5072
Marat Dukhanc068bb62019-10-04 13:24:39 -07005073xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005074 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005075 srcs = [
5076 "bench/f32-dwconv-e2e.cc",
5077 "bench/end2end.h",
5078 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005079 deps = MICROKERNEL_BENCHMARK_DEPS + [
5080 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005081 ":fp32_mobilenet_v1",
5082 ":fp32_mobilenet_v2",
5083 ":fp32_mobilenet_v3_large",
5084 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005085 ],
5086)
5087
5088xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005089 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005090 srcs = [
5091 "bench/f32-gemm-e2e.cc",
5092 "bench/end2end.h",
5093 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005094 deps = MICROKERNEL_BENCHMARK_DEPS + [
5095 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005096 ":fp32_mobilenet_v1",
5097 ":fp32_mobilenet_v2",
5098 ":fp32_mobilenet_v3_large",
5099 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005100 ],
5101)
5102
5103xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005104 name = "end2end_bench",
5105 srcs = ["bench/end2end.cc"],
5106 deps = [
5107 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005108 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005109 ":fp16_mobilenet_v1",
5110 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005111 ":fp16_mobilenet_v3_large",
5112 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005113 ":fp32_mobilenet_v1",
5114 ":fp32_mobilenet_v2",
5115 ":fp32_mobilenet_v3_large",
5116 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005117 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005118 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005119 "@pthreadpool",
5120 ],
5121)
5122
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005123#################### Accuracy evaluation for math functions ####################
5124
5125xnnpack_benchmark(
5126 name = "f32_exp_eval",
5127 srcs = [
5128 "eval/f32-exp.cc",
5129 "src/xnnpack/AlignedAllocator.h",
5130 ] + ACCURACY_EVAL_HDRS,
5131 deps = ACCURACY_EVAL_DEPS,
5132)
5133
Marat Dukhan515c9772019-10-17 18:07:57 -07005134xnnpack_benchmark(
5135 name = "f32_expminus_eval",
5136 srcs = [
5137 "eval/f32-expminus.cc",
5138 "src/xnnpack/AlignedAllocator.h",
5139 ] + ACCURACY_EVAL_HDRS,
5140 deps = ACCURACY_EVAL_DEPS,
5141)
5142
Marat Dukhan98ba4412019-10-23 02:14:28 -07005143xnnpack_benchmark(
Marat Dukhana438aca2020-11-20 15:45:01 -08005144 name = "f32_expm1minus_eval",
5145 srcs = [
5146 "eval/f32-expm1minus.cc",
5147 "src/xnnpack/AlignedAllocator.h",
5148 ] + ACCURACY_EVAL_HDRS,
5149 deps = ACCURACY_EVAL_DEPS,
5150)
5151
5152xnnpack_benchmark(
Marat Dukhan98ba4412019-10-23 02:14:28 -07005153 name = "f32_extexp_eval",
5154 srcs = [
5155 "eval/f32-extexp.cc",
5156 "src/xnnpack/AlignedAllocator.h",
5157 ] + ACCURACY_EVAL_HDRS,
5158 deps = ACCURACY_EVAL_DEPS,
5159)
5160
Marat Dukhan8853b822020-05-07 12:19:01 -07005161xnnpack_unit_test(
5162 name = "f32_roundne_eval",
5163 srcs = [
5164 "eval/f32-roundne.cc",
5165 "src/xnnpack/AlignedAllocator.h",
5166 "src/xnnpack/math-stubs.h",
5167 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005168 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005169 deps = MICROKERNEL_TEST_DEPS,
5170)
5171
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005172xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005173 name = "f32_roundd_eval",
5174 srcs = [
5175 "eval/f32-roundd.cc",
5176 "src/xnnpack/AlignedAllocator.h",
5177 "src/xnnpack/math-stubs.h",
5178 ] + MICROKERNEL_TEST_HDRS,
5179 automatic = False,
5180 deps = MICROKERNEL_TEST_DEPS,
5181)
5182
5183xnnpack_unit_test(
5184 name = "f32_roundu_eval",
5185 srcs = [
5186 "eval/f32-roundu.cc",
5187 "src/xnnpack/AlignedAllocator.h",
5188 "src/xnnpack/math-stubs.h",
5189 ] + MICROKERNEL_TEST_HDRS,
5190 automatic = False,
5191 deps = MICROKERNEL_TEST_DEPS,
5192)
5193
5194xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005195 name = "f32_roundz_eval",
5196 srcs = [
5197 "eval/f32-roundz.cc",
5198 "src/xnnpack/AlignedAllocator.h",
5199 "src/xnnpack/math-stubs.h",
5200 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005201 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005202 deps = MICROKERNEL_TEST_DEPS,
5203)
5204
Marat Dukhan346a9e52019-11-15 09:06:30 -08005205xnnpack_benchmark(
5206 name = "f32_sigmoid_eval",
5207 srcs = [
5208 "eval/f32-sigmoid.cc",
5209 "src/xnnpack/AlignedAllocator.h",
5210 ] + ACCURACY_EVAL_HDRS,
5211 deps = ACCURACY_EVAL_DEPS,
5212)
5213
Marat Dukhan84000762020-06-29 18:38:43 -07005214xnnpack_benchmark(
5215 name = "f32_sqrt_eval",
5216 srcs = [
5217 "eval/f32-sqrt.cc",
5218 "src/xnnpack/AlignedAllocator.h",
5219 ] + ACCURACY_EVAL_HDRS,
5220 deps = ACCURACY_EVAL_DEPS + [
5221 ":bench_utils",
5222 ],
5223)
5224
Marat Dukhan08c4a432019-10-03 09:29:21 -07005225######################### Unit tests for micro-kernels #########################
5226
5227xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005228 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005229 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005230 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005231 "test/gemm-microkernel-tester.h",
5232 "src/xnnpack/AlignedAllocator.h",
5233 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005234 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005235)
5236
5237xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005238 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005239 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005240 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005241 "test/spmm-microkernel-tester.h",
5242 "src/xnnpack/AlignedAllocator.h",
5243 ] + MICROKERNEL_TEST_HDRS,
5244 deps = MICROKERNEL_TEST_DEPS,
5245)
5246
5247xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005248 name = "f32_argmaxpool_test",
5249 srcs = [
5250 "test/f32-argmaxpool.cc",
5251 "test/argmaxpool-microkernel-tester.h",
5252 "src/xnnpack/AlignedAllocator.h",
5253 ] + MICROKERNEL_TEST_HDRS,
5254 deps = MICROKERNEL_TEST_DEPS,
5255)
5256
5257xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005258 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005259 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005260 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005261 "test/avgpool-microkernel-tester.h",
5262 "src/xnnpack/AlignedAllocator.h",
5263 ] + MICROKERNEL_TEST_HDRS,
5264 deps = MICROKERNEL_TEST_DEPS,
5265)
5266
5267xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005268 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005269 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005270 "test/f32-ibilinear.cc",
5271 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005272 "src/xnnpack/AlignedAllocator.h",
5273 ] + MICROKERNEL_TEST_HDRS,
5274 deps = MICROKERNEL_TEST_DEPS,
5275)
5276
5277xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005278 name = "f32_ibilinear_chw_test",
5279 srcs = [
5280 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005281 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005282 "src/xnnpack/AlignedAllocator.h",
5283 ] + MICROKERNEL_TEST_HDRS,
5284 deps = MICROKERNEL_TEST_DEPS,
5285)
5286
5287xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005288 name = "f16_clamp_test",
5289 srcs = [
5290 "test/f16-clamp.cc",
5291 "test/clamp-microkernel-tester.h",
5292 ] + MICROKERNEL_TEST_HDRS,
5293 deps = MICROKERNEL_TEST_DEPS,
5294)
5295
5296xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005297 name = "f32_clamp_test",
5298 srcs = [
5299 "test/f32-clamp.cc",
5300 "test/clamp-microkernel-tester.h",
5301 ] + MICROKERNEL_TEST_HDRS,
5302 deps = MICROKERNEL_TEST_DEPS,
5303)
5304
5305xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005306 name = "f32_igemm_test",
5307 srcs = [
5308 "test/f32-igemm.cc",
5309 "test/gemm-microkernel-tester.h",
5310 "src/xnnpack/AlignedAllocator.h",
5311 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005312 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005313)
5314
