blob: 5b4802be4c22391f7ca37786b3d23a76109dccd9 [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 Dukhanc60742b2020-11-23 12:33:27 -0800453 "src/math/expm1minus-scalar-rr2-lut16-p3.c",
454 "src/math/expm1minus-scalar-rr2-lut16-p4.c",
455 "src/math/expm1minus-scalar-rr2-p5.c",
456 "src/math/expm1minus-scalar-rr2-p6.c",
Frank Barchard22136062020-11-24 18:44:46 -0800457 "src/math/expminus-scalar-rr2-lut64-p2.c",
458 "src/math/expminus-scalar-rr2-lut2048-p1.c",
459 "src/math/expminus-scalar-rr2-p5.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 Dukhan048931b2020-11-24 20:53:54 -0800521 "src/xx-copy/memcpy.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700522]
523
Marat Dukhan436ebe62019-12-04 15:10:12 -0800524WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700525 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
526 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700527 "src/f32-clamp/gen/wasm-x1.c",
528 "src/f32-clamp/gen/wasm-x2.c",
529 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700530 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
531 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700532 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
533 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700534 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
535 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700536 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
537 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700538 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
539 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700540 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
541 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700542 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
543 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700544 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
545 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700546 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
547 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700548 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
549 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700550 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
551 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700552 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
553 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700554 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
555 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700556 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
557 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
558 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
559 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700560 "src/f32-gemm/gen/1x4-relu-wasm.c",
561 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700562 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700563 "src/f32-gemm/gen/2x4-relu-wasm.c",
564 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700565 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700566 "src/f32-gemm/gen/4x2-relu-wasm.c",
567 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700568 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700569 "src/f32-gemm/gen/4x4-relu-wasm.c",
570 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700571 "src/f32-hswish/gen/hswish-wasm-x1.c",
572 "src/f32-hswish/gen/hswish-wasm-x2.c",
573 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700574 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700575 "src/f32-igemm/gen/1x4-relu-wasm.c",
576 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700577 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700578 "src/f32-igemm/gen/2x4-relu-wasm.c",
579 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700580 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700581 "src/f32-igemm/gen/4x2-relu-wasm.c",
582 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700583 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700584 "src/f32-igemm/gen/4x4-relu-wasm.c",
585 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700586 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
587 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
588 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700589 "src/f32-prelu/gen/wasm-2x1.c",
590 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700591 "src/f32-relu/gen/wasm-x1.c",
592 "src/f32-relu/gen/wasm-x2.c",
593 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700594 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700595 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
596 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
597 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700598 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700599 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
600 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
601 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700602 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700603 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
604 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
605 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
606 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700607 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
608 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
609 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700610 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700611 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
612 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
613 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
614 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700615 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
616 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
617 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700618 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700619 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
620 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
621 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
622 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700623 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
624 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
625 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700626 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800627 "src/f32-vbinary/gen/vmax-wasm-x1.c",
628 "src/f32-vbinary/gen/vmax-wasm-x2.c",
629 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700630 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800631 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
632 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
633 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700634 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800635 "src/f32-vbinary/gen/vmin-wasm-x1.c",
636 "src/f32-vbinary/gen/vmin-wasm-x2.c",
637 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700638 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800639 "src/f32-vbinary/gen/vminc-wasm-x1.c",
640 "src/f32-vbinary/gen/vminc-wasm-x2.c",
641 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700642 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700643 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
644 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
645 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700646 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700647 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
648 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
649 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700650 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700651 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
652 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
653 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
654 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700655 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
656 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
657 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700658 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700659 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
660 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
661 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
662 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700663 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
664 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
665 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700666 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700667 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
668 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
669 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
670 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700671 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
672 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
673 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700674 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700675 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
676 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
677 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
678 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700679 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
680 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
681 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700682 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700683 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
684 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
685 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
686 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700687 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
688 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
689 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700690 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700691 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
692 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
693 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700694 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
695 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
696 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800697]
698
Marat Dukhan290055c2020-06-09 12:24:29 -0700699WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700700 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
701 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
702 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700703 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
704 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
705 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
706 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700707 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
708 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
709 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
710 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Frank Barchard22136062020-11-24 18:44:46 -0800711 "src/f32-conv-hwc2chw/3x3s2p1c3x4-wasmsimd-2x2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700712 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700713 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700714 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700715 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700716 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
717 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700718 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700719 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700720 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700721 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
722 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
723 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
724 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
725 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
726 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
727 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700728 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700729 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700730 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700731 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
732 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700733 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700734 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700735 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700736 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
737 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
738 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
739 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
740 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
741 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
Frank Barchard3b800452020-11-22 12:12:35 -0800742 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc2.c",
743 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc3.c",
744 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc4.c",
745 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4.c",
746 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-2x4-acc2.c",
747 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-2x4.c",
748 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-3x4.c",
749 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-4x4.c",
750 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-5x4.c",
751 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-6x4.c",
752 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc2.c",
753 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc3.c",
754 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc4.c",
755 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4.c",
756 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-2x4-acc2.c",
757 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-2x4.c",
758 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-3x4.c",
759 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-4x4.c",
760 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-5x4.c",
761 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-6x4.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -0800762 "src/f32-dwconv2d-chw/gen/3x3s2p1-wasmsimd-1x4-acc3-arm.c",
Frank Barcharddb5c32d2020-11-16 23:58:42 -0800763 "src/f32-dwconv2d-chw/gen/3x3s2p1-wasmsimd-1x4-acc3-x86.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -0800764 "src/f32-dwconv2d-chw/gen/5x5p2-wasmsimd-3x4-arm.c",
Frank Barcharddb5c32d2020-11-16 23:58:42 -0800765 "src/f32-dwconv2d-chw/gen/5x5p2-wasmsimd-3x4-x86.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -0800766 "src/f32-dwconv2d-chw/gen/5x5s2p2-wasmsimd-1x4-acc2-arm.c",
Frank Barcharddb5c32d2020-11-16 23:58:42 -0800767 "src/f32-dwconv2d-chw/gen/5x5s2p2-wasmsimd-1x4-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700768 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
769 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700770 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
771 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
772 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
773 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700774 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
775 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
776 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
777 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
778 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
779 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
780 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
781 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
782 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
783 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
784 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
785 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
786 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
787 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
788 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
789 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
790 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
791 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
792 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
793 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
794 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
795 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
796 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
797 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
798 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
799 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
800 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
801 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
802 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
803 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
804 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
805 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
806 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
807 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
808 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
809 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
810 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
811 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
812 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
813 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
814 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
815 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
816 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
817 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
818 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
819 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
820 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
821 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
822 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
823 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
824 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
825 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
826 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
827 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
828 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
829 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
830 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
831 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
832 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
833 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
834 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
835 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
836 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
837 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
838 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
839 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
840 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
841 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
842 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
843 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700844 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
845 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
846 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
XNNPACK Team965272b2020-10-23 21:10:15 -0700847 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
848 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700849 "src/f32-ibilinear/gen/wasmsimd-c4.c",
850 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700851 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
852 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
853 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
854 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
855 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
856 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
857 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
858 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
859 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
860 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
861 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
862 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
863 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
864 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
865 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
866 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
867 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
868 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
869 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
870 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
871 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
872 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
873 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
874 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
875 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
876 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
877 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
878 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
879 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
880 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
881 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
882 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
883 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
884 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
885 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
886 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
887 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
888 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
889 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
890 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700891 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
892 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700893 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
894 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
895 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
896 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700897 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
898 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700899 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
900 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
901 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700902 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
903 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700904 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
905 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
906 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
907 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
908 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
909 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
910 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700911 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
912 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700913 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
914 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
915 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
916 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700917 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700918 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700919 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700920 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
921 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700922 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700923 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
924 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700925 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700926 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
927 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700928 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700929 "src/f32-relu/gen/wasmsimd-x4.c",
930 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700931 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700932 "src/f32-rmax/wasmsimd-arm.c",
933 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700934 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
935 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
936 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
937 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
938 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
939 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700940 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
941 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
942 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
943 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
944 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
945 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800946 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined-x2.c",
947 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700948 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c",
949 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700950 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800951 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined-x2.c",
952 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700953 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c",
954 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700955 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800956 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined-x2.c",
957 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700958 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c",
959 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700960 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800961 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined-x2.c",
962 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700963 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c",
964 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700965 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800966 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined-x2.c",
967 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700968 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c",
969 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700970 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800971 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined-x2.c",
972 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700973 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c",
974 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700975 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800976 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined-x2.c",
977 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700978 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c",
979 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -0700980 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -0800981 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined-x2.c",
982 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -0700983 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c",
984 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -0700985 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700986 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
987 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700988 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700989 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
990 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700991 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700992 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
993 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700994 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700995 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
996 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700997 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700998 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
999 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001000 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001001 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
1002 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001003 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001004 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
1005 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001006 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001007 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
1008 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001009 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001010 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
1011 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001012 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001013 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
1014 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001015 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001016 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
1017 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001018 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001019 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
1020 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001021 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001022 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
1023 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001024 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001025 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
1026 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001027 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001028 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
1029 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001030 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001031 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
1032 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001033 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001034 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
1035 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001036 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001037 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
1038 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001039 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001040 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
1041 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001042 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001043 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
1044 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001045 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001046 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
1047 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001048 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001049 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
1050 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001051 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001052 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
1053 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001054 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001055 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
1056 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001057 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001058 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
1059 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001060 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001061 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
1062 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001063 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001064 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
1065 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001066 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001067 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
1068 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001069 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001070 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
1071 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001072 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001073 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
1074 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001075 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001076 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
1077 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001078 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001079 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
1080 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001081 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001082 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
1083 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001084 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001085 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
1086 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001087 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001088 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
1089 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001090 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001091 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
1092 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001093 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001094 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1095 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001096 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001097 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1098 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001099 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001100 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1101 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001102 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001103 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1104 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001105 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001106 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1107 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001108 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001109 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1110 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001111 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001112 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1113 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001114 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001115 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1116 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001117 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001118 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1119 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001120 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001121 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1122 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001123 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001124 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1125 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001126 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001127 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1128 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001129 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001130 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1131 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001132 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001133 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1134 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001135 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001136 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1137 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1138 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1139 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001140 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001141 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001142 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001143 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001144 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1145 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1146 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1147 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001148 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1149 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1150 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1151 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1152 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1153 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1154 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1155 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1156 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1157 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001158 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1159 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001160 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1161 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1162 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1163 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1164 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1165 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001166 "src/math/roundd-wasmsimd-addsub.c",
1167 "src/math/roundd-wasmsimd-cvt.c",
1168 "src/math/roundne-wasmsimd-addsub.c",
1169 "src/math/roundu-wasmsimd-addsub.c",
1170 "src/math/roundu-wasmsimd-cvt.c",
1171 "src/math/roundz-wasmsimd-addsub.c",
1172 "src/math/roundz-wasmsimd-cvt.c",
1173 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1174 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001175 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1176 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1177 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001178 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1179 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1180 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001181 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1182 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1183 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001184 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001185 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001186 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001187 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1188 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001189 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001190 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1191 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001192 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1193 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001194 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001195 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001196 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001197 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001198 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001199 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001200 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001201 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1202 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1203 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1204 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1205 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1206 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1207 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1208 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001209 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001210 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001211 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001212 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001213 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001214 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001215 "src/x32-zip/x2-wasmsimd.c",
1216 "src/x32-zip/x3-wasmsimd.c",
1217 "src/x32-zip/x4-wasmsimd.c",
1218 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001219]
1220
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001221PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001222 "src/qs8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001223 "src/qs8-requantization/precise-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001224 "src/qu8-requantization/fp32-psimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001225 "src/qu8-requantization/precise-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001226]
1227
Marat Dukhan08c4a432019-10-03 09:29:21 -07001228# ISA-specific micro-kernels
1229NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001230 "src/f32-argmaxpool/4x-neon-c4.c",
1231 "src/f32-argmaxpool/9p8x-neon-c4.c",
1232 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001233 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1234 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001235 "src/f32-clamp/gen/neon-x4.c",
1236 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001237 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001238 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001239 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001240 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001241 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001242 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001243 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001244 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001245 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001246 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001247 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001248 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001249 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001250 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1251 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1252 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1253 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1254 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001255 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001256 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001257 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1258 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1259 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001260 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001261 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001262 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1263 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1264 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1265 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1266 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001267 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1268 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1269 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001270 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001271 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001272 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1273 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1274 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001275 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1276 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1277 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1278 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001279 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001280 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1281 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001282 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001283 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001284 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001285 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001286 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1287 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001288 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1289 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1290 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1291 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1292 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1293 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1294 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1295 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001296 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001297 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001298 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001299 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1300 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001301 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001302 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1303 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001304 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001305 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1306 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1307 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1308 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1309 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001310 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1311 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001312 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1313 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001314 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1315 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001316 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1317 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1318 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1319 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1320 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1321 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1322 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1323 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1324 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1325 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1326 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1327 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1328 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1329 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1330 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1331 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001332 "src/f32-hswish/gen/hswish-neon-x4.c",
1333 "src/f32-hswish/gen/hswish-neon-x8.c",
1334 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001335 "src/f32-ibilinear/gen/neon-c4.c",
1336 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001337 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001338 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001339 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001340 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1341 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001342 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001343 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1344 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1345 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1346 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001347 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1348 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001349 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1350 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001351 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1352 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001353 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1354 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1355 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001356 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1357 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001358 "src/f32-prelu/gen/neon-1x4.c",
1359 "src/f32-prelu/gen/neon-1x8.c",
1360 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001361 "src/f32-prelu/gen/neon-2x4.c",
1362 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001363 "src/f32-prelu/gen/neon-2x16.c",
1364 "src/f32-prelu/gen/neon-4x4.c",
1365 "src/f32-prelu/gen/neon-4x8.c",
1366 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001367 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001368 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001369 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001370 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1371 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001372 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001373 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1374 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001375 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001376 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1377 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001378 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1379 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1380 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1381 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1382 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1383 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1384 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1385 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1386 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1387 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1388 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1389 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1390 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001391 "src/f32-relu/gen/neon-x4.c",
1392 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001393 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001394 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001395 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1396 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1397 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1398 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1399 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1400 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1401 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1402 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1403 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1404 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1405 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1406 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001407 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1408 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1409 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1410 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1411 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1412 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001413 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1414 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1415 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1416 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001417 "src/f32-vbinary/gen/vmax-neon-x4.c",
1418 "src/f32-vbinary/gen/vmax-neon-x8.c",
1419 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1420 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1421 "src/f32-vbinary/gen/vmin-neon-x4.c",
1422 "src/f32-vbinary/gen/vmin-neon-x8.c",
1423 "src/f32-vbinary/gen/vminc-neon-x4.c",
1424 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001425 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1426 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1427 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1428 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1429 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1430 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001431 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1432 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1433 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1434 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001435 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1436 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1437 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1438 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001439 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1440 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001441 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1442 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001443 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1444 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001445 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1446 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1447 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1448 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1449 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1450 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001451 "src/f32-vunary/gen/vabs-neon-x4.c",
1452 "src/f32-vunary/gen/vabs-neon-x8.c",
1453 "src/f32-vunary/gen/vneg-neon-x4.c",
1454 "src/f32-vunary/gen/vneg-neon-x8.c",
1455 "src/f32-vunary/gen/vsqr-neon-x4.c",
1456 "src/f32-vunary/gen/vsqr-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001457 "src/math/roundd-neon-addsub.c",
1458 "src/math/roundd-neon-cvt.c",
1459 "src/math/roundne-neon-addsub.c",
1460 "src/math/roundu-neon-addsub.c",
1461 "src/math/roundu-neon-cvt.c",
1462 "src/math/roundz-neon-addsub.c",
1463 "src/math/roundz-neon-cvt.c",
1464 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
1465 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1466 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1467 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1468 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1469 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1470 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1471 "src/math/sqrt-neon-nr1rsqrts.c",
1472 "src/math/sqrt-neon-nr2rsqrts.c",
1473 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001474 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1475 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1476 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1477 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001478 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1479 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1480 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1481 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001482 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1483 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1484 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1485 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001486 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001487 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001488 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001489 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1490 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001491 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001492 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001493 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001494 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001495 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001496 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001497 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1498 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1499 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1500 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1501 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1502 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1503 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1504 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001505 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1506 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1507 "src/qu8-dwconv/up8x9-minmax-neon.c",
1508 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1509 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1510 "src/qu8-gemm/4x8-minmax-neon.c",
1511 "src/qu8-gemm/8x8-minmax-neon.c",
1512 "src/qu8-igemm/4x8-minmax-neon.c",
1513 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001514 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001515 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001516 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001517 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001518 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001519 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001520 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001521 "src/x8-zip/x2-neon.c",
1522 "src/x8-zip/x3-neon.c",
1523 "src/x8-zip/x4-neon.c",
1524 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001525 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001526 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001527 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001528 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001529 "src/x32-zip/x2-neon.c",
1530 "src/x32-zip/x3-neon.c",
1531 "src/x32-zip/x4-neon.c",
1532 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001533]
1534
1535NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001536 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1537 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1538 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1539 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1540 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1541 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1542 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1543 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1544 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1545 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1546 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1547 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1548 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1549 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1550 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1551 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1552 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1553 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1554 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1555 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1556 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1557 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1558 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1559 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1560 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1561 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1562 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1563 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1564 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1565 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001566 "src/f32-ibilinear/gen/neonfma-c4.c",
1567 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001568 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001569 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001570 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001571 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1572 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001573 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1574 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001575 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1576 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001577 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1578 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001579 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001580 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001581 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001582 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1583 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001584 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001585 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1586 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001587 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001588 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1589 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001590 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1591 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1592 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1593 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1594 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1595 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1596 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1597 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1598 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1599 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1600 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1601 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1602 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001603 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1604 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1605 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1606 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1607 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1608 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001609 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1610 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1611 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1612 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1613 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1614 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001615 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1616 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1617 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1618 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1619 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1620 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001621 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1622 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1623 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1624 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1625 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1626 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001627 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1628 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1629 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1630 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1631 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1632 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001633 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1634 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1635 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1636 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1637 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1638 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001639 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1640 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1641 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1642 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1643 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1644 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1645 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1646 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1647 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1648 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1649 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1650 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1651 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1652 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1653 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1654 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1655 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1656 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001657 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1658 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001659 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1660 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1661 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1662 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1663 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1664 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1665 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1666 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1667 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1668 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1669 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1670 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1671 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1672 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1673 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1674 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1675 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1676 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1677 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1678 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08001679 "src/math/exp-neonfma-rr2-lut64-p2.c",
1680 "src/math/exp-neonfma-rr2-p5.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08001681 "src/math/expminus-neonfma-rr2-lut64-p2.c",
1682 "src/math/expminus-neonfma-rr2-lut2048-p1.c",
1683 "src/math/expminus-neonfma-rr2-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001684 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1685 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1686 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001687 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1688 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1689 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001690 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1691 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1692 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001693 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1694 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1695 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001696 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1697 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1698 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001699 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1700 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1701 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001702 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001703 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001704 "src/math/sqrt-neonfma-nr2fma.c",
1705 "src/math/sqrt-neonfma-nr2fma1adj.c",
1706 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001707]
1708
1709AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001710 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001711 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001712 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001713 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001714 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001715 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001716 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001717 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001718 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001719 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
1720 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
1721 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001722 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001723 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001724 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
1725 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
1726 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
1727 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
1728 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001729 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
1730 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
1731 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001732 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001733 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001734 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
1735 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
1736 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001737 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
1738 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
1739 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
1740 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001741 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001742 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