5315xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005316 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005317 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005318 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005319 "test/gemm-microkernel-tester.h",
5320 "src/xnnpack/AlignedAllocator.h",
5321 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005322 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005323)
5324
5325xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005326 name = "f32_igemm_minmax_test",
5327 srcs = [
5328 "test/f32-igemm-minmax.cc",
5329 "test/gemm-microkernel-tester.h",
5330 "src/xnnpack/AlignedAllocator.h",
5331 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005332 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005333)
5334
5335xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005336 name = "f16_igemm_minmax_test",
5337 srcs = [
5338 "test/f16-igemm-minmax.cc",
5339 "test/gemm-microkernel-tester.h",
5340 "src/xnnpack/AlignedAllocator.h",
5341 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005342 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005343)
5344
5345xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005346 name = "f32_conv_hwc_test",
5347 srcs = [
5348 "test/f32-conv-hwc.cc",
5349 "test/conv-hwc-microkernel-tester.h",
5350 "src/xnnpack/AlignedAllocator.h",
5351 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005352 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005353)
5354
5355xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005356 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005357 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005358 "test/f32-conv-hwc2chw.cc",
5359 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005360 "src/xnnpack/AlignedAllocator.h",
5361 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005362 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005363)
5364
5365xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005366 name = "f32_dwconv_test",
5367 srcs = [
5368 "test/f32-dwconv.cc",
5369 "test/dwconv-microkernel-tester.h",
5370 "src/xnnpack/AlignedAllocator.h",
5371 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005372 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005373)
5374
5375xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005376 name = "f16_dwconv_minmax_test",
5377 srcs = [
5378 "test/f16-dwconv-minmax.cc",
5379 "test/dwconv-microkernel-tester.h",
5380 "src/xnnpack/AlignedAllocator.h",
5381 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005382 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005383)
5384
5385xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005386 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005387 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005388 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005389 "test/dwconv-microkernel-tester.h",
5390 "src/xnnpack/AlignedAllocator.h",
5391 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005392 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005393)
5394
5395xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005396 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005397 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005398 "test/f32-dwconv2d-chw.cc",
5399 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005400 "src/xnnpack/AlignedAllocator.h",
5401 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005402 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005403)
5404
5405xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005406 name = "f16_gavgpool_minmax_test",
5407 srcs = [
5408 "test/f16-gavgpool-minmax.cc",
5409 "test/gavgpool-microkernel-tester.h",
5410 "src/xnnpack/AlignedAllocator.h",
5411 ] + MICROKERNEL_TEST_HDRS,
5412 deps = MICROKERNEL_TEST_DEPS,
5413)
5414
5415xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005416 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005417 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005418 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005419 "test/gavgpool-microkernel-tester.h",
5420 "src/xnnpack/AlignedAllocator.h",
5421 ] + MICROKERNEL_TEST_HDRS,
5422 deps = MICROKERNEL_TEST_DEPS,
5423)
5424
5425xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005426 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005427 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005428 "test/f32-gavgpool-cw.cc",
5429 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005430 "src/xnnpack/AlignedAllocator.h",
5431 ] + MICROKERNEL_TEST_HDRS,
5432 deps = MICROKERNEL_TEST_DEPS,
5433)
5434
5435xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005436 name = "f32_gemm_test",
5437 srcs = [
5438 "test/f32-gemm.cc",
5439 "test/gemm-microkernel-tester.h",
5440 "src/xnnpack/AlignedAllocator.h",
5441 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005442 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005443)
5444
5445xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005446 name = "f32_gemm_relu_test",
5447 srcs = [
5448 "test/f32-gemm-relu.cc",
5449 "test/gemm-microkernel-tester.h",
5450 "src/xnnpack/AlignedAllocator.h",
5451 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005452 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005453)
5454
5455xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005456 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005457 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005458 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005459 "test/gemm-microkernel-tester.h",
5460 "src/xnnpack/AlignedAllocator.h",
5461 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005462 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005463)
5464
5465xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005466 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005467 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005468 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005469 "test/gemm-microkernel-tester.h",
5470 "src/xnnpack/AlignedAllocator.h",
5471 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005472 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005473)
5474
5475xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005476 name = "f16_hswish_test",
5477 srcs = [
5478 "test/f16-hswish.cc",
5479 "test/hswish-microkernel-tester.h",
5480 ] + MICROKERNEL_TEST_HDRS,
5481 deps = MICROKERNEL_TEST_DEPS,
5482)
5483
5484xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005485 name = "f32_hswish_test",
5486 srcs = [
5487 "test/f32-hswish.cc",
5488 "test/hswish-microkernel-tester.h",
5489 ] + MICROKERNEL_TEST_HDRS,
5490 deps = MICROKERNEL_TEST_DEPS,
5491)
5492
5493xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005494 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005495 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005496 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005497 "test/maxpool-microkernel-tester.h",
5498 ] + MICROKERNEL_TEST_HDRS,
5499 deps = MICROKERNEL_TEST_DEPS,
5500)
5501
5502xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005503 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005504 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005505 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005506 "test/avgpool-microkernel-tester.h",
5507 "src/xnnpack/AlignedAllocator.h",
5508 ] + MICROKERNEL_TEST_HDRS,
5509 deps = MICROKERNEL_TEST_DEPS,
5510)
5511
5512xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005513 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005514 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005515 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005516 "test/gemm-microkernel-tester.h",
5517 "src/xnnpack/AlignedAllocator.h",
5518 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005519 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005520)
5521
5522xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005523 name = "f16_prelu_test",
5524 srcs = [
5525 "test/f16-prelu.cc",
5526 "test/prelu-microkernel-tester.h",
5527 "src/xnnpack/AlignedAllocator.h",
5528 ] + MICROKERNEL_TEST_HDRS,
5529 deps = MICROKERNEL_TEST_DEPS,
5530)
5531
5532xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005533 name = "f32_prelu_test",
5534 srcs = [
5535 "test/f32-prelu.cc",
5536 "test/prelu-microkernel-tester.h",