1743 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001744 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001745 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001746 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001747 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001748 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
1749 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001750 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
1751 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
1752 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
1753 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
1754 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
1755 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
1756 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
1757 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001758 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001759 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001760 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1761 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1762 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1763 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1764 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1765 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1766 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1767 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1768 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1769 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1770 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1771 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1772 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1773 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1774 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1775 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1776 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1777 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1778 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1779 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001780 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1781 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1782 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1783 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1784 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1785 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1786 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1787 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1788 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1789 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1790 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1791 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001792 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1793 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1794 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1795 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1796 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1797 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001798 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001799 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001800 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1801 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1802 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1803 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001804 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001805 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1806 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1807 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001808 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1809 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1810 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1811 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001812 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001813 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1814 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1815 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001816 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001817 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001818 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
1819 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
1820 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001821 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1822 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1823 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1824 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1825 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1826 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001827 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1828 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001829 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001830 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001831 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001832 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001833 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001834 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001835]
1836
Marat Dukhan8853b822020-05-07 12:19:01 -07001837NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001838 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1839 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001840 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1841 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1842 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1843 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1844 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1845 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001846 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001847 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001848 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001849 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001850]
1851
Marat Dukhan08c4a432019-10-03 09:29:21 -07001852AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001853 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001854 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001855 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1856 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1857 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1858 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001859 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1860 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1861 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1862 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1863 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1864 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1865 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1866 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001867 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1868 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001869 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1870 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1871 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1872 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1873 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1874 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1875 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
1876 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1877 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1878 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1879 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1880 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1881 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1882 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1883 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1884 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001885 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001886 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1887 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1888 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1889 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1890 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1891 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1892 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1893 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1894 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001895 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001896 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001897 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001898 "src/f16-relu/gen/neonfp16arith-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001899 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001900 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001901 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001902 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001903 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001904 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001905 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001906 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1907 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1908 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1909 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1910 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1911 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1912 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1913 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1914 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1915 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1916 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1917 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1918 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1919 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1920 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1921 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1922 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1923 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1924 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1925 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1926 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1927 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1928 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1929 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1930 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1931 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
1932 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1933 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1934 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1935 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
1936 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001937]
1938
Benoit Jacoba9644732020-08-13 12:48:55 -07001939NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07001940 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001941 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001942 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001943 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001944 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001945 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001946 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001947 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001948 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001949 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001950 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001951 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001952 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001953 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07001954 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001955 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07001956 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001957 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001958]
1959
Marat Dukhan08c4a432019-10-03 09:29:21 -07001960SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001961 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1962 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001963 "src/f32-clamp/gen/sse-x4.c",
1964 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001965 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1966 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001967 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1968 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1969 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1970 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001971 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1972 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001973 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1974 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1975 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1976 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001977 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1978 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001979 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
1980 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
1981 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07001982 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001983 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07001984 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
1985 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
1986 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
1987 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
1988 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07001989 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
1990 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
1991 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001992 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07001993 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001994 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
1995 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
1996 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07001997 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
1998 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
1999 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
2000 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
2001 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
2002 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
2003 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
2004 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
2005 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
2006 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
2007 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
2008 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
2009 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002010 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
2011 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
2012 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
2013 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
2014 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
2015 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
2016 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
2017 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002018 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08002019 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002020 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002021 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
2022 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002023 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
2024 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
2025 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
2026 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
2027 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
2028 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002029 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
2030 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
2031 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
2032 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
2033 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
2034 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
2035 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002036 "src/f32-hswish/gen/hswish-sse-x4.c",
2037 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07002038 "src/f32-ibilinear/gen/sse-c4.c",
2039 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002040 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
2041 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
2042 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
2043 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2044 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2045 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2046 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002047 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2048 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2049 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002050 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002051 "src/f32-prelu/gen/sse-2x4.c",
2052 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002053 "src/f32-relu/gen/sse-x4.c",
2054 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002055 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002056 "src/f32-spmm/gen/4x1-minmax-sse.c",
2057 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002058 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002059 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002060 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2061 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2062 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2063 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2064 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2065 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2066 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2067 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002068 "src/f32-vbinary/gen/vmax-sse-x4.c",
2069 "src/f32-vbinary/gen/vmax-sse-x8.c",
2070 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2071 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2072 "src/f32-vbinary/gen/vmin-sse-x4.c",
2073 "src/f32-vbinary/gen/vmin-sse-x8.c",
2074 "src/f32-vbinary/gen/vminc-sse-x4.c",
2075 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002076 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2077 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2078 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2079 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2080 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2081 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2082 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2083 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002084 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2085 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2086 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2087 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002088 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2089 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2090 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2091 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002092 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2093 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002094 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2095 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002096 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2097 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002098 "src/f32-vunary/gen/vabs-sse-x4.c",
2099 "src/f32-vunary/gen/vabs-sse-x8.c",
2100 "src/f32-vunary/gen/vneg-sse-x4.c",
2101 "src/f32-vunary/gen/vneg-sse-x8.c",
2102 "src/f32-vunary/gen/vsqr-sse-x4.c",
2103 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002104 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002105 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002106 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002107 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002108 "src/math/sqrt-sse-hh1mac.c",
2109 "src/math/sqrt-sse-nr1mac.c",
2110 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002111 "src/x32-fill/sse.c",
2112 "src/x32-packx/x4-sse.c",
2113 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002114]
2115
2116SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002117 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002118 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002119 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002120 "src/f32-prelu/gen/sse2-2x4.c",
2121 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002122 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002123 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002124 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002125 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2126 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002127 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002128 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2129 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002130 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002131 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2132 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002133 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002134 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
2135 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
2136 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
2137 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
2138 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
2139 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002140 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002141 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002142 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002143 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002144 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
2145 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002146 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2147 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002148 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2149 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002150 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2151 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2152 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2153 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2154 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2155 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002156 "src/math/exp-sse2-rr2-lut64-p2.c",
2157 "src/math/exp-sse2-rr2-p5.c",
Marat Dukhand6d321e2020-11-20 18:35:55 -08002158 "src/math/expm1minus-sse2-rr2-p5.c",
Marat Dukhana438aca2020-11-20 15:45:01 -08002159 "src/math/expm1minus-sse2-rr2-p6.c",
Frank Barchard3b800452020-11-22 12:12:35 -08002160 "src/math/expminus-sse2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002161 "src/math/roundd-sse2-cvt.c",
2162 "src/math/roundne-sse2-cvt.c",
2163 "src/math/roundu-sse2-cvt.c",
2164 "src/math/roundz-sse2-cvt.c",
2165 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2166 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2167 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2168 "src/math/sigmoid-sse2-rr2-p5-div.c",
2169 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2170 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002171 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2172 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2173 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002174 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2175 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2176 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002177 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2178 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2179 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002180 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002181 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002182 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002183 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002184 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002185 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002186 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2187 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002188 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002189 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2190 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002191 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002192 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2193 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2194 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002195 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002196 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002197 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002198 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002199 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002200 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002201 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002202 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002203 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2204 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002205 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002206 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002207 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002208 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2209 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2210 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2211 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002212 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2213 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2214 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2215 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002216 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2217 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002218 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2219 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2220 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2221 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2222 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002223 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002224 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002225 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002226 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002227 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002228 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002229 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002230 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002231 "src/x8-zip/x2-sse2.c",
2232 "src/x8-zip/x3-sse2.c",
2233 "src/x8-zip/x4-sse2.c",
2234 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002235 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002236 "src/x32-zip/x2-sse2.c",
2237 "src/x32-zip/x3-sse2.c",
2238 "src/x32-zip/x4-sse2.c",
2239 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002240]
2241
2242SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002243 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2244 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2245 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002246 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002247 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002248 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2249 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2250 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2251 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2252 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002253 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2254 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2255 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002256 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2257 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2258 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002259 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2260 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2261 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002262 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002263 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002264 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002265 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002266 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002267 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002268 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2269 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002270 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002271 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2272 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002273 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002274 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2275 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2276 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002277 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002278 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002279 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002280 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002281 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002282 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002283 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002284 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002285 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2286 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002287 "src/qs8-requantization/precise-ssse3.c",
2288 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002289 "src/qu8-requantization/precise-ssse3.c",
2290 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002291]
2292
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002293SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002294 "src/f32-prelu/gen/sse41-2x4.c",
2295 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002296 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2297 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2298 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2299 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2300 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2301 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002302 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002303 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002304 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002305 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002306 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2307 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002308 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2309 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002310 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2311 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002312 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2313 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2314 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2315 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2316 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2317 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2318 "src/math/roundd-sse41.c",
2319 "src/math/roundne-sse41.c",
2320 "src/math/roundu-sse41.c",
2321 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002322 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2323 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2324 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002325 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2326 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2327 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002328 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2329 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2330 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002331 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002332 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002333 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002334 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002335 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002336 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002337 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2338 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002339 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002340 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2341 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002342 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002343 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2344 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2345 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002346 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002347 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002348 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002349 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002350 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002351 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002352 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002353 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002354 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2355 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002356 "src/qs8-requantization/fp32-sse4.c",
2357 "src/qs8-requantization/precise-sse4.c",
2358 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002359 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2360 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2361 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2362 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002363 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2364 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2365 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2366 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002367 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2368 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2369 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2370 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002371 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2372 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2373 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2374 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002375 "src/qu8-requantization/precise-sse4.c",