5537 "src/xnnpack/AlignedAllocator.h",
5538 ] + MICROKERNEL_TEST_HDRS,
5539 deps = MICROKERNEL_TEST_DEPS,
5540)
5541
5542xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005543 name = "f32_raddexpminusmax_test",
5544 srcs = [
5545 "test/f32-raddexpminusmax.cc",
5546 "test/raddexpminusmax-microkernel-tester.h",
5547 ] + MICROKERNEL_TEST_HDRS,
5548 deps = MICROKERNEL_TEST_DEPS,
5549)
5550
5551xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005552 name = "f32_raddextexp_test",
5553 srcs = [
5554 "test/f32-raddextexp.cc",
5555 "test/raddextexp-microkernel-tester.h",
5556 ] + MICROKERNEL_TEST_HDRS,
5557 deps = MICROKERNEL_TEST_DEPS,
5558)
5559
5560xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005561 name = "f32_raddstoreexpminusmax_test",
5562 srcs = [
5563 "test/f32-raddstoreexpminusmax.cc",
5564 "test/raddstoreexpminusmax-microkernel-tester.h",
5565 ] + MICROKERNEL_TEST_HDRS,
5566 deps = MICROKERNEL_TEST_DEPS,
5567)
5568
5569xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005570 name = "f32_relu_test",
5571 srcs = [
5572 "test/f32-relu.cc",
5573 "test/vunary-microkernel-tester.h",
5574 ] + MICROKERNEL_TEST_HDRS,
5575 deps = MICROKERNEL_TEST_DEPS,
5576)
5577
5578xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005579 name = "f32_rmax_test",
5580 srcs = [
5581 "test/f32-rmax.cc",
5582 "test/rmax-microkernel-tester.h",
5583 ] + MICROKERNEL_TEST_HDRS,
5584 deps = MICROKERNEL_TEST_DEPS,
5585)
5586
5587xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005588 name = "f32_sigmoid_test",
5589 srcs = [
5590 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005591 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005592 ] + MICROKERNEL_TEST_HDRS,
5593 deps = MICROKERNEL_TEST_DEPS,
5594)
5595
5596xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005597 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005598 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005599 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005600 "test/spmm-microkernel-tester.h",
5601 "src/xnnpack/AlignedAllocator.h",
5602 ] + MICROKERNEL_TEST_HDRS,
5603 deps = MICROKERNEL_TEST_DEPS,
5604)
5605
5606xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005607 name = "f16_vadd_minmax_test",
5608 srcs = [
5609 "test/f16-vadd-minmax.cc",
5610 "test/vbinary-microkernel-tester.h",
5611 ] + MICROKERNEL_TEST_HDRS,
5612 deps = MICROKERNEL_TEST_DEPS,
5613)
5614
5615xnnpack_unit_test(
5616 name = "f16_vaddc_minmax_test",
5617 srcs = [
5618 "test/f16-vaddc-minmax.cc",
5619 "test/vbinaryc-microkernel-tester.h",
5620 ] + MICROKERNEL_TEST_HDRS,
5621 deps = MICROKERNEL_TEST_DEPS,
5622)
5623
5624xnnpack_unit_test(
5625 name = "f16_vdiv_minmax_test",
5626 srcs = [
5627 "test/f16-vdiv-minmax.cc",
5628 "test/vbinary-microkernel-tester.h",
5629 ] + MICROKERNEL_TEST_HDRS,
5630 deps = MICROKERNEL_TEST_DEPS,
5631)
5632
5633xnnpack_unit_test(
5634 name = "f16_vdivc_minmax_test",
5635 srcs = [
5636 "test/f16-vdivc-minmax.cc",
5637 "test/vbinaryc-microkernel-tester.h",
5638 ] + MICROKERNEL_TEST_HDRS,
5639 deps = MICROKERNEL_TEST_DEPS,
5640)
5641
5642xnnpack_unit_test(
5643 name = "f16_vrdivc_minmax_test",
5644 srcs = [
5645 "test/f16-vrdivc-minmax.cc",
5646 "test/vbinaryc-microkernel-tester.h",
5647 ] + MICROKERNEL_TEST_HDRS,
5648 deps = MICROKERNEL_TEST_DEPS,
5649)
5650
5651xnnpack_unit_test(
5652 name = "f16_vmax_test",
5653 srcs = [
5654 "test/f16-vmax.cc",
5655 "test/vbinary-microkernel-tester.h",
5656 ] + MICROKERNEL_TEST_HDRS,
5657 deps = MICROKERNEL_TEST_DEPS,
5658)
5659
5660xnnpack_unit_test(
5661 name = "f16_vmaxc_test",
5662 srcs = [
5663 "test/f16-vmaxc.cc",
5664 "test/vbinaryc-microkernel-tester.h",
5665 ] + MICROKERNEL_TEST_HDRS,
5666 deps = MICROKERNEL_TEST_DEPS,
5667)
5668
5669xnnpack_unit_test(
5670 name = "f16_vmin_test",
5671 srcs = [
5672 "test/f16-vmin.cc",
5673 "test/vbinary-microkernel-tester.h",
5674 ] + MICROKERNEL_TEST_HDRS,
5675 deps = MICROKERNEL_TEST_DEPS,
5676)
5677
5678xnnpack_unit_test(
5679 name = "f16_vminc_test",
5680 srcs = [
5681 "test/f16-vminc.cc",
5682 "test/vbinaryc-microkernel-tester.h",
5683 ] + MICROKERNEL_TEST_HDRS,
5684 deps = MICROKERNEL_TEST_DEPS,
5685)
5686
5687xnnpack_unit_test(
5688 name = "f16_vmul_minmax_test",
5689 srcs = [
5690 "test/f16-vmul-minmax.cc",
5691 "test/vbinary-microkernel-tester.h",
5692 ] + MICROKERNEL_TEST_HDRS,
5693 deps = MICROKERNEL_TEST_DEPS,
5694)
5695
5696xnnpack_unit_test(
5697 name = "f16_vmulc_minmax_test",
5698 srcs = [
5699 "test/f16-vmulc-minmax.cc",
5700 "test/vbinaryc-microkernel-tester.h",
5701 ] + MICROKERNEL_TEST_HDRS,
5702 deps = MICROKERNEL_TEST_DEPS,
5703)
5704
5705xnnpack_unit_test(
5706 name = "f16_vsub_minmax_test",
5707 srcs = [
5708 "test/f16-vsub-minmax.cc",
5709 "test/vbinary-microkernel-tester.h",
5710 ] + MICROKERNEL_TEST_HDRS,
5711 deps = MICROKERNEL_TEST_DEPS,
5712)
5713
5714xnnpack_unit_test(
5715 name = "f16_vsubc_minmax_test",
5716 srcs = [
5717 "test/f16-vsubc-minmax.cc",
5718 "test/vbinaryc-microkernel-tester.h",
5719 ] + MICROKERNEL_TEST_HDRS,
5720 deps = MICROKERNEL_TEST_DEPS,
5721)
5722
5723xnnpack_unit_test(
5724 name = "f16_vrsubc_minmax_test",
5725 srcs = [
5726 "test/f16-vrsubc-minmax.cc",
5727 "test/vbinaryc-microkernel-tester.h",
5728 ] + MICROKERNEL_TEST_HDRS,
5729 deps = MICROKERNEL_TEST_DEPS,
5730)
5731
5732xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005733 name = "f32_vabs_test",
5734 srcs = [
5735 "test/f32-vabs.cc",
5736 "test/vunary-microkernel-tester.h",
5737 ] + MICROKERNEL_TEST_HDRS,
5738 deps = MICROKERNEL_TEST_DEPS,
5739)
5740
5741xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005742 name = "f32_vadd_test",
5743 srcs = [
5744 "test/f32-vadd.cc",
5745 "test/vbinary-microkernel-tester.h",
5746 ] + MICROKERNEL_TEST_HDRS,
5747 deps = MICROKERNEL_TEST_DEPS,
5748)
5749
5750xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005751 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005752 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005753 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005754 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005755 ] + MICROKERNEL_TEST_HDRS,
5756 deps = MICROKERNEL_TEST_DEPS,
5757)
5758
5759xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005760 name = "f32_vadd_relu_test",
5761 srcs = [
5762 "test/f32-vadd-relu.cc",
5763 "test/vbinary-microkernel-tester.h",
5764 ] + MICROKERNEL_TEST_HDRS,
5765 deps = MICROKERNEL_TEST_DEPS,
5766)
5767
5768xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005769 name = "f32_vaddc_test",
5770 srcs = [
5771 "test/f32-vaddc.cc",
5772 "test/vbinaryc-microkernel-tester.h",
5773 ] + MICROKERNEL_TEST_HDRS,
5774 deps = MICROKERNEL_TEST_DEPS,
5775)
5776
5777xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005778 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005779 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005780 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005781 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005782 ] + MICROKERNEL_TEST_HDRS,
5783 deps = MICROKERNEL_TEST_DEPS,
5784)
5785
5786xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005787 name = "f32_vaddc_relu_test",
5788 srcs = [
5789 "test/f32-vaddc-relu.cc",
5790 "test/vbinaryc-microkernel-tester.h",
5791 ] + MICROKERNEL_TEST_HDRS,
5792 deps = MICROKERNEL_TEST_DEPS,
5793)
5794
5795xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005796 name = "f32_vdiv_test",
5797 srcs = [
5798 "test/f32-vdiv.cc",
5799 "test/vbinary-microkernel-tester.h",
5800 ] + MICROKERNEL_TEST_HDRS,
5801 deps = MICROKERNEL_TEST_DEPS,
5802)
5803
5804xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005805 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005806 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005807 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005808 "test/vbinary-microkernel-tester.h",
5809 ] + MICROKERNEL_TEST_HDRS,
5810 deps = MICROKERNEL_TEST_DEPS,
5811)
5812
5813xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005814 name = "f32_vdiv_relu_test",
5815 srcs = [
5816 "test/f32-vdiv-relu.cc",
5817 "test/vbinary-microkernel-tester.h",
5818 ] + MICROKERNEL_TEST_HDRS,
5819 deps = MICROKERNEL_TEST_DEPS,
5820)
5821
5822xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005823 name = "f32_vdivc_test",
5824 srcs = [
5825 "test/f32-vdivc.cc",
5826 "test/vbinaryc-microkernel-tester.h",
5827 ] + MICROKERNEL_TEST_HDRS,
5828 deps = MICROKERNEL_TEST_DEPS,
5829)
5830
5831xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005832 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005833 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005834 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005835 "test/vbinaryc-microkernel-tester.h",
5836 ] + MICROKERNEL_TEST_HDRS,
5837 deps = MICROKERNEL_TEST_DEPS,
5838)
5839
5840xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005841 name = "f32_vdivc_relu_test",
5842 srcs = [
5843 "test/f32-vdivc-relu.cc",
5844 "test/vbinaryc-microkernel-tester.h",
5845 ] + MICROKERNEL_TEST_HDRS,
5846 deps = MICROKERNEL_TEST_DEPS,
5847)
5848
5849xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005850 name = "f32_vrdivc_test",
5851 srcs = [
5852 "test/f32-vrdivc.cc",
5853 "test/vbinaryc-microkernel-tester.h",
5854 ] + MICROKERNEL_TEST_HDRS,
5855 deps = MICROKERNEL_TEST_DEPS,
5856)
5857
5858xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005859 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005860 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005861 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005862 "test/vbinaryc-microkernel-tester.h",
5863 ] + MICROKERNEL_TEST_HDRS,
5864 deps = MICROKERNEL_TEST_DEPS,
5865)
5866
5867xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005868 name = "f32_vrdivc_relu_test",
5869 srcs = [
5870 "test/f32-vrdivc-relu.cc",
5871 "test/vbinaryc-microkernel-tester.h",
5872 ] + MICROKERNEL_TEST_HDRS,
5873 deps = MICROKERNEL_TEST_DEPS,
5874)
5875
5876xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005877 name = "f32_vmax_test",
5878 srcs = [
5879 "test/f32-vmax.cc",
5880 "test/vbinary-microkernel-tester.h",
5881 ] + MICROKERNEL_TEST_HDRS,
5882 deps = MICROKERNEL_TEST_DEPS,
5883)
5884
5885xnnpack_unit_test(
5886 name = "f32_vmaxc_test",
5887 srcs = [
5888 "test/f32-vmaxc.cc",
5889 "test/vbinaryc-microkernel-tester.h",
5890 ] + MICROKERNEL_TEST_HDRS,
5891 deps = MICROKERNEL_TEST_DEPS,
5892)
5893
5894xnnpack_unit_test(
5895 name = "f32_vmin_test",
5896 srcs = [
5897 "test/f32-vmin.cc",
5898 "test/vbinary-microkernel-tester.h",
5899 ] + MICROKERNEL_TEST_HDRS,
5900 deps = MICROKERNEL_TEST_DEPS,
5901)
5902
5903xnnpack_unit_test(
5904 name = "f32_vminc_test",
5905 srcs = [
5906 "test/f32-vminc.cc",
5907 "test/vbinaryc-microkernel-tester.h",
5908 ] + MICROKERNEL_TEST_HDRS,
5909 deps = MICROKERNEL_TEST_DEPS,
5910)
5911
5912xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005913 name = "f32_vmul_test",
5914 srcs = [
5915 "test/f32-vmul.cc",
5916 "test/vbinary-microkernel-tester.h",
5917 ] + MICROKERNEL_TEST_HDRS,
5918 deps = MICROKERNEL_TEST_DEPS,
5919)
5920
5921xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005922 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005923 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005924 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005925 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005926 ] + MICROKERNEL_TEST_HDRS,
5927 deps = MICROKERNEL_TEST_DEPS,
5928)
5929
5930xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005931 name = "f32_vmul_relu_test",
5932 srcs = [
5933 "test/f32-vmul-relu.cc",
5934 "test/vbinary-microkernel-tester.h",
5935 ] + MICROKERNEL_TEST_HDRS,
5936 deps = MICROKERNEL_TEST_DEPS,
5937)
5938
5939xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005940 name = "f32_vmulc_test",
5941 srcs = [
5942 "test/f32-vmulc.cc",
5943 "test/vbinaryc-microkernel-tester.h",
5944 ] + MICROKERNEL_TEST_HDRS,
5945 deps = MICROKERNEL_TEST_DEPS,
5946)
5947
5948xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005949 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005950 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005951 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005952 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005953 ] + MICROKERNEL_TEST_HDRS,
5954 deps = MICROKERNEL_TEST_DEPS,
5955)
5956
5957xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005958 name = "f32_vmulc_relu_test",
5959 srcs = [
5960 "test/f32-vmulc-relu.cc",
5961 "test/vbinaryc-microkernel-tester.h",
5962 ] + MICROKERNEL_TEST_HDRS,
5963 deps = MICROKERNEL_TEST_DEPS,
5964)
5965
5966xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005967 name = "f16_vmulcaddc_minmax_test",
5968 srcs = [
5969 "test/f16-vmulcaddc-minmax.cc",
5970 "test/vmulcaddc-microkernel-tester.h",
5971 "src/xnnpack/AlignedAllocator.h",
5972 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005973 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005974)
5975
5976xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005977 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005978 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005979 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005980 "test/vmulcaddc-microkernel-tester.h",
5981 "src/xnnpack/AlignedAllocator.h",
5982 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005983 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005984)
5985
5986xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005987 name = "f32_vlrelu_test",
5988 srcs = [
5989 "test/f32-vlrelu.cc",
5990 "test/vunary-microkernel-tester.h",
5991 ] + MICROKERNEL_TEST_HDRS,
5992 deps = MICROKERNEL_TEST_DEPS,
5993)
5994
5995xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005996 name = "f32_vneg_test",
5997 srcs = [
5998 "test/f32-vneg.cc",
5999 "test/vunary-microkernel-tester.h",
6000 ] + MICROKERNEL_TEST_HDRS,
6001 deps = MICROKERNEL_TEST_DEPS,
6002)
6003
6004xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07006005 name = "f32_vrndne_test",
6006 srcs = [
6007 "test/f32-vrndne.cc",
6008 "test/vunary-microkernel-tester.h",
6009 ] + MICROKERNEL_TEST_HDRS,
6010 deps = MICROKERNEL_TEST_DEPS,
6011)
6012
6013xnnpack_unit_test(
6014 name = "f32_vrndz_test",
6015 srcs = [
6016 "test/f32-vrndz.cc",
6017 "test/vunary-microkernel-tester.h",
6018 ] + MICROKERNEL_TEST_HDRS,
6019 deps = MICROKERNEL_TEST_DEPS,
6020)
6021
6022xnnpack_unit_test(
6023 name = "f32_vrndu_test",
6024 srcs = [
6025 "test/f32-vrndu.cc",
6026 "test/vunary-microkernel-tester.h",
6027 ] + MICROKERNEL_TEST_HDRS,
6028 deps = MICROKERNEL_TEST_DEPS,
6029)
6030
6031xnnpack_unit_test(
6032 name = "f32_vrndd_test",
6033 srcs = [
6034 "test/f32-vrndd.cc",
6035 "test/vunary-microkernel-tester.h",
6036 ] + MICROKERNEL_TEST_HDRS,
6037 deps = MICROKERNEL_TEST_DEPS,
6038)
6039
6040xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07006041 name = "f32_vscale_test",
6042 srcs = [
6043 "test/f32-vscale.cc",
6044 "test/vscale-microkernel-tester.h",
6045 ] + MICROKERNEL_TEST_HDRS,
6046 deps = MICROKERNEL_TEST_DEPS,
6047)
6048
6049xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006050 name = "f32_vscaleexpminusmax_test",
6051 srcs = [
6052 "test/f32-vscaleexpminusmax.cc",
6053 "test/vscaleexpminusmax-microkernel-tester.h",
6054 ] + MICROKERNEL_TEST_HDRS,
6055 deps = MICROKERNEL_TEST_DEPS,
6056)
6057
6058xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006059 name = "f32_vscaleextexp_test",
6060 srcs = [
6061 "test/f32-vscaleextexp.cc",
6062 "test/vscaleextexp-microkernel-tester.h",
6063 ] + MICROKERNEL_TEST_HDRS,
6064 deps = MICROKERNEL_TEST_DEPS,
6065)
6066
6067xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006068 name = "f32_vsqr_test",
6069 srcs = [
6070 "test/f32-vsqr.cc",
6071 "test/vunary-microkernel-tester.h",
6072 ] + MICROKERNEL_TEST_HDRS,
6073 deps = MICROKERNEL_TEST_DEPS,
6074)
6075
6076xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006077 name = "f32_vsqrdiff_test",
6078 srcs = [
6079 "test/f32-vsqrdiff.cc",
6080 "test/vbinary-microkernel-tester.h",
6081 ] + MICROKERNEL_TEST_HDRS,
6082 deps = MICROKERNEL_TEST_DEPS,
6083)
6084
6085xnnpack_unit_test(
6086 name = "f32_vsqrdiffc_test",
6087 srcs = [
6088 "test/f32-vsqrdiffc.cc",
6089 "test/vbinaryc-microkernel-tester.h",
6090 ] + MICROKERNEL_TEST_HDRS,
6091 deps = MICROKERNEL_TEST_DEPS,
6092)
6093
6094xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006095 name = "f32_vsqrt_test",
6096 srcs = [
6097 "test/f32-vsqrt.cc",
6098 "test/vunary-microkernel-tester.h",