2376 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002377]
2378
Marat Dukhan08c4a432019-10-03 09:29:21 -07002379AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002380 "src/f32-clamp/gen/avx-x8.c",
2381 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002382 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2383 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002384 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2385 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002386 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2387 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002388 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2389 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2390 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2391 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2392 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2393 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002394 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002395 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2396 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002397 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002398 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002399 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002400 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002401 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2402 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2403 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2404 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2405 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2406 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2407 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2408 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2409 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2410 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2411 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002412 "src/f32-hswish/gen/hswish-avx-x8.c",
2413 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002414 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002415 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2416 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002417 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002418 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002419 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002420 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002421 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2422 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002423 "src/f32-prelu/gen/avx-2x8.c",
2424 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002425 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002426 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002427 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002428 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2429 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2430 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2431 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2432 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2433 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2434 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2435 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2436 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2437 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2438 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2439 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2440 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2441 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2442 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2443 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2444 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2445 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2446 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2447 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002448 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2449 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2450 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2451 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2452 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2453 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2454 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2455 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002456 "src/f32-vbinary/gen/vmax-avx-x8.c",
2457 "src/f32-vbinary/gen/vmax-avx-x16.c",
2458 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2459 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2460 "src/f32-vbinary/gen/vmin-avx-x8.c",
2461 "src/f32-vbinary/gen/vmin-avx-x16.c",
2462 "src/f32-vbinary/gen/vminc-avx-x8.c",
2463 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002464 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2465 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2466 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2467 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2468 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2469 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2470 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2471 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002472 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2473 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2474 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2475 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002476 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2477 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2478 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2479 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002480 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2481 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002482 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2483 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002484 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2485 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2486 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2487 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2488 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2489 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002490 "src/f32-vscale/avx-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002491 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2492 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002493 "src/f32-vunary/gen/vabs-avx-x8.c",
2494 "src/f32-vunary/gen/vabs-avx-x16.c",
2495 "src/f32-vunary/gen/vneg-avx-x8.c",
2496 "src/f32-vunary/gen/vneg-avx-x16.c",
2497 "src/f32-vunary/gen/vsqr-avx-x8.c",
2498 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002499 "src/math/exp-avx-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002500 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002501 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002502 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2503 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002504]
2505
Marat Dukhan1566fee2020-08-02 21:55:41 -07002506XOP_UKERNELS = [
2507 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002508 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002509 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002510 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002511 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002512 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002513 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2514 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002515 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002516 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2517 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002518 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002519 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2520 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2521 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002522 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002523 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002524 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002525 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002526 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002527 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002528 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002529 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002530 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2531 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002532 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2533 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2534 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2535 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2536 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2537 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2538 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2539 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002540]
2541
Marat Dukhanfda12b82019-11-21 12:27:59 -08002542FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002543 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2544 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002545 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2546 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002547 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2548 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002549 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2550 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2551 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2552 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2553 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2554 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002555 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002556 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2557 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2558 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2559 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002560 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002561 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2562 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002563 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002564 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2565 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002566 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2567 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2568 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002569 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2570 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2571 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2572 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2573 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2574 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2575 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2576 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2577 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2578 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2579 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2580 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2581 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2582 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002583 "src/f32-hswish/gen/hswish-fma3-x8.c",
2584 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002585 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002586 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2587 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2588 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2589 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002590 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002591 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2592 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002593 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002594 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2595 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002596 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2597 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2598 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002599 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2600 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2601 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2602 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2603 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2604 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2605 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2606 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002607 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002608 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002609 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002610]
2611
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002612AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002613 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2614 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002615 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002616 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002617 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002618 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2619 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002620 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002621 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2622 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2623 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002624 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002625 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2626 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002627 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002628 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002629 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002630 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2631 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002632 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002633 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2634 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2635 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002636 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002637 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2638 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002639 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002640 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002641 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002642 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2643 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002644 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002645 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2646 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2647 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002648 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002649 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2650 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2651 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2652 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2653 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2654 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2655 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2656 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2657 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2658 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2659 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2660 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2661 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2662 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2663 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2664 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2665 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2666 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2667 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2668 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2669 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2670 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2671 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2672 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2673 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2674 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2675 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2676 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2677 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2678 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002679 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2680 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2681 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2682 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2683 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2684 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2685 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2686 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2687 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2688 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2689 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2690 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2691 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2692 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2693 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2694 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2695 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2696 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2697 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2698 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2699 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2700 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2701 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2702 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002703 "src/math/exp-avx2-rr2-lut8-p3-perm.c",
2704 "src/math/exp-avx2-rr2-lut8-p4-perm.c",
2705 "src/math/exp-avx2-rr2-p5.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08002706 "src/math/expminus-avx2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002707 "src/math/extexp-avx2-p5.c",
2708 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2709 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2710 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2711 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2712 "src/math/sigmoid-avx2-rr1-p5-div.c",
2713 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
2714 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2715 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2716 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2717 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2718 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2719 "src/math/sigmoid-avx2-rr2-p5-div.c",
2720 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2721 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002722 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002723 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002724 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2725 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002726 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002727 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002728 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002729 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002730 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002731 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002732 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002733 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002734 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2735 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2736 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002737 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2738 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2739 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2740 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2741 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2742 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2743 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2744 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002745]
2746
Marat Dukhan08c4a432019-10-03 09:29:21 -07002747AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002748 "src/f32-clamp/gen/avx512f-x16.c",
2749 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002750 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2751 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002752 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2753 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002754 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2755 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002756 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2757 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2758 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2759 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2760 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2761 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002762 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2763 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2764 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2765 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2766 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2767 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002768 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2769 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2770 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2771 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2772 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2773 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002774 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2775 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002776 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2777 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2778 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2779 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2780 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2781 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002782 "src/f32-prelu/gen/avx512f-2x16.c",
2783 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002784 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2785 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002786 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002787 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002788 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002789 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2790 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002791 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002792 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2793 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2794 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002795 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002796 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2797 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002798 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002799 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002800 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002801 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2802 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002803 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002804 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2805 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2806 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002807 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002808 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2809 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002810 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002811 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002812 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002813 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2814 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002815 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002816 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2817 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2818 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002819 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002820 "src/f32-relu/gen/avx512f-x16.c",
2821 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002822 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002823 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2824 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2825 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2826 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2827 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2828 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2829 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2830 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2831 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2832 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2833 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2834 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2835 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2836 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2837 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2838 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002839 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2840 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2841 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2842 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2843 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2844 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2845 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2846 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2847 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2848 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2849 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2850 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2851 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2852 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2853 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2854 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07002855 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2856 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2857 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2858 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2859 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2860 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2861 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2862 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2863 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2864 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2865 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2866 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2867 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2868 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2869 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2870 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002871 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2872 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2873 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2874 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2875 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2876 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2877 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2878 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002879 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2880 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2881 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2882 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2883 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2884 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2885 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2886 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002887 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2888 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2889 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2890 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2891 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2892 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2893 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2894 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002895 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2896 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2897 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2898 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002899 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2900 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2901 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2902 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002903 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2904 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002905 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2906 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
2907 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2908 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2909 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2910 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2911 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2912 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002913 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002914 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2915 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2916 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2917 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2918 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2919 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2920 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2921 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2922 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2923 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2924 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2925 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2926 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2927 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2928 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2929 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2930 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2931 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2932 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2933 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2934 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2935 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2936 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2937 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002938 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2939 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2940 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2941 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2942 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2943 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2944 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2945 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002946 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2947 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2948 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2949 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2950 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2951 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002952 "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
2953 "src/math/exp-avx512f-rr2-lut16-p3-perm.c",
2954 "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
2955 "src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
2956 "src/math/exp-avx512f-rr2-p5-scalef.c",
2957 "src/math/exp-avx512f-rr2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002958 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002959 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002960 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002961 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002962 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07002963 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002964 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002965 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002966 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002967 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
2968 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
2969 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
2970 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
2971 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2972 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2973 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2974 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2975 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
2976 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002977 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07002978 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002979 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
2980 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
2981 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
2982 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002983 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002984 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002985 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002986]
2987