6099 ] + MICROKERNEL_TEST_HDRS,
6100 deps = MICROKERNEL_TEST_DEPS,
6101)
6102
6103xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006104 name = "f32_vsub_test",
6105 srcs = [
6106 "test/f32-vsub.cc",
6107 "test/vbinary-microkernel-tester.h",
6108 ] + MICROKERNEL_TEST_HDRS,
6109 deps = MICROKERNEL_TEST_DEPS,
6110)
6111
6112xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006113 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006114 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006115 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006116 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006117 ] + MICROKERNEL_TEST_HDRS,
6118 deps = MICROKERNEL_TEST_DEPS,
6119)
6120
6121xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006122 name = "f32_vsub_relu_test",
6123 srcs = [
6124 "test/f32-vsub-relu.cc",
6125 "test/vbinary-microkernel-tester.h",
6126 ] + MICROKERNEL_TEST_HDRS,
6127 deps = MICROKERNEL_TEST_DEPS,
6128)
6129
6130xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006131 name = "f32_vsubc_test",
6132 srcs = [
6133 "test/f32-vsubc.cc",
6134 "test/vbinaryc-microkernel-tester.h",
6135 ] + MICROKERNEL_TEST_HDRS,
6136 deps = MICROKERNEL_TEST_DEPS,
6137)
6138
6139xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006140 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006141 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006142 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006143 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006144 ] + MICROKERNEL_TEST_HDRS,
6145 deps = MICROKERNEL_TEST_DEPS,
6146)
6147
6148xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006149 name = "f32_vsubc_relu_test",
6150 srcs = [
6151 "test/f32-vsubc-relu.cc",
6152 "test/vbinaryc-microkernel-tester.h",
6153 ] + MICROKERNEL_TEST_HDRS,
6154 deps = MICROKERNEL_TEST_DEPS,
6155)
6156
6157xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006158 name = "f32_vrsubc_test",
6159 srcs = [
6160 "test/f32-vrsubc.cc",
6161 "test/vbinaryc-microkernel-tester.h",
6162 ] + MICROKERNEL_TEST_HDRS,
6163 deps = MICROKERNEL_TEST_DEPS,
6164)
6165
6166xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006167 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006168 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006169 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006170 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006171 ] + MICROKERNEL_TEST_HDRS,
6172 deps = MICROKERNEL_TEST_DEPS,
6173)
6174
6175xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006176 name = "f32_vrsubc_relu_test",
6177 srcs = [
6178 "test/f32-vrsubc-relu.cc",
6179 "test/vbinaryc-microkernel-tester.h",
6180 ] + MICROKERNEL_TEST_HDRS,
6181 deps = MICROKERNEL_TEST_DEPS,
6182)
6183
6184xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006185 name = "qs8_dwconv_minmax_test",
6186 srcs = [
6187 "test/qs8-dwconv-minmax.cc",
6188 "test/dwconv-microkernel-tester.h",
6189 "src/xnnpack/AlignedAllocator.h",
6190 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6191 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6192)
6193
6194xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006195 name = "qs8_gavgpool_minmax_test",
6196 srcs = [
6197 "test/qs8-gavgpool-minmax.cc",
6198 "test/gavgpool-microkernel-tester.h",
6199 "src/xnnpack/AlignedAllocator.h",
6200 ] + MICROKERNEL_TEST_HDRS,
6201 deps = MICROKERNEL_TEST_DEPS,
6202)
6203
6204xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006205 name = "qs8_gemm_minmax_test",
6206 srcs = [
6207 "test/qs8-gemm-minmax.cc",
6208 "test/gemm-microkernel-tester.h",
6209 "src/xnnpack/AlignedAllocator.h",
6210 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6211 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6212)
6213
6214xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006215 name = "qs8_igemm_minmax_test",
6216 srcs = [
6217 "test/qs8-igemm-minmax.cc",
6218 "test/gemm-microkernel-tester.h",
6219 "src/xnnpack/AlignedAllocator.h",
6220 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6221 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6222)
6223
6224xnnpack_unit_test(
6225 name = "qs8_requantization_test",
6226 srcs = [
6227 "src/xnnpack/requantization-stubs.h",
6228 "test/qs8-requantization.cc",
6229 "test/requantization-tester.h",
6230 ] + MICROKERNEL_TEST_HDRS,
6231 deps = MICROKERNEL_TEST_DEPS,
6232)
6233
6234xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006235 name = "qs8_vadd_minmax_test",
6236 srcs = [
6237 "test/qs8-vadd-minmax.cc",
6238 "test/vadd-microkernel-tester.h",
6239 ] + MICROKERNEL_TEST_HDRS,
6240 deps = MICROKERNEL_TEST_DEPS,
6241)
6242
6243xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006244 name = "qs8_vaddc_minmax_test",
6245 srcs = [
6246 "test/qs8-vaddc-minmax.cc",
6247 "test/vaddc-microkernel-tester.h",
6248 ] + MICROKERNEL_TEST_HDRS,
6249 deps = MICROKERNEL_TEST_DEPS,
6250)
6251
6252xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006253 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006254 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006255 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006256 "test/avgpool-microkernel-tester.h",
6257 "src/xnnpack/AlignedAllocator.h",
6258 ] + MICROKERNEL_TEST_HDRS,
6259 deps = MICROKERNEL_TEST_DEPS,
6260)
6261
6262xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006263 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006264 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006265 "test/qu8-dwconv-minmax.cc",
6266 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006267 "src/xnnpack/AlignedAllocator.h",
6268 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006269 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006270)
6271
6272xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006273 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006274 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006275 "test/qu8-igemm-minmax.cc",
6276 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006277 "src/xnnpack/AlignedAllocator.h",
6278 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006279 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006280)
6281
6282xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006283 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006284 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006285 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006286 "test/gavgpool-microkernel-tester.h",
6287 "src/xnnpack/AlignedAllocator.h",
6288 ] + MICROKERNEL_TEST_HDRS,
6289 deps = MICROKERNEL_TEST_DEPS,
6290)
6291
6292xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006293 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006294 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006295 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006296 "test/gemm-microkernel-tester.h",
6297 "src/xnnpack/AlignedAllocator.h",
6298 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006299 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006300)
6301
6302xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006303 name = "qu8_requantization_test",
6304 srcs = [
6305 "src/xnnpack/requantization-stubs.h",
6306 "test/qu8-requantization.cc",
6307 "test/requantization-tester.h",
6308 ] + MICROKERNEL_TEST_HDRS,
6309 deps = MICROKERNEL_TEST_DEPS,
6310)
6311
6312xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006313 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006314 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006315 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006316 "test/vadd-microkernel-tester.h",
6317 ] + MICROKERNEL_TEST_HDRS,
6318 deps = MICROKERNEL_TEST_DEPS,
6319)
6320
6321xnnpack_unit_test(
6322 name = "u8_clamp_test",
6323 srcs = [
6324 "test/u8-clamp.cc",
6325 "test/clamp-microkernel-tester.h",
6326 ] + MICROKERNEL_TEST_HDRS,
6327 deps = MICROKERNEL_TEST_DEPS,
6328)
6329
6330xnnpack_unit_test(
6331 name = "u8_lut32norm_test",
6332 srcs = [
6333 "test/u8-lut32norm.cc",
6334 "test/lut-norm-microkernel-tester.h",
6335 ] + MICROKERNEL_TEST_HDRS,
6336 deps = MICROKERNEL_TEST_DEPS,
6337)
6338
6339xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006340 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006341 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006342 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006343 "test/maxpool-microkernel-tester.h",
6344 ] + MICROKERNEL_TEST_HDRS,
6345 deps = MICROKERNEL_TEST_DEPS,
6346)
6347
6348xnnpack_unit_test(
6349 name = "u8_rmax_test",
6350 srcs = [
6351 "test/u8-rmax.cc",
6352 "test/rmax-microkernel-tester.h",
6353 ] + MICROKERNEL_TEST_HDRS,
6354 deps = MICROKERNEL_TEST_DEPS,
6355)
6356
6357xnnpack_unit_test(
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006358 name = "x32_depthtospace2d_chw2hwc_test",
Yury Kartynnike7841862020-11-04 18:22:18 -08006359 srcs = [
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006360 "test/x32-depthtospace2d-chw2hwc.cc",
6361 "test/depthtospace-microkernel-tester.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08006362 ] + MICROKERNEL_TEST_HDRS,
6363 deps = MICROKERNEL_TEST_DEPS,
6364)
6365
6366xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006367 name = "x32_fill_test",
6368 srcs = [
6369 "test/x32-fill.cc",
6370 "test/fill-microkernel-tester.h",
6371 ] + MICROKERNEL_TEST_HDRS,
6372 deps = MICROKERNEL_TEST_DEPS,
6373)
6374
6375xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006376 name = "x32_packx_test",
6377 srcs = [
6378 "test/x32-packx.cc",
6379 "test/pack-microkernel-tester.h",
6380 "src/xnnpack/AlignedAllocator.h",
6381 ] + MICROKERNEL_TEST_HDRS,
6382 deps = MICROKERNEL_TEST_DEPS,
6383)
6384
6385xnnpack_unit_test(
6386 name = "x32_pad_test",
6387 srcs = [
6388 "test/x32-pad.cc",
6389 "test/pad-microkernel-tester.h",
6390 ] + MICROKERNEL_TEST_HDRS,
6391 deps = MICROKERNEL_TEST_DEPS,
6392)
6393
6394xnnpack_unit_test(
6395 name = "x32_unpool_test",
6396 srcs = [
6397 "test/x32-unpool.cc",
6398 "test/unpool-microkernel-tester.h",
6399 ] + MICROKERNEL_TEST_HDRS,
6400 deps = MICROKERNEL_TEST_DEPS,
6401)
6402
6403xnnpack_unit_test(
6404 name = "x32_zip_test",
6405 srcs = [
6406 "test/x32-zip.cc",
6407 "test/zip-microkernel-tester.h",
6408 ] + MICROKERNEL_TEST_HDRS,
6409 deps = MICROKERNEL_TEST_DEPS,
6410)
6411
6412xnnpack_unit_test(
6413 name = "x8_lut_test",
6414 srcs = [
6415 "test/x8-lut.cc",
6416 "test/lut-microkernel-tester.h",
6417 ] + MICROKERNEL_TEST_HDRS,
6418 deps = MICROKERNEL_TEST_DEPS,
6419)
6420
6421xnnpack_unit_test(
6422 name = "x8_zip_test",
6423 srcs = [
6424 "test/x8-zip.cc",
6425 "test/zip-microkernel-tester.h",
6426 ] + MICROKERNEL_TEST_HDRS,
6427 deps = MICROKERNEL_TEST_DEPS,
6428)
6429
Marat Dukhan20c3b922020-03-10 03:45:06 -07006430########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006431
6432xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006433 name = "operator_size_test",
6434 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006435 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006436)
6437
Marat Dukhan20c3b922020-03-10 03:45:06 -07006438xnnpack_binary(
6439 name = "subgraph_size_test",
6440 srcs = ["test/subgraph-size.c"],
6441 deps = [":XNNPACK"],
6442)
6443
6444########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006445
6446xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006447 name = "abs_nc_test",
6448 srcs = [
6449 "test/abs-nc.cc",
6450 "test/abs-operator-tester.h",
6451 ],
6452 deps = OPERATOR_TEST_DEPS,
6453)
6454
6455xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006456 name = "add_nd_test",
6457 srcs = [
6458 "test/add-nd.cc",
6459 "test/binary-elementwise-operator-tester.h",
6460 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006461 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006462)
6463
6464xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006465 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006466 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006467 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006468 "test/argmax-pooling-operator-tester.h",
6469 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006470 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006471)
6472
6473xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006474 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006475 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006476 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006477 "test/average-pooling-operator-tester.h",
6478 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006479 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006480)
6481
6482xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006483 name = "bankers_rounding_nc_test",
6484 srcs = [
6485 "test/bankers-rounding-nc.cc",
6486 "test/bankers-rounding-operator-tester.h",
6487 ],
6488 deps = OPERATOR_TEST_DEPS,
6489)
6490
6491xnnpack_unit_test(
6492 name = "ceiling_nc_test",
6493 srcs = [
6494 "test/ceiling-nc.cc",
6495 "test/ceiling-operator-tester.h",
6496 ],
6497 deps = OPERATOR_TEST_DEPS,
6498)
6499
6500xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006501 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006502 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006503 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006504 "test/channel-shuffle-operator-tester.h",
6505 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006506 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006507)
6508
6509xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006510 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006511 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006512 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006513 "test/clamp-operator-tester.h",
6514 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006515 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006516)
6517
6518xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006519 name = "constant_pad_nd_test",
6520 srcs = [
6521 "test/constant-pad-nd.cc",
6522 "test/constant-pad-operator-tester.h",
6523 ],
6524 deps = OPERATOR_TEST_DEPS,
6525)
6526
6527xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006528 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006529 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006530 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006531 "test/convolution-operator-tester.h",
6532 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006533 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006534)
6535
6536xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006537 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006538 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006539 "test/convolution-nchw.cc",
6540 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006541 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006542 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006543)
6544
6545xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006546 name = "copy_nc_test",
6547 srcs = [
6548 "test/copy-nc.cc",
6549 "test/copy-operator-tester.h",
6550 ],
6551 deps = OPERATOR_TEST_DEPS,
6552)
6553
6554xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006555 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006556 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006557 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006558 "test/deconvolution-operator-tester.h",
6559 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006560 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006561)
6562
6563xnnpack_unit_test(
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08006564 name = "depth_to_space_test",
6565 srcs = [
6566 "test/depth-to-space.cc",
6567 "test/depth-to-space-operator-tester.h",
6568 ] + OPERATOR_TEST_PARAMS_HDRS,
6569 deps = OPERATOR_TEST_DEPS,
6570)
6571
6572xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006573 name = "divide_nd_test",
6574 srcs = [
6575 "test/binary-elementwise-operator-tester.h",
6576 "test/divide-nd.cc",
6577 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006578 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006579)
6580
6581xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006582 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006583 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006584 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006585 "test/fully-connected-operator-tester.h",