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002988AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002989 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2990 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002991 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2992 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2993 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2994 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2995 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2996 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2997 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2998 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2999]
3000
Frank Barchardbcedc082020-08-17 18:00:51 -07003001WASM32_ASM_UKERNELS = [
3002 "src/f32-relu/wasm_shr_x1.S",
3003 "src/f32-relu/wasm_shr_x2.S",
3004 "src/f32-relu/wasm_shr_x4.S",
3005]
3006
Marat Dukhan08c4a432019-10-03 09:29:21 -07003007AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07003008 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07003009 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003010 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3011 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003012 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003013 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07003014 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003015 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003016 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3017 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003018 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
3019 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
3020 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
3021 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003022]
3023
3024AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003025 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003026 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003027 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003028 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003029 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003030 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07003031 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003032 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
3033 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
3034 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
3035 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
3036 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
3037 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
3038 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003039 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
3040 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003041 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3042 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3043 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003044 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
3045 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003046 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3047 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3048 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3049 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003050 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003051 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3052 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003053 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3054 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3055 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3056 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003057 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003058 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003059 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003060 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003061 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3062 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3063 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3064 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3065 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3066 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3067 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3068 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3069 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3070 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3071 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3072 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3073 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3074 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3075 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3076 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3077 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3078 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3079 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3080 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3081 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3082 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003083 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003084 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003085 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3086 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003087 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3088 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3089 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3090 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3091 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3092 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003093 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3094 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3095 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3096 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003097 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3098 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003099 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07003100 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07003101 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003102]
3103
Marat Dukhan1b354632020-03-23 12:50:22 -07003104INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003105 "src/xnnpack/argmaxpool.h",
3106 "src/xnnpack/avgpool.h",
3107 "src/xnnpack/clamp.h",
3108 "src/xnnpack/common.h",
3109 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003110 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003111 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003112 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003113 "src/xnnpack/gavgpool.h",
3114 "src/xnnpack/gemm.h",
3115 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003116 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003117 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003118 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003119 "src/xnnpack/lut.h",
3120 "src/xnnpack/math.h",
3121 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003122 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003123 "src/xnnpack/packx.h",
3124 "src/xnnpack/pad.h",
3125 "src/xnnpack/params.h",
3126 "src/xnnpack/pavgpool.h",
3127 "src/xnnpack/ppmm.h",
3128 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003129 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003130 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003131 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003132 "src/xnnpack/rmax.h",
3133 "src/xnnpack/scalar-utils.h",
3134 "src/xnnpack/spmm.h",
3135 "src/xnnpack/unpool.h",
3136 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003137 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003138 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003139 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003140 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003141 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003142 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003143 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003144]
3145
3146INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003147 "include/xnnpack.h",
3148 "src/xnnpack/allocator.h",
3149 "src/xnnpack/compute.h",
3150 "src/xnnpack/im2col.h",
3151 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003152 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003153 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003154 "src/xnnpack/operator.h",
3155 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003156 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003157 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003158 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003159 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003160]
3161
Marat Dukhan1b354632020-03-23 12:50:22 -07003162ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003163 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003164]
3165
Marat Dukhan1b354632020-03-23 12:50:22 -07003166MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003167 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003168 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003169]
3170
Marat Dukhan1b354632020-03-23 12:50:22 -07003171MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003172 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003173 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003174 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003175 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003176]
3177
3178OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003179 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003180 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003181]
3182
3183WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003184 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003185 "src/xnnpack/operator.h",
3186 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003187]
3188
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003189LOGGING_COPTS = select({
3190 # No logging in optimized mode
3191 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3192 # Full logging in debug mode
3193 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3194 # Error-only logging in default (fastbuild) mode
3195 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3196})
3197
Marat Dukhan3b59de22020-06-03 20:15:19 -07003198LOGGING_SRCS = select({
3199 # No logging in optimized mode
3200 ":optimized_build": [],
3201 "//conditions:default": [
3202 "src/operator-strings.c",
3203 "src/subgraph-strings.c",
3204 ],
3205})
3206
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003207LOGGING_HDRS = [
3208 "src/xnnpack/log.h",
3209]
3210
Marat Dukhan08c4a432019-10-03 09:29:21 -07003211xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003212 name = "tables",
3213 srcs = TABLE_SRCS,
3214 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003215 gcc_copts = xnnpack_gcc_std_copts(),
3216 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003217)
3218
3219xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003220 name = "scalar_ukernels",
3221 srcs = SCALAR_UKERNELS,
3222 hdrs = INTERNAL_HDRS,
3223 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003224 gcc_copts = xnnpack_gcc_std_copts(),
3225 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003226 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003227 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003228 "@FP16",
3229 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003230 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003231 ],
3232)
3233
3234xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003235 name = "scalar_ukernels_test_mode",
3236 srcs = SCALAR_UKERNELS,
3237 hdrs = INTERNAL_HDRS,
3238 aarch32_copts = ["-marm"],
3239 copts = [
3240 "-UNDEBUG",
3241 "-DXNN_TEST_MODE=1",
3242 ],
3243 gcc_copts = xnnpack_gcc_std_copts(),
3244 msvc_copts = xnnpack_msvc_std_copts(),
3245 deps = [
3246 ":tables",
3247 "@FP16",
3248 "@FXdiv",
3249 "@pthreadpool",
3250 ],
3251)
3252
3253xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003254 name = "wasm_ukernels",
3255 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003256 gcc_copts = xnnpack_gcc_std_copts(),
3257 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003258 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003259 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003260 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003261 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003262 "@FP16",
3263 "@FXdiv",
3264 "@pthreadpool",
3265 ],
3266)
3267
3268xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003269 name = "wasm_ukernels_test_mode",
3270 hdrs = INTERNAL_HDRS,
3271 copts = [
3272 "-UNDEBUG",
3273 "-DXNN_TEST_MODE=1",
3274 ],
3275 gcc_copts = xnnpack_gcc_std_copts(),
3276 msvc_copts = xnnpack_msvc_std_copts(),
3277 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003278 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003279 deps = [
3280 ":tables",
3281 "@FP16",
3282 "@FXdiv",
3283 "@pthreadpool",
3284 ],
3285)
3286
3287xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003288 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003289 hdrs = INTERNAL_HDRS,
3290 aarch32_copts = [
3291 "-marm",
3292 "-mfpu=neon",
3293 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003294 gcc_copts = xnnpack_gcc_std_copts(),
3295 gcc_x86_copts = ["-msse2"],
3296 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003297 optimized_copts = [
3298 "-O3",
3299 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003300 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003301 deps = [
3302 ":tables",
3303 "@FP16",
3304 "@psimd",
3305 "@pthreadpool",
3306 ],
3307)
3308
3309xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003310 name = "psimd_accmath_ukernels_test_mode",
3311 hdrs = INTERNAL_HDRS,
3312 aarch32_copts = [
3313 "-marm",
3314 "-mfpu=neon",
3315 ],
3316 copts = [
3317 "-UNDEBUG",
3318 "-DXNN_TEST_MODE=1",
3319 ],
3320 gcc_copts = xnnpack_gcc_std_copts(),
3321 gcc_x86_copts = ["-msse2"],
3322 msvc_copts = xnnpack_msvc_std_copts(),
3323 optimized_copts = [
3324 "-O3",
3325 ],
3326 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3327 deps = [
3328 ":tables",
3329 "@FP16",
3330 "@psimd",
3331 "@pthreadpool",
3332 ],
3333)
3334
3335xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003336 name = "neon_ukernels",
3337 hdrs = INTERNAL_HDRS,
3338 aarch32_copts = [
3339 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003340 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003341 "-mfpu=neon",
3342 ],
3343 aarch32_srcs = NEON_UKERNELS,
3344 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003345 gcc_copts = xnnpack_gcc_std_copts(),
3346 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003347 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003348 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003349 "@FP16",
3350 "@pthreadpool",
3351 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003352)
3353
3354xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003355 name = "neon_ukernels_test_mode",
3356 hdrs = INTERNAL_HDRS,
3357 aarch32_copts = [
3358 "-marm",
3359 "-march=armv7-a",
3360 "-mfpu=neon",
3361 ],
3362 aarch32_srcs = NEON_UKERNELS,
3363 aarch64_srcs = NEON_UKERNELS,
3364 copts = [
3365 "-UNDEBUG",
3366 "-DXNN_TEST_MODE=1",
3367 ],
3368 gcc_copts = xnnpack_gcc_std_copts(),
3369 msvc_copts = xnnpack_msvc_std_copts(),
3370 deps = [
3371 ":tables",
3372 "@FP16",
3373 "@pthreadpool",
3374 ],
3375)
3376
3377xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003378 name = "neonfma_ukernels",
3379 hdrs = INTERNAL_HDRS,
3380 aarch32_copts = [
3381 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003382 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003383 "-mfpu=neon-vfpv4",
3384 ],
3385 aarch32_srcs = NEONFMA_UKERNELS,
3386 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003387 apple_aarch32_copts = [
3388 "-mcpu=swift",
3389 "-mtune=generic",
3390 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003391 gcc_copts = xnnpack_gcc_std_copts(),
3392 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003393 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003394 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003395 "@FP16",
3396 "@pthreadpool",
3397 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003398)
3399
3400xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003401 name = "neonfma_ukernels_test_mode",
3402 hdrs = INTERNAL_HDRS,
3403 aarch32_copts = [
3404 "-marm",
3405 "-march=armv7-a",
3406 "-mfpu=neon-vfpv4",
3407 ],
3408 aarch32_srcs = NEONFMA_UKERNELS,
3409 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003410 apple_aarch32_copts = [
3411 "-mcpu=swift",
3412 "-mtune=generic",
3413 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003414 copts = [
3415 "-UNDEBUG",
3416 "-DXNN_TEST_MODE=1",
3417 ],
3418 gcc_copts = xnnpack_gcc_std_copts(),
3419 msvc_copts = xnnpack_msvc_std_copts(),
3420 deps = [
3421 ":tables",
3422 "@FP16",
3423 "@pthreadpool",
3424 ],
3425)
3426
3427xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003428 name = "neonv8_ukernels",
3429 hdrs = INTERNAL_HDRS,
3430 aarch32_copts = [
3431 "-marm",
3432 "-march=armv8-a",
3433 "-mfpu=neon-fp-armv8",
3434 ],
3435 aarch32_srcs = NEONV8_UKERNELS,
3436 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003437 apple_aarch32_copts = [
3438 "-mcpu=cyclone",
3439 "-mtune=generic",
3440 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003441 gcc_copts = xnnpack_gcc_std_copts(),
3442 msvc_copts = xnnpack_msvc_std_copts(),
3443 deps = [
3444 ":tables",
3445 "@FP16",
3446 "@pthreadpool",
3447 ],
3448)
3449
3450xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003451 name = "neonv8_ukernels_test_mode",
3452 hdrs = INTERNAL_HDRS,
3453 aarch32_copts = [
3454 "-marm",
3455 "-march=armv8-a",
3456 "-mfpu=neon-fp-armv8",
3457 ],
3458 aarch32_srcs = NEONV8_UKERNELS,
3459 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003460 apple_aarch32_copts = [
3461 "-mcpu=cyclone",
3462 "-mtune=generic",
3463 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003464 copts = [
3465 "-UNDEBUG",
3466 "-DXNN_TEST_MODE=1",
3467 ],
3468 gcc_copts = xnnpack_gcc_std_copts(),
3469 msvc_copts = xnnpack_msvc_std_copts(),
3470 deps = [
3471 ":tables",
3472 "@FP16",
3473 "@pthreadpool",
3474 ],
3475)
3476
3477xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003478 name = "neonfp16arith_ukernels",
3479 hdrs = INTERNAL_HDRS,
3480 aarch64_copts = ["-march=armv8.2-a+fp16"],
3481 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003482 gcc_copts = xnnpack_gcc_std_copts(),
3483 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003484 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003485 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003486 "@FP16",
3487 "@pthreadpool",
3488 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003489)
3490
3491xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003492 name = "neonfp16arith_ukernels_test_mode",
3493 hdrs = INTERNAL_HDRS,
3494 aarch64_copts = ["-march=armv8.2-a+fp16"],
3495 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3496 copts = [
3497 "-UNDEBUG",
3498 "-DXNN_TEST_MODE=1",
3499 ],
3500 gcc_copts = xnnpack_gcc_std_copts(),
3501 msvc_copts = xnnpack_msvc_std_copts(),
3502 deps = [
3503 ":tables",
3504 "@FP16",
3505 "@pthreadpool",
3506 ],
3507)
3508
3509xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003510 name = "neondot_ukernels",
3511 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003512 aarch32_copts = [
3513 "-marm",
3514 "-march=armv8.2-a+dotprod",
3515 "-mfpu=neon-fp-armv8",
3516 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003517 aarch32_srcs = NEONDOT_UKERNELS,
3518 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3519 aarch64_srcs = NEONDOT_UKERNELS,
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(
3530 name = "neondot_ukernels_test_mode",
3531 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003532 aarch32_copts = [
3533 "-marm",
3534 "-march=armv8.2-a+dotprod",
3535 "-mfpu=neon-fp-armv8",
3536 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003537 aarch32_srcs = NEONDOT_UKERNELS,
3538 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3539 aarch64_srcs = NEONDOT_UKERNELS,
3540 copts = [
3541 "-UNDEBUG",
3542 "-DXNN_TEST_MODE=1",
3543 ],
3544 gcc_copts = xnnpack_gcc_std_copts(),
3545 msvc_copts = xnnpack_msvc_std_copts(),
3546 deps = [
3547 ":tables",
3548 "@FP16",
3549 "@pthreadpool",
3550 ],
3551)
3552
3553xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003554 name = "sse2_ukernels",
3555 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003556 gcc_copts = xnnpack_gcc_std_copts(),
3557 gcc_x86_copts = ["-msse2"],
3558 msvc_copts = xnnpack_msvc_std_copts(),
3559 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003560 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003561 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003562 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003563 "@FP16",
3564 "@pthreadpool",
3565 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003566)
3567
3568xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003569 name = "sse2_ukernels_test_mode",
3570 hdrs = INTERNAL_HDRS,
3571 copts = [
3572 "-UNDEBUG",
3573 "-DXNN_TEST_MODE=1",
3574 ],
3575 gcc_copts = xnnpack_gcc_std_copts(),
3576 gcc_x86_copts = ["-msse2"],
3577 msvc_copts = xnnpack_msvc_std_copts(),
3578 msvc_x86_32_copts = ["/arch:SSE2"],
3579 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3580 deps = [
3581 ":tables",
3582 "@FP16",
3583 "@pthreadpool",
3584 ],
3585)
3586
3587xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003588 name = "ssse3_ukernels",
3589 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003590 gcc_copts = xnnpack_gcc_std_copts(),
3591 gcc_x86_copts = ["-mssse3"],
3592 msvc_copts = xnnpack_msvc_std_copts(),
3593 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003594 x86_srcs = SSSE3_UKERNELS,
3595 deps = [
3596 ":tables",
3597 "@FP16",
3598 "@pthreadpool",
3599 ],
3600)
3601
3602xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003603 name = "ssse3_ukernels_test_mode",
3604 hdrs = INTERNAL_HDRS,
3605 copts = [
3606 "-UNDEBUG",
3607 "-DXNN_TEST_MODE=1",
3608 ],
3609 gcc_copts = xnnpack_gcc_std_copts(),
3610 gcc_x86_copts = ["-mssse3"],
3611 msvc_copts = xnnpack_msvc_std_copts(),
3612 msvc_x86_32_copts = ["/arch:SSE2"],
3613 x86_srcs = SSSE3_UKERNELS,
3614 deps = [
3615 ":tables",
3616 "@FP16",
3617 "@pthreadpool",
3618 ],
3619)
3620
3621xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003622 name = "sse41_ukernels",
3623 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003624 gcc_copts = xnnpack_gcc_std_copts(),
3625 gcc_x86_copts = ["-msse4.1"],
3626 msvc_copts = xnnpack_msvc_std_copts(),
3627 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003628 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003629 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003630 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003631 "@FP16",
3632 "@pthreadpool",
3633 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003634)
3635
3636xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003637 name = "sse41_ukernels_test_mode",
3638 hdrs = INTERNAL_HDRS,
3639 copts = [
3640 "-UNDEBUG",
3641 "-DXNN_TEST_MODE=1",
3642 ],
3643 gcc_copts = xnnpack_gcc_std_copts(),
3644 gcc_x86_copts = ["-msse4.1"],
3645 msvc_copts = xnnpack_msvc_std_copts(),
3646 msvc_x86_32_copts = ["/arch:SSE2"],
3647 x86_srcs = SSE41_UKERNELS,
3648 deps = [
3649 ":tables",
3650 "@FP16",
3651 "@pthreadpool",
3652 ],
3653)
3654
3655xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003656 name = "avx_ukernels",
3657 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003658 gcc_copts = xnnpack_gcc_std_copts(),
3659 gcc_x86_copts = ["-mavx"],
3660 msvc_copts = xnnpack_msvc_std_copts(),
3661 msvc_x86_32_copts = ["/arch:AVX"],
3662 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003663 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003664 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003665 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003666 "@FP16",
3667 "@pthreadpool",
3668 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003669)
3670
3671xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003672 name = "avx_ukernels_test_mode",
3673 hdrs = INTERNAL_HDRS,
3674 copts = [
3675 "-UNDEBUG",
3676 "-DXNN_TEST_MODE=1",
3677 ],
3678 gcc_copts = xnnpack_gcc_std_copts(),
3679 gcc_x86_copts = ["-mavx"],
3680 msvc_copts = xnnpack_msvc_std_copts(),
3681 msvc_x86_32_copts = ["/arch:AVX"],
3682 msvc_x86_64_copts = ["/arch:AVX"],
3683 x86_srcs = AVX_UKERNELS,
3684 deps = [
3685 ":tables",
3686 "@FP16",
3687 "@pthreadpool",
3688 ],
3689)
3690
3691xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003692 name = "xop_ukernels",
3693 hdrs = INTERNAL_HDRS,
3694 gcc_copts = xnnpack_gcc_std_copts(),
3695 gcc_x86_copts = ["-mxop"],
3696 msvc_copts = xnnpack_msvc_std_copts(),
3697 msvc_x86_32_copts = ["/arch:AVX"],
3698 msvc_x86_64_copts = ["/arch:AVX"],
3699 x86_srcs = XOP_UKERNELS,
3700 deps = [
3701 ":tables",
3702 "@FP16",
3703 "@pthreadpool",
3704 ],
3705)
3706
3707xnnpack_cc_library(
3708 name = "xop_ukernels_test_mode",
3709 hdrs = INTERNAL_HDRS,
3710 copts = [
3711 "-UNDEBUG",
3712 "-DXNN_TEST_MODE=1",
3713 ],
3714 gcc_copts = xnnpack_gcc_std_copts(),
3715 gcc_x86_copts = ["-mxop"],
3716 msvc_copts = xnnpack_msvc_std_copts(),
3717 msvc_x86_32_copts = ["/arch:AVX"],
3718 msvc_x86_64_copts = ["/arch:AVX"],
3719 x86_srcs = XOP_UKERNELS,
3720 deps = [
3721 ":tables",
3722 "@FP16",
3723 "@pthreadpool",
3724 ],
3725)
3726
3727xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003728 name = "fma3_ukernels",
3729 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003730 gcc_copts = xnnpack_gcc_std_copts(),
3731 gcc_x86_copts = ["-mfma"],
3732 msvc_copts = xnnpack_msvc_std_copts(),
3733 msvc_x86_32_copts = ["/arch:AVX"],
3734 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003735 x86_srcs = FMA3_UKERNELS,
3736 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003737 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003738 "@FP16",
3739 "@pthreadpool",
3740 ],
3741)
3742
3743xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003744 name = "fma3_ukernels_test_mode",
3745 hdrs = INTERNAL_HDRS,
3746 copts = [
3747 "-UNDEBUG",
3748 "-DXNN_TEST_MODE=1",
3749 ],
3750 gcc_copts = xnnpack_gcc_std_copts(),
3751 gcc_x86_copts = ["-mfma"],
3752 msvc_copts = xnnpack_msvc_std_copts(),
3753 msvc_x86_32_copts = ["/arch:AVX"],
3754 msvc_x86_64_copts = ["/arch:AVX"],
3755 x86_srcs = FMA3_UKERNELS,
3756 deps = [
3757 ":tables",
3758 "@FP16",
3759 "@pthreadpool",
3760 ],
3761)
3762
3763xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003764 name = "avx2_ukernels",
3765 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003766 gcc_copts = xnnpack_gcc_std_copts(),
3767 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003768 "-mfma",
3769 "-mavx2",
3770 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003771 msvc_copts = xnnpack_msvc_std_copts(),
3772 msvc_x86_32_copts = ["/arch:AVX2"],
3773 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003774 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003775 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003776 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003777 "@FP16",
3778 "@pthreadpool",
3779 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003780)
3781
3782xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003783 name = "avx2_ukernels_test_mode",
3784 hdrs = INTERNAL_HDRS,
3785 copts = [
3786 "-UNDEBUG",
3787 "-DXNN_TEST_MODE=1",
3788 ],
3789 gcc_copts = xnnpack_gcc_std_copts(),
3790 gcc_x86_copts = [
3791 "-mfma",
3792 "-mavx2",
3793 ],
3794 msvc_copts = xnnpack_msvc_std_copts(),
3795 msvc_x86_32_copts = ["/arch:AVX2"],
3796 msvc_x86_64_copts = ["/arch:AVX2"],
3797 x86_srcs = AVX2_UKERNELS,
3798 deps = [
3799 ":tables",
3800 "@FP16",
3801 "@pthreadpool",
3802 ],
3803)
3804
3805xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003806 name = "avx512f_ukernels",
3807 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003808 gcc_copts = xnnpack_gcc_std_copts(),
3809 gcc_x86_copts = ["-mavx512f"],
3810 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3811 msvc_copts = xnnpack_msvc_std_copts(),
3812 msvc_x86_32_copts = ["/arch:AVX512"],
3813 msvc_x86_64_copts = ["/arch:AVX512"],
3814 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003815 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003816 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003817 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003818 "@FP16",
3819 "@pthreadpool",
3820 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003821)
3822
3823xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003824 name = "avx512f_ukernels_test_mode",
3825 hdrs = INTERNAL_HDRS,
3826 copts = [
3827 "-UNDEBUG",
3828 "-DXNN_TEST_MODE=1",
3829 ],
3830 gcc_copts = xnnpack_gcc_std_copts(),
3831 gcc_x86_copts = ["-mavx512f"],
3832 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3833 msvc_copts = xnnpack_msvc_std_copts(),
3834 msvc_x86_32_copts = ["/arch:AVX512"],
3835 msvc_x86_64_copts = ["/arch:AVX512"],
3836 msys_copts = ["-fno-asynchronous-unwind-tables"],
3837 x86_srcs = AVX512F_UKERNELS,
3838 deps = [
3839 ":tables",
3840 "@FP16",
3841 "@pthreadpool",
3842 ],
3843)
3844
3845xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003846 name = "avx512skx_ukernels",
3847 hdrs = INTERNAL_HDRS,
3848 gcc_copts = xnnpack_gcc_std_copts(),
3849 gcc_x86_copts = [
3850 "-mavx512f",
3851 "-mavx512cd",
3852 "-mavx512bw",
3853 "-mavx512dq",
3854 "-mavx512vl",
3855 ],
3856 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3857 msvc_copts = xnnpack_msvc_std_copts(),
3858 msvc_x86_32_copts = ["/arch:AVX512"],
3859 msvc_x86_64_copts = ["/arch:AVX512"],
3860 msys_copts = ["-fno-asynchronous-unwind-tables"],
3861 x86_srcs = AVX512SKX_UKERNELS,
3862 deps = [
3863 ":tables",
3864 "@FP16",
3865 "@pthreadpool",
3866 ],
3867)
3868
3869xnnpack_cc_library(
3870 name = "avx512skx_ukernels_test_mode",
3871 hdrs = INTERNAL_HDRS,
3872 copts = [
3873 "-UNDEBUG",
3874 "-DXNN_TEST_MODE=1",
3875 ],
3876 gcc_copts = xnnpack_gcc_std_copts(),
3877 gcc_x86_copts = [
3878 "-mavx512f",
3879 "-mavx512cd",
3880 "-mavx512bw",
3881 "-mavx512dq",
3882 "-mavx512vl",
3883 ],
3884 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3885 msvc_copts = xnnpack_msvc_std_copts(),
3886 msvc_x86_32_copts = ["/arch:AVX512"],
3887 msvc_x86_64_copts = ["/arch:AVX512"],
3888 msys_copts = ["-fno-asynchronous-unwind-tables"],
3889 x86_srcs = AVX512SKX_UKERNELS,
3890 deps = [
3891 ":tables",
3892 "@FP16",
3893 "@pthreadpool",
3894 ],
3895)
3896
3897xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003898 name = "asm_ukernels",
3899 hdrs = ["src/xnnpack/assembly.h"],
3900 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07003901 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003902 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003903 wasm_srcs = WASM32_ASM_UKERNELS,
3904 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003905)
3906
Marat Dukhan3b59de22020-06-03 20:15:19 -07003907xnnpack_cc_library(
3908 name = "logging_utils",
3909 srcs = LOGGING_SRCS,
3910 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3911 copts = LOGGING_COPTS + [
3912 "-Isrc",
3913 "-Iinclude",
3914 ] + select({
3915 ":debug_build": [],
3916 "//conditions:default": xnnpack_min_size_copts(),
3917 }),
3918 gcc_copts = xnnpack_gcc_std_copts(),
3919 msvc_copts = xnnpack_msvc_std_copts(),
3920 visibility = xnnpack_visibility(),
3921 deps = [
3922 "@FP16",
3923 "@clog",
3924 "@pthreadpool",
3925 ],
3926)
3927
Marat Dukhan08c4a432019-10-03 09:29:21 -07003928xnnpack_aggregate_library(
3929 name = "ukernels",
3930 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003931 ":neon_ukernels",
3932 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003933 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003934 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003935 ":asm_ukernels",
3936 ],
3937 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003938 ":neon_ukernels",
3939 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003940 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003941 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003942 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003943 ":asm_ukernels",
3944 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003945 generic_deps = [
3946 ":scalar_ukernels",
3947 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003948 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003949 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003950 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003951 wasm_deps = [
3952 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003953 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003954 ],
3955 wasmsimd_deps = [
3956 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003957 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003958 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003959 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003960 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003961 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003962 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003963 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003964 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003965 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003966 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003967 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003968 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003969 ],
3970)
3971
Marat Dukhan33fcf782020-05-24 14:27:15 -07003972xnnpack_aggregate_library(
3973 name = "ukernels_test_mode",
3974 aarch32_deps = [
3975 ":neon_ukernels_test_mode",
3976 ":neonfma_ukernels_test_mode",
3977 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003978 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003979 ":asm_ukernels",
3980 ],
3981 aarch64_deps = [
3982 ":neon_ukernels_test_mode",
3983 ":neonfma_ukernels_test_mode",
3984 ":neonv8_ukernels_test_mode",
3985 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003986 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003987 ":asm_ukernels",
3988 ],
3989 generic_deps = [
3990 ":scalar_ukernels_test_mode",
3991 ],
3992 psimd_deps = [
Marat Dukhan33fcf782020-05-24 14:27:15 -07003993 ":psimd_accmath_ukernels_test_mode",
3994 ],
3995 wasm_deps = [
3996 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003997 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003998 ],
3999 wasmsimd_deps = [
4000 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004001 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004002 ],
4003 x86_deps = [
4004 ":sse2_ukernels_test_mode",
4005 ":ssse3_ukernels_test_mode",
4006 ":sse41_ukernels_test_mode",
4007 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004008 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004009 ":fma3_ukernels_test_mode",
4010 ":avx2_ukernels_test_mode",
4011 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004012 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004013 ],
4014)
4015
Marat Dukhan86df3842020-06-17 14:03:31 -07004016# This target must be compatible with Android NDK r17 and compile with Clang 6.