6586 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006587 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006588)
6589
6590xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006591 name = "floor_nc_test",
6592 srcs = [
6593 "test/floor-nc.cc",
6594 "test/floor-operator-tester.h",
6595 ],
6596 deps = OPERATOR_TEST_DEPS,
6597)
6598
6599xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006600 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006601 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006602 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006603 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006604 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006605 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006606)
6607
6608xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006609 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006610 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006611 "test/global-average-pooling-ncw.cc",
6612 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006613 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006614 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006615)
6616
6617xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006618 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006619 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006620 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006621 "test/hardswish-operator-tester.h",
6622 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006623 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006624)
6625
6626xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006627 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006628 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006629 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006630 "test/leaky-relu-operator-tester.h",
6631 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006632 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006633)
6634
6635xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006636 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006637 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006638 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006639 "test/max-pooling-operator-tester.h",
6640 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006641 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006642)
6643
6644xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006645 name = "maximum_nd_test",
6646 srcs = [
6647 "test/binary-elementwise-operator-tester.h",
6648 "test/maximum-nd.cc",
6649 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006650 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006651)
6652
6653xnnpack_unit_test(
6654 name = "minimum_nd_test",
6655 srcs = [
6656 "test/binary-elementwise-operator-tester.h",
6657 "test/minimum-nd.cc",
6658 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006659 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006660)
6661
6662xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006663 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006664 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006665 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006666 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006667 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006668 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006669)
6670
6671xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006672 name = "negate_nc_test",
6673 srcs = [
6674 "test/negate-nc.cc",
6675 "test/negate-operator-tester.h",
6676 ],
6677 deps = OPERATOR_TEST_DEPS,
6678)
6679
6680xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006681 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006682 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006683 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006684 "test/prelu-operator-tester.h",
6685 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006686 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006687)
6688
6689xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006690 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006691 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006692 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006693 "test/resize-bilinear-operator-tester.h",
6694 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006695 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006696)
6697
6698xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07006699 name = "resize_bilinear_nchw_test",
6700 srcs = [
6701 "test/resize-bilinear-nchw.cc",
6702 "test/resize-bilinear-operator-tester.h",
6703 ] + OPERATOR_TEST_PARAMS_HDRS,
6704 deps = OPERATOR_TEST_DEPS,
6705)
6706
6707xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006708 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006709 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006710 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006711 "test/sigmoid-operator-tester.h",
6712 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006713 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006714)
6715
6716xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006717 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006718 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006719 "test/softmax-nc.cc",
6720 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006721 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006722 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006723)
6724
6725xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006726 name = "square_nc_test",
6727 srcs = [
6728 "test/square-nc.cc",
6729 "test/square-operator-tester.h",
6730 ],
6731 deps = OPERATOR_TEST_DEPS,
6732)
6733
6734xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006735 name = "square_root_nc_test",
6736 srcs = [
6737 "test/square-root-nc.cc",
6738 "test/square-root-operator-tester.h",
6739 ],
6740 deps = OPERATOR_TEST_DEPS,
6741)
6742
6743xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006744 name = "squared_difference_nd_test",
6745 srcs = [
6746 "test/binary-elementwise-operator-tester.h",
6747 "test/squared-difference-nd.cc",
6748 ],
6749 deps = OPERATOR_TEST_DEPS,
6750)
6751
6752xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006753 name = "subtract_nd_test",
6754 srcs = [
6755 "test/binary-elementwise-operator-tester.h",
6756 "test/subtract-nd.cc",
6757 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006758 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006759)
6760
6761xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006762 name = "truncation_nc_test",
6763 srcs = [
6764 "test/truncation-nc.cc",
6765 "test/truncation-operator-tester.h",
6766 ],
6767 deps = OPERATOR_TEST_DEPS,
6768)
6769
6770xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006771 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006772 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006773 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006774 "test/unpooling-operator-tester.h",
6775 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006776 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006777)
6778
Chao Mei6ddfc602020-05-13 22:29:36 -07006779############################### Misc unit tests ###############################
6780
6781xnnpack_unit_test(
6782 name = "memory_planner_test",
6783 srcs = [
6784 "test/memory-planner-test.cc",
6785 ],
6786 deps = [
6787 ":XNNPACK",
6788 ":memory_planner",
6789 ],
6790)
6791
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07006792xnnpack_unit_test(
6793 name = "subgraph_nchw_test",
6794 srcs = [
6795 "src/xnnpack/subgraph.h",
6796 "test/subgraph-nchw.cc",
6797 "test/subgraph-tester.h",
6798 ],
6799 deps = [
6800 ":XNNPACK",
6801 ],
6802)
6803
Marat Dukhan08c4a432019-10-03 09:29:21 -07006804############################# Build configurations #############################
6805
Marat Dukhanb8642352019-10-30 15:43:02 -07006806# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006807config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006808 name = "xnn_enable_assembly_explicit_true",
6809 define_values = {"xnn_enable_assembly": "true"},
6810)
6811
6812# Disables usage of assembly kernels.