4017# It is similar to :ukernels target, but excludes microkernels which require
4018# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07004019xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07004020 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004021 aarch32_deps = [
4022 ":neon_ukernels",
4023 ":neonfma_ukernels",
4024 ":neonv8_ukernels",
4025 ":asm_ukernels",
4026 ],
4027 aarch64_deps = [
4028 ":neon_ukernels",
4029 ":neonfma_ukernels",
4030 ":neonv8_ukernels",
4031 ":asm_ukernels",
4032 ],
4033 generic_deps = [
4034 ":scalar_ukernels",
4035 ],
4036 psimd_deps = [
Marat Dukhanae046f52020-06-15 13:16:14 -07004037 ":psimd_accmath_ukernels",
4038 ],
4039 wasm_deps = [
4040 ":wasm_ukernels",
4041 ],
4042 wasmsimd_deps = [
4043 ":wasm_ukernels",
4044 ],
4045 x86_deps = [
4046 ":sse2_ukernels",
4047 ":ssse3_ukernels",
4048 ":sse41_ukernels",
4049 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004050 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004051 ":fma3_ukernels",
4052 ":avx2_ukernels",
4053 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004054 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004055 ],
4056)
4057
Marat Dukhan08c4a432019-10-03 09:29:21 -07004058xnnpack_cc_library(
4059 name = "im2col",
4060 srcs = ["src/im2col.c"],
4061 hdrs = [
4062 "src/xnnpack/common.h",
4063 "src/xnnpack/im2col.h",
4064 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004065 gcc_copts = xnnpack_gcc_std_copts(),
4066 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004067)
4068
4069xnnpack_cc_library(
4070 name = "indirection",
4071 srcs = ["src/indirection.c"],
4072 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004073 gcc_copts = xnnpack_gcc_std_copts(),
4074 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004075 deps = [
4076 "@FP16",
4077 "@FXdiv",
4078 "@pthreadpool",
4079 ],
4080)
4081
4082xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004083 name = "indirection_test_mode",
4084 srcs = ["src/indirection.c"],
4085 hdrs = INTERNAL_HDRS,
4086 copts = [
4087 "-UNDEBUG",
4088 "-DXNN_TEST_MODE=1",
4089 ],
4090 gcc_copts = xnnpack_gcc_std_copts(),
4091 msvc_copts = xnnpack_msvc_std_copts(),
4092 deps = [
4093 "@FP16",
4094 "@FXdiv",
4095 "@pthreadpool",
4096 ],
4097)
4098
4099xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004100 name = "packing",
4101 srcs = ["src/packing.c"],
4102 hdrs = INTERNAL_HDRS,
4103 gcc_copts = xnnpack_gcc_std_copts(),
4104 msvc_copts = xnnpack_msvc_std_copts(),
4105 deps = [
4106 "@FP16",
4107 "@FXdiv",
4108 "@pthreadpool",
4109 ],
4110)
4111
4112xnnpack_cc_library(
4113 name = "packing_test_mode",
4114 srcs = ["src/packing.c"],
4115 hdrs = INTERNAL_HDRS,
4116 copts = [
4117 "-UNDEBUG",
4118 "-DXNN_TEST_MODE=1",
4119 ],
4120 gcc_copts = xnnpack_gcc_std_copts(),
4121 msvc_copts = xnnpack_msvc_std_copts(),
4122 deps = [
4123 "@FP16",
4124 "@FXdiv",
4125 "@pthreadpool",
4126 ],
4127)
4128
4129xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004130 name = "operator_run",
4131 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004132 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004133 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004134 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4135 "//conditions:default": [],
4136 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004137 gcc_copts = xnnpack_gcc_std_copts(),
4138 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004139 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004140 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004141 "@FP16",
4142 "@FXdiv",
4143 "@clog",
4144 "@pthreadpool",
4145 ],
4146)
4147
Chao Mei6ddfc602020-05-13 22:29:36 -07004148xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004149 name = "operator_run_test_mode",
4150 srcs = ["src/operator-run.c"],
4151 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4152 copts = LOGGING_COPTS + [
4153 "-UNDEBUG",
4154 "-DXNN_TEST_MODE=1",
4155 ] + select({
4156 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4157 "//conditions:default": [],
4158 }),
4159 gcc_copts = xnnpack_gcc_std_copts(),
4160 msvc_copts = xnnpack_msvc_std_copts(),
4161 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004162 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004163 "@FP16",
4164 "@FXdiv",
4165 "@clog",
4166 "@pthreadpool",
4167 ],
4168)
4169
4170xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004171 name = "memory_planner",
4172 srcs = ["src/memory-planner.c"],
4173 hdrs = INTERNAL_HDRS,
4174 defines = select({
4175 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4176 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4177 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4178 }),
4179 gcc_copts = xnnpack_gcc_std_copts(),
4180 msvc_copts = xnnpack_msvc_std_copts(),
4181 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004182 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004183 "@pthreadpool",
4184 ],
4185)
4186
Marat Dukhan33fcf782020-05-24 14:27:15 -07004187xnnpack_cc_library(
4188 name = "memory_planner_test_mode",
4189 srcs = ["src/memory-planner.c"],
4190 hdrs = INTERNAL_HDRS,
4191 copts = [
4192 "-UNDEBUG",
4193 "-DXNN_TEST_MODE=1",
4194 ],
4195 defines = select({
4196 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4197 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4198 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4199 }),
4200 gcc_copts = xnnpack_gcc_std_copts(),
4201 msvc_copts = xnnpack_msvc_std_copts(),
4202 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004203 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004204 "@pthreadpool",
4205 ],
4206)
4207
Marat Dukhan08c4a432019-10-03 09:29:21 -07004208cc_library(
4209 name = "enable_assembly",
4210 defines = select({
4211 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4212 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004213 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004214 }),
4215)
4216
Marat Dukhan9de90e02020-06-18 16:04:12 -07004217cc_library(
4218 name = "enable_sparse",
4219 defines = select({
4220 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4221 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07004222 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004223 }),
4224)
4225
Marat Dukhancf056b22019-10-07 10:26:29 -07004226xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004227 name = "operators",
4228 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004229 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004230 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004231 ],
4232 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004233 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004234 "-Isrc",
4235 "-Iinclude",
4236 ] + select({
4237 ":debug_build": [],
4238 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004239 }) + select({
4240 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4241 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004242 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004243 gcc_copts = xnnpack_gcc_std_copts(),
4244 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004245 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004246 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004247 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004248 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004249 "@FP16",
4250 "@FXdiv",
4251 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004252 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004253 ],
4254)
4255
Marat Dukhan10a38082020-04-17 03:58:35 -07004256xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004257 name = "operators_test_mode",
4258 srcs = OPERATOR_SRCS + [
4259 "src/memory.c",
4260 "src/operator-delete.c",
4261 ],
4262 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4263 copts = LOGGING_COPTS + [
4264 "-Isrc",
4265 "-Iinclude",
4266 "-UNDEBUG",
4267 "-DXNN_TEST_MODE=1",
4268 ] + select({
4269 ":debug_build": [],
4270 "//conditions:default": xnnpack_min_size_copts(),
4271 }) + select({
4272 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4273 "//conditions:default": [],
4274 }),
4275 gcc_copts = xnnpack_gcc_std_copts(),
4276 msvc_copts = xnnpack_msvc_std_copts(),
4277 deps = [
4278 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004279 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004280 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004281 "@FP16",
4282 "@FXdiv",
4283 "@clog",
4284 "@pthreadpool",
4285 ],
4286)
4287
4288xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004289 name = "XNNPACK",
4290 srcs = [
4291 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004292 "src/runtime.c",
4293 "src/subgraph.c",
4294 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004295 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004296 hdrs = ["include/xnnpack.h"],
4297 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004298 "-Isrc",
4299 "-Iinclude",
4300 ] + select({
4301 ":debug_build": [],
4302 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004303 }) + select({
4304 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4305 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004306 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004307 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004308 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004309 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004310 visibility = xnnpack_visibility(),
4311 deps = [
4312 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004313 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004314 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004315 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004316 ":operator_run",
4317 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004318 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004319 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004320 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004321 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004322 ] + select({
4323 ":emscripten": [],
4324 "//conditions:default": ["@cpuinfo"],
4325 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004326)
4327
Marat Dukhan10a38082020-04-17 03:58:35 -07004328xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004329 name = "XNNPACK_test_mode",
4330 srcs = [
4331 "src/init.c",
4332 "src/runtime.c",
4333 "src/subgraph.c",
4334 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004335 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004336 hdrs = ["include/xnnpack.h"],
4337 copts = LOGGING_COPTS + [
4338 "-Isrc",
4339 "-Iinclude",
4340 "-UNDEBUG",
4341 "-DXNN_TEST_MODE=1",
4342 ] + select({
4343 ":debug_build": [],
4344 "//conditions:default": xnnpack_min_size_copts(),
4345 }) + select({
4346 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4347 "//conditions:default": [],
4348 }),
4349 gcc_copts = xnnpack_gcc_std_copts(),
4350 includes = ["include"],
4351 msvc_copts = xnnpack_msvc_std_copts(),
4352 visibility = xnnpack_visibility(),
4353 deps = [
4354 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004355 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004356 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004357 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004358 ":operator_run_test_mode",
4359 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004360 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004361 "@clog",
4362 "@FP16",
4363 "@pthreadpool",
4364 ] + select({
4365 ":emscripten": [],
4366 "//conditions:default": ["@cpuinfo"],
4367 }),
4368)
4369
4370xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004371 name = "xnnpack_f32",
4372 srcs = [
4373 "src/init.c",
4374 "src/runtime.c",
4375 "src/subgraph.c",
4376 "src/tensor.c",
4377 ] + SUBGRAPH_SRCS,
4378 hdrs = ["include/xnnpack.h"],
4379 copts = LOGGING_COPTS + [
4380 "-Isrc",
4381 "-Iinclude",
4382 ] + select({
4383 ":debug_build": [],
4384 "//conditions:default": xnnpack_min_size_copts(),
4385 }) + select({
4386 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4387 "//conditions:default": [],
4388 }),
4389 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004390 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004391 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004392 "XNN_NO_U8_OPERATORS",
4393 "XNN_NO_X8_OPERATORS",
4394 "XNN_NO_F16_OPERATORS",
4395 "XNN_NO_X16_OPERATORS",
4396 ],
4397 gcc_copts = xnnpack_gcc_std_copts(),
4398 includes = ["include"],
4399 msvc_copts = xnnpack_msvc_std_copts(),
4400 visibility = xnnpack_visibility(),
4401 deps = [
4402 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004403 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004404 ":logging_utils",
4405 ":memory_planner",
4406 ":operator_run",
4407 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004408 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004409 "@clog",
4410 "@FP16",
4411 "@pthreadpool",
4412 ] + select({
4413 ":emscripten": [],
4414 "//conditions:default": ["@cpuinfo"],
4415 }),
4416)
4417
4418xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004419 name = "xnnpack_operators_nhwc_f32",
4420 srcs = [
4421 "src/init.c",
4422 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004423 hdrs = ["include/xnnpack.h"],
4424 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004425 "-Isrc",
4426 "-Iinclude",
4427 ] + select({
4428 ":debug_build": [],
4429 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004430 }) + select({
4431 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4432 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004433 }),
4434 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004435 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004436 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004437 "XNN_NO_U8_OPERATORS",
4438 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004439 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004440 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004441 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004442 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004443 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004444 visibility = xnnpack_visibility(),
4445 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004446 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004447 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004448 ":operator_run",
4449 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004450 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004451 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004452 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004453 ] + select({
4454 ":emscripten": [],
4455 "//conditions:default": ["@cpuinfo"],
4456 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004457)
4458
Marat Dukhancf056b22019-10-07 10:26:29 -07004459xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004460 name = "bench_utils",
4461 srcs = ["bench/utils.cc"],
4462 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004463 deps = [
4464 "@com_google_benchmark//:benchmark",
4465 "@cpuinfo",
4466 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004467)
4468
Frank Barchard7e955972019-10-11 10:34:25 -07004469######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004470
4471xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004472 name = "qs8_gemm_bench",
4473 srcs = [
4474 "bench/gemm.h",
4475 "bench/qs8-gemm.cc",
4476 "src/xnnpack/AlignedAllocator.h",
4477 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004478 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4479 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004480)
4481
4482xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004483 name = "qs8_requantization_bench",
4484 srcs = [
4485 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004486 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004487 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004488 ] + MICROKERNEL_BENCHMARK_HDRS,
4489 deps = MICROKERNEL_BENCHMARK_DEPS,
4490)
4491
4492xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004493 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004494 srcs = [
4495 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004496 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004497 "src/xnnpack/AlignedAllocator.h",
4498 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004499 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004500 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004501)
4502
4503xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004504 name = "qu8_requantization_bench",
4505 srcs = [
4506 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004507 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004508 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004509 ] + MICROKERNEL_BENCHMARK_HDRS,
4510 deps = MICROKERNEL_BENCHMARK_DEPS,
4511)
4512
4513xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004514 name = "f16_igemm_bench",
4515 srcs = [
4516 "bench/f16-igemm.cc",
4517 "bench/conv.h",
4518 "bench/google/conv.h",
4519 "src/xnnpack/AlignedAllocator.h",
4520 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004521 deps = MICROKERNEL_BENCHMARK_DEPS + [
4522 ":indirection",
4523 ":packing",
4524 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004525)
4526
4527xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004528 name = "f16_gemm_bench",
4529 srcs = [
4530 "bench/f16-gemm.cc",
4531 "bench/gemm.h",
4532 "src/xnnpack/AlignedAllocator.h",
4533 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004534 deps = MICROKERNEL_BENCHMARK_DEPS + [
4535 ":packing",
4536 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004537)
4538
4539xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004540 name = "f16_spmm_bench",
4541 srcs = [
4542 "bench/f16-spmm.cc",
4543 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004544 "src/xnnpack/AlignedAllocator.h",
4545 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004546 deps = MICROKERNEL_BENCHMARK_DEPS,
4547)
4548
4549xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004550 name = "f32_igemm_bench",
4551 srcs = [
4552 "bench/f32-igemm.cc",
4553 "bench/conv.h",
4554 "src/xnnpack/AlignedAllocator.h",
4555 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004556 deps = MICROKERNEL_BENCHMARK_DEPS + [
4557 ":indirection",
4558 ":packing",
4559 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004560)
4561
4562xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004563 name = "f16_relu_bench",
4564 srcs = [
4565 "bench/f16-relu.cc",
4566 "src/xnnpack/AlignedAllocator.h",
4567 ] + MICROKERNEL_BENCHMARK_HDRS,
4568 deps = MICROKERNEL_BENCHMARK_DEPS,
4569)
4570
4571xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004572 name = "f32_conv_hwc_bench",
4573 srcs = [
4574 "bench/f32-conv-hwc.cc",
4575 "bench/dconv.h",
4576 "src/xnnpack/AlignedAllocator.h",
4577 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004578 deps = MICROKERNEL_BENCHMARK_DEPS + [
4579 ":packing",
4580 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004581)
4582
4583xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004584 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004585 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004586 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004587 "bench/dconv.h",
4588 "src/xnnpack/AlignedAllocator.h",
4589 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004590 deps = MICROKERNEL_BENCHMARK_DEPS + [
4591 ":packing",
4592 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004593)
4594
4595xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004596 name = "f16_dwconv_bench",
4597 srcs = [
4598 "bench/f16-dwconv.cc",
4599 "bench/dwconv.h",
4600 "bench/google/dwconv.h",
4601 "src/xnnpack/AlignedAllocator.h",
4602 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004603 deps = MICROKERNEL_BENCHMARK_DEPS + [
4604 ":indirection",
4605 ":packing",
4606 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004607)
4608
4609xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004610 name = "f32_dwconv_bench",
4611 srcs = [
4612 "bench/f32-dwconv.cc",
4613 "bench/dwconv.h",
4614 "src/xnnpack/AlignedAllocator.h",
4615 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004616 deps = MICROKERNEL_BENCHMARK_DEPS + [
4617 ":indirection",
4618 ":packing",
4619 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004620)
4621
4622xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07004623 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004624 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07004625 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004626 "bench/dwconv.h",
4627 "src/xnnpack/AlignedAllocator.h",
4628 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004629 deps = MICROKERNEL_BENCHMARK_DEPS + [
4630 ":indirection",
4631 ":packing",
4632 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004633)
4634
4635xnnpack_benchmark(
4636 name = "f32_gemm_bench",
4637 srcs = [
4638 "bench/f32-gemm.cc",
4639 "bench/gemm.h",
4640 "src/xnnpack/AlignedAllocator.h",
4641 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004642 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004643 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004644)
4645
4646xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004647 name = "f32_hswish_bench",
4648 srcs = [
4649 "bench/f32-hswish.cc",
4650 "src/xnnpack/AlignedAllocator.h",
4651 ] + MICROKERNEL_BENCHMARK_HDRS,
4652 deps = MICROKERNEL_BENCHMARK_DEPS,
4653)
4654
4655xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004656 name = "f32_raddexpminusmax_bench",
4657 srcs = [
4658 "bench/f32-raddexpminusmax.cc",
4659 "src/xnnpack/AlignedAllocator.h",
4660 ] + MICROKERNEL_BENCHMARK_HDRS,
4661 deps = MICROKERNEL_BENCHMARK_DEPS,
4662)
4663
4664xnnpack_benchmark(
4665 name = "f32_raddextexp_bench",
4666 srcs = [
4667 "bench/f32-raddextexp.cc",
4668 "src/xnnpack/AlignedAllocator.h",
4669 ] + MICROKERNEL_BENCHMARK_HDRS,
4670 deps = MICROKERNEL_BENCHMARK_DEPS,
4671)
4672
4673xnnpack_benchmark(
4674 name = "f32_raddstoreexpminusmax_bench",
4675 srcs = [
4676 "bench/f32-raddstoreexpminusmax.cc",
4677 "src/xnnpack/AlignedAllocator.h",
4678 ] + MICROKERNEL_BENCHMARK_HDRS,
4679 deps = MICROKERNEL_BENCHMARK_DEPS,
4680)
4681
4682xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004683 name = "f32_relu_bench",
4684 srcs = [
4685 "bench/f32-relu.cc",
4686 "src/xnnpack/AlignedAllocator.h",
4687 ] + MICROKERNEL_BENCHMARK_HDRS,
4688 deps = MICROKERNEL_BENCHMARK_DEPS,
4689)
4690
4691xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004692 name = "f32_rmax_bench",
4693 srcs = [
4694 "bench/f32-rmax.cc",
4695 "src/xnnpack/AlignedAllocator.h",
4696 ] + MICROKERNEL_BENCHMARK_HDRS,
4697 deps = MICROKERNEL_BENCHMARK_DEPS,
4698)
4699
4700xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004701 name = "f32_sigmoid_bench",
4702 srcs = [
4703 "bench/f32-sigmoid.cc",
4704 "src/xnnpack/AlignedAllocator.h",
4705 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004706 deps = MICROKERNEL_BENCHMARK_DEPS,
4707)
4708
4709xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004710 name = "f32_spmm_bench",
4711 srcs = [
4712 "bench/f32-spmm.cc",
4713 "bench/gemm.h",
4714 "src/xnnpack/AlignedAllocator.h",
4715 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004716 deps = MICROKERNEL_BENCHMARK_DEPS,
4717)
4718
4719xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004720 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004721 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004722 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004723 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004724 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004725 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004726)
4727
4728xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004729 name = "f32_vscaleexpminusmax_bench",
4730 srcs = [
4731 "bench/f32-vscaleexpminusmax.cc",
4732 "src/xnnpack/AlignedAllocator.h",
4733 ] + MICROKERNEL_BENCHMARK_HDRS,
4734 deps = MICROKERNEL_BENCHMARK_DEPS,
4735)
4736
4737xnnpack_benchmark(
4738 name = "f32_vscaleextexp_bench",
4739 srcs = [
4740 "bench/f32-vscaleextexp.cc",
4741 "src/xnnpack/AlignedAllocator.h",
4742 ] + MICROKERNEL_BENCHMARK_HDRS,
4743 deps = MICROKERNEL_BENCHMARK_DEPS,
4744)
4745
4746xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004747 name = "f32_vsqrt_bench",
4748 srcs = [
4749 "bench/f32-vsqrt.cc",
4750 "src/xnnpack/AlignedAllocator.h",
4751 ] + MICROKERNEL_BENCHMARK_HDRS,
4752 deps = MICROKERNEL_BENCHMARK_DEPS,
4753)
4754
4755xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004756 name = "f32_im2col_gemm_bench",
4757 srcs = [
4758 "bench/f32-im2col-gemm.cc",
4759 "bench/conv.h",
4760 "src/xnnpack/AlignedAllocator.h",
4761 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004762 deps = MICROKERNEL_BENCHMARK_DEPS + [
4763 ":im2col",
4764 ":packing",
4765 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004766)
4767
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004768xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004769 name = "rounding_bench",
4770 srcs = [
4771 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004772 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004773 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004774 ] + MICROKERNEL_BENCHMARK_HDRS,
4775 deps = MICROKERNEL_BENCHMARK_DEPS,
4776)
4777
Marat Dukhan08c4a432019-10-03 09:29:21 -07004778########################### Benchmarks for operators ###########################
4779
4780xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004781 name = "average_pooling_bench",
4782 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004783 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004784 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004785 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004786)
4787
4788xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004789 name = "bankers_rounding_bench",
4790 srcs = ["bench/bankers-rounding.cc"],
4791 copts = xnnpack_optional_tflite_copts(),
4792 tags = ["nowin32"],
4793 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4794)
4795
4796xnnpack_benchmark(
4797 name = "ceiling_bench",
4798 srcs = ["bench/ceiling.cc"],
4799 copts = xnnpack_optional_tflite_copts(),
4800 tags = ["nowin32"],
4801 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4802)
4803
4804xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004805 name = "channel_shuffle_bench",
4806 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004807 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004808)
4809
4810xnnpack_benchmark(
4811 name = "convolution_bench",
4812 srcs = ["bench/convolution.cc"],
4813 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004814 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004815 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004816)
4817
4818xnnpack_benchmark(
4819 name = "deconvolution_bench",
4820 srcs = ["bench/deconvolution.cc"],
4821 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004822 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004823 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004824)
4825
4826xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004827 name = "floor_bench",
4828 srcs = ["bench/floor.cc"],
4829 copts = xnnpack_optional_tflite_copts(),
4830 tags = ["nowin32"],
4831 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4832)
4833
4834xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004835 name = "global_average_pooling_bench",
4836 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004837 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004838)
4839
4840xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004841 name = "hardswish_bench",
4842 srcs = ["bench/hardswish.cc"],
4843 copts = xnnpack_optional_tflite_copts(),
4844 tags = ["nowin32"],
4845 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4846)
4847
4848xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004849 name = "max_pooling_bench",
4850 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004851 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004852)
4853
4854xnnpack_benchmark(
4855 name = "sigmoid_bench",
4856 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004857 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004858 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004859 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004860)
4861
4862xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004863 name = "prelu_bench",
4864 srcs = ["bench/prelu.cc"],
4865 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004866 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004867 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004868)
4869
4870xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004871 name = "softmax_bench",
4872 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004873 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004874 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004875 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004876)
4877
Marat Dukhan87727142020-06-24 15:24:10 -07004878xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004879 name = "square_root_bench",