6813config_setting(
6814 name = "xnn_enable_assembly_explicit_false",
6815 define_values = {"xnn_enable_assembly": "false"},
6816)
6817
Marat Dukhan9de90e02020-06-18 16:04:12 -07006818# Enables usage of sparse inference.
6819config_setting(
6820 name = "xnn_enable_sparse_explicit_true",
6821 define_values = {"xnn_enable_sparse": "true"},
6822)
6823
6824# Disables usage of sparse inference.
6825config_setting(
6826 name = "xnn_enable_sparse_explicit_false",
6827 define_values = {"xnn_enable_sparse": "false"},
6828)
6829
Marat Dukhan05702cf2020-03-26 15:41:33 -07006830# Disables usage of HMP-aware optimizations.
6831config_setting(
6832 name = "xnn_enable_hmp_explicit_false",
6833 define_values = {"xnn_enable_hmp": "false"},
6834)
6835
Chao Mei6ddfc602020-05-13 22:29:36 -07006836# Enable usage of optimized memory allocation
6837config_setting(
6838 name = "xnn_enable_memopt_explicit_true",
6839 define_values = {"xnnpack_enable_memopt": "true"},
6840)
6841
6842# Disable usage of optimized memory allocation
6843config_setting(
6844 name = "xnn_enable_memopt_explicit_false",
6845 define_values = {"xnnpack_enable_memopt": "false"},
6846)
6847
Marat Dukhanb8642352019-10-30 15:43:02 -07006848# Builds with -c dbg
6849config_setting(
6850 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006851 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006852 "compilation_mode": "dbg",
6853 },
6854)
6855
6856# Builds with -c opt
6857config_setting(
6858 name = "optimized_build",
6859 values = {
6860 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006861 },
6862)
6863
6864config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006865 name = "linux_k8",
6866 values = {"cpu": "k8"},
6867)
6868
6869config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006870 name = "linux_arm",
6871 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006872)
6873
6874config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006875 name = "linux_armeabi",
6876 values = {"cpu": "armeabi"},
6877)
6878
6879config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006880 name = "linux_armhf",
6881 values = {"cpu": "armhf"},
6882)
6883
6884config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006885 name = "linux_armv7a",
6886 values = {"cpu": "armv7a"},
6887)
6888
6889config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006890 name = "linux_aarch64",
6891 values = {"cpu": "aarch64"},
6892)
6893
6894config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006895 name = "android",
6896 values = {"crosstool_top": "//external:android/crosstool"},
6897)
6898
6899config_setting(
6900 name = "android_armv7",
6901 values = {
6902 "crosstool_top": "//external:android/crosstool",
6903 "cpu": "armeabi-v7a",
6904 },
6905)
6906
6907config_setting(
6908 name = "android_arm64",
6909 values = {
6910 "crosstool_top": "//external:android/crosstool",
6911 "cpu": "arm64-v8a",
6912 },
6913)
6914
6915config_setting(
6916 name = "android_x86",
6917 values = {
6918 "crosstool_top": "//external:android/crosstool",
6919 "cpu": "x86",
6920 },
6921)
6922
6923config_setting(
6924 name = "android_x86_64",
6925 values = {
6926 "crosstool_top": "//external:android/crosstool",
6927 "cpu": "x86_64",
6928 },
6929)
6930
6931config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006932 name = "windows_x86_64",
6933 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006934)
6935
6936config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006937 name = "windows_x86_64_clang",
6938 values = {
6939 "compiler": "clang-cl",
6940 "cpu": "x64_windows",
6941 },
6942)
6943
6944config_setting(
6945 name = "windows_x86_64_mingw",
6946 values = {
6947 "compiler": "mingw-gcc",
6948 "cpu": "x64_windows",
6949 },
6950)
6951
6952config_setting(
6953 name = "windows_x86_64_msys",
6954 values = {
6955 "compiler": "msys-gcc",
6956 "cpu": "x64_windows",
6957 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006958)
6959
6960config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006961 name = "macos_x86_64",
6962 values = {
6963 "apple_platform_type": "macos",
6964 "cpu": "darwin",
6965 },
6966)
6967
6968config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006969 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006970 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006971)
6972
6973config_setting(
6974 name = "emscripten_wasm",
6975 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006976 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006977 "cpu": "wasm",
6978 },
6979)
6980
6981config_setting(
6982 name = "emscripten_wasmsimd",
6983 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006984 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006985 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006986 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006987 },
6988)
6989
6990config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006991 name = "ios_armv7",
6992 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006993 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006994 "cpu": "ios_armv7",
6995 },
6996)
6997
6998config_setting(
6999 name = "ios_arm64",
7000 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007001 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007002 "cpu": "ios_arm64",
7003 },
7004)
7005
7006config_setting(
7007 name = "ios_arm64e",
7008 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007009 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007010 "cpu": "ios_arm64e",
7011 },
7012)
7013
7014config_setting(
7015 name = "ios_x86",
7016 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007017 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007018 "cpu": "ios_i386",
7019 },
7020)
7021
7022config_setting(
7023 name = "ios_x86_64",
7024 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007025 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007026 "cpu": "ios_x86_64",
7027 },
7028)
7029
7030config_setting(
7031 name = "watchos_armv7k",
7032 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007033 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007034 "cpu": "watchos_armv7k",
7035 },
7036)
7037
7038config_setting(
7039 name = "watchos_arm64_32",
7040 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007041 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007042 "cpu": "watchos_arm64_32",
7043 },
7044)
7045
7046config_setting(
7047 name = "watchos_x86",
7048 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007049 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007050 "cpu": "watchos_i386",
7051 },
7052)
7053
7054config_setting(
7055 name = "watchos_x86_64",
7056 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007057 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007058 "cpu": "watchos_x86_64",
7059 },
7060)
7061
7062config_setting(
7063 name = "tvos_arm64",
7064 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007065 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007066 "cpu": "tvos_arm64",
7067 },
7068)
7069
7070config_setting(
7071 name = "tvos_x86_64",
7072 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007073 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007074 "cpu": "tvos_x86_64",
7075 },
7076)