4880 srcs = ["bench/square-root.cc"],
4881 copts = xnnpack_optional_tflite_copts(),
4882 tags = ["nowin32"],
4883 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4884)
4885
4886xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004887 name = "truncation_bench",
4888 srcs = ["bench/truncation.cc"],
4889 deps = OPERATOR_BENCHMARK_DEPS,
4890)
4891
Marat Dukhanc068bb62019-10-04 13:24:39 -07004892############################# End-to-end benchmarks ############################
4893
4894cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004895 name = "fp32_mobilenet_v1",
4896 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004897 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004898 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004899 linkstatic = True,
4900 deps = [
4901 ":XNNPACK",
4902 "@pthreadpool",
4903 ],
4904)
4905
4906cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004907 name = "fp16_mobilenet_v1",
4908 srcs = ["models/fp16-mobilenet-v1.cc"],
4909 hdrs = ["models/models.h"],
4910 copts = xnnpack_std_cxxopts(),
4911 linkstatic = True,
4912 deps = [
4913 ":XNNPACK",
4914 "@FP16",
4915 "@pthreadpool",
4916 ],
4917)
4918
4919cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004920 name = "qs8_mobilenet_v1",
4921 srcs = ["models/qs8-mobilenet-v1.cc"],
4922 hdrs = ["models/models.h"],
4923 copts = xnnpack_std_cxxopts(),
4924 linkstatic = True,
4925 deps = [
4926 ":XNNPACK",
4927 "@pthreadpool",
4928 ],
4929)
4930
4931cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004932 name = "qs8_mobilenet_v2",
4933 srcs = ["models/qs8-mobilenet-v2.cc"],
4934 hdrs = ["models/models.h"],
4935 copts = xnnpack_std_cxxopts(),
4936 linkstatic = True,
4937 deps = [
4938 ":XNNPACK",
4939 "@pthreadpool",
4940 ],
4941)
4942
4943cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004944 name = "fp32_mobilenet_v2",
4945 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004946 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004947 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004948 linkstatic = True,
4949 deps = [
4950 ":XNNPACK",
4951 "@pthreadpool",
4952 ],
4953)
4954
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004955cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004956 name = "fp16_mobilenet_v2",
4957 srcs = ["models/fp16-mobilenet-v2.cc"],
4958 hdrs = ["models/models.h"],
4959 copts = xnnpack_std_cxxopts(),
4960 linkstatic = True,
4961 deps = [
4962 ":XNNPACK",
4963 "@FP16",
4964 "@pthreadpool",
4965 ],
4966)
4967
4968cc_library(
4969 name = "fp32_mobilenet_v3_large",
4970 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004971 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004972 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004973 linkstatic = True,
4974 deps = [
4975 ":XNNPACK",
4976 "@pthreadpool",
4977 ],
4978)
4979
4980cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004981 name = "fp16_mobilenet_v3_large",
4982 srcs = ["models/fp16-mobilenet-v3-large.cc"],
4983 hdrs = ["models/models.h"],
4984 copts = xnnpack_std_cxxopts(),
4985 linkstatic = True,
4986 deps = [
4987 ":XNNPACK",
4988 "@FP16",
4989 "@pthreadpool",
4990 ],
4991)
4992
4993cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004994 name = "fp32_mobilenet_v3_small",
4995 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004996 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004997 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004998 linkstatic = True,
4999 deps = [
5000 ":XNNPACK",
5001 "@pthreadpool",
5002 ],
5003)
5004
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005005cc_library(
5006 name = "fp16_mobilenet_v3_small",
5007 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5008 hdrs = ["models/models.h"],
5009 copts = xnnpack_std_cxxopts(),
5010 linkstatic = True,
5011 deps = [
5012 ":XNNPACK",
5013 "@FP16",
5014 "@pthreadpool",
5015 ],
5016)
5017
Marat Dukhanc068bb62019-10-04 13:24:39 -07005018xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005019 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005020 srcs = [
5021 "bench/f32-dwconv-e2e.cc",
5022 "bench/end2end.h",
5023 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005024 deps = MICROKERNEL_BENCHMARK_DEPS + [
5025 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005026 ":fp32_mobilenet_v1",
5027 ":fp32_mobilenet_v2",
5028 ":fp32_mobilenet_v3_large",
5029 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005030 ],
5031)
5032
5033xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005034 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005035 srcs = [
5036 "bench/f32-gemm-e2e.cc",
5037 "bench/end2end.h",
5038 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005039 deps = MICROKERNEL_BENCHMARK_DEPS + [
5040 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005041 ":fp32_mobilenet_v1",
5042 ":fp32_mobilenet_v2",
5043 ":fp32_mobilenet_v3_large",
5044 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005045 ],
5046)
5047
5048xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005049 name = "end2end_bench",
5050 srcs = ["bench/end2end.cc"],
5051 deps = [
5052 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005053 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005054 ":fp16_mobilenet_v1",
5055 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005056 ":fp16_mobilenet_v3_large",
5057 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005058 ":fp32_mobilenet_v1",
5059 ":fp32_mobilenet_v2",
5060 ":fp32_mobilenet_v3_large",
5061 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005062 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005063 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005064 "@pthreadpool",
5065 ],
5066)
5067
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005068#################### Accuracy evaluation for math functions ####################
5069
5070xnnpack_benchmark(
5071 name = "f32_exp_eval",
5072 srcs = [
5073 "eval/f32-exp.cc",
5074 "src/xnnpack/AlignedAllocator.h",
5075 ] + ACCURACY_EVAL_HDRS,
5076 deps = ACCURACY_EVAL_DEPS,
5077)
5078
Marat Dukhan515c9772019-10-17 18:07:57 -07005079xnnpack_benchmark(
5080 name = "f32_expminus_eval",
5081 srcs = [
5082 "eval/f32-expminus.cc",
5083 "src/xnnpack/AlignedAllocator.h",
5084 ] + ACCURACY_EVAL_HDRS,
5085 deps = ACCURACY_EVAL_DEPS,
5086)
5087
Marat Dukhan98ba4412019-10-23 02:14:28 -07005088xnnpack_benchmark(
Marat Dukhana438aca2020-11-20 15:45:01 -08005089 name = "f32_expm1minus_eval",
5090 srcs = [
5091 "eval/f32-expm1minus.cc",
5092 "src/xnnpack/AlignedAllocator.h",
5093 ] + ACCURACY_EVAL_HDRS,
5094 deps = ACCURACY_EVAL_DEPS,
5095)
5096
5097xnnpack_benchmark(
Marat Dukhan98ba4412019-10-23 02:14:28 -07005098 name = "f32_extexp_eval",
5099 srcs = [
5100 "eval/f32-extexp.cc",
5101 "src/xnnpack/AlignedAllocator.h",
5102 ] + ACCURACY_EVAL_HDRS,
5103 deps = ACCURACY_EVAL_DEPS,
5104)
5105
Marat Dukhan8853b822020-05-07 12:19:01 -07005106xnnpack_unit_test(
5107 name = "f32_roundne_eval",
5108 srcs = [
5109 "eval/f32-roundne.cc",
5110 "src/xnnpack/AlignedAllocator.h",
5111 "src/xnnpack/math-stubs.h",
5112 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005113 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005114 deps = MICROKERNEL_TEST_DEPS,
5115)
5116
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005117xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005118 name = "f32_roundd_eval",
5119 srcs = [
5120 "eval/f32-roundd.cc",
5121 "src/xnnpack/AlignedAllocator.h",
5122 "src/xnnpack/math-stubs.h",
5123 ] + MICROKERNEL_TEST_HDRS,
5124 automatic = False,
5125 deps = MICROKERNEL_TEST_DEPS,
5126)
5127
5128xnnpack_unit_test(
5129 name = "f32_roundu_eval",
5130 srcs = [
5131 "eval/f32-roundu.cc",
5132 "src/xnnpack/AlignedAllocator.h",
5133 "src/xnnpack/math-stubs.h",
5134 ] + MICROKERNEL_TEST_HDRS,
5135 automatic = False,
5136 deps = MICROKERNEL_TEST_DEPS,
5137)
5138
5139xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005140 name = "f32_roundz_eval",
5141 srcs = [
5142 "eval/f32-roundz.cc",
5143 "src/xnnpack/AlignedAllocator.h",
5144 "src/xnnpack/math-stubs.h",
5145 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005146 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005147 deps = MICROKERNEL_TEST_DEPS,
5148)
5149
Marat Dukhan346a9e52019-11-15 09:06:30 -08005150xnnpack_benchmark(
5151 name = "f32_sigmoid_eval",
5152 srcs = [
5153 "eval/f32-sigmoid.cc",
5154 "src/xnnpack/AlignedAllocator.h",
5155 ] + ACCURACY_EVAL_HDRS,
5156 deps = ACCURACY_EVAL_DEPS,
5157)
5158
Marat Dukhan84000762020-06-29 18:38:43 -07005159xnnpack_benchmark(
5160 name = "f32_sqrt_eval",
5161 srcs = [
5162 "eval/f32-sqrt.cc",
5163 "src/xnnpack/AlignedAllocator.h",
5164 ] + ACCURACY_EVAL_HDRS,
5165 deps = ACCURACY_EVAL_DEPS + [
5166 ":bench_utils",
5167 ],
5168)
5169
Marat Dukhan08c4a432019-10-03 09:29:21 -07005170######################### Unit tests for micro-kernels #########################
5171
5172xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005173 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005174 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005175 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005176 "test/gemm-microkernel-tester.h",
5177 "src/xnnpack/AlignedAllocator.h",
5178 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005179 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005180)
5181
5182xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005183 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005184 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005185 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005186 "test/spmm-microkernel-tester.h",
5187 "src/xnnpack/AlignedAllocator.h",
5188 ] + MICROKERNEL_TEST_HDRS,
5189 deps = MICROKERNEL_TEST_DEPS,
5190)
5191
5192xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005193 name = "f32_argmaxpool_test",
5194 srcs = [
5195 "test/f32-argmaxpool.cc",
5196 "test/argmaxpool-microkernel-tester.h",
5197 "src/xnnpack/AlignedAllocator.h",
5198 ] + MICROKERNEL_TEST_HDRS,
5199 deps = MICROKERNEL_TEST_DEPS,
5200)
5201
5202xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005203 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005204 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005205 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005206 "test/avgpool-microkernel-tester.h",
5207 "src/xnnpack/AlignedAllocator.h",
5208 ] + MICROKERNEL_TEST_HDRS,
5209 deps = MICROKERNEL_TEST_DEPS,
5210)
5211
5212xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005213 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005214 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005215 "test/f32-ibilinear.cc",
5216 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005217 "src/xnnpack/AlignedAllocator.h",
5218 ] + MICROKERNEL_TEST_HDRS,
5219 deps = MICROKERNEL_TEST_DEPS,
5220)
5221
5222xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005223 name = "f32_ibilinear_chw_test",
5224 srcs = [
5225 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005226 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005227 "src/xnnpack/AlignedAllocator.h",
5228 ] + MICROKERNEL_TEST_HDRS,
5229 deps = MICROKERNEL_TEST_DEPS,
5230)
5231
5232xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005233 name = "f16_clamp_test",
5234 srcs = [
5235 "test/f16-clamp.cc",
5236 "test/clamp-microkernel-tester.h",
5237 ] + MICROKERNEL_TEST_HDRS,
5238 deps = MICROKERNEL_TEST_DEPS,
5239)
5240
5241xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005242 name = "f32_clamp_test",
5243 srcs = [
5244 "test/f32-clamp.cc",
5245 "test/clamp-microkernel-tester.h",
5246 ] + MICROKERNEL_TEST_HDRS,
5247 deps = MICROKERNEL_TEST_DEPS,
5248)
5249
5250xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005251 name = "f32_igemm_test",
5252 srcs = [
5253 "test/f32-igemm.cc",
5254 "test/gemm-microkernel-tester.h",
5255 "src/xnnpack/AlignedAllocator.h",
5256 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005257 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005258)
5259
5260xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005261 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005262 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005263 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005264 "test/gemm-microkernel-tester.h",
5265 "src/xnnpack/AlignedAllocator.h",
5266 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005267 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005268)
5269
5270xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005271 name = "f32_igemm_minmax_test",
5272 srcs = [
5273 "test/f32-igemm-minmax.cc",
5274 "test/gemm-microkernel-tester.h",
5275 "src/xnnpack/AlignedAllocator.h",
5276 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005277 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005278)
5279
5280xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005281 name = "f16_igemm_minmax_test",
5282 srcs = [
5283 "test/f16-igemm-minmax.cc",
5284 "test/gemm-microkernel-tester.h",
5285 "src/xnnpack/AlignedAllocator.h",
5286 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005287 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005288)
5289
5290xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005291 name = "f32_conv_hwc_test",
5292 srcs = [
5293 "test/f32-conv-hwc.cc",
5294 "test/conv-hwc-microkernel-tester.h",
5295 "src/xnnpack/AlignedAllocator.h",
5296 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005297 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005298)
5299
5300xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005301 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005302 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005303 "test/f32-conv-hwc2chw.cc",
5304 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005305 "src/xnnpack/AlignedAllocator.h",
5306 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005307 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005308)
5309
5310xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005311 name = "f32_dwconv_test",
5312 srcs = [
5313 "test/f32-dwconv.cc",
5314 "test/dwconv-microkernel-tester.h",
5315 "src/xnnpack/AlignedAllocator.h",
5316 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005317 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005318)
5319
5320xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005321 name = "f16_dwconv_minmax_test",
5322 srcs = [
5323 "test/f16-dwconv-minmax.cc",
5324 "test/dwconv-microkernel-tester.h",
5325 "src/xnnpack/AlignedAllocator.h",
5326 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005327 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005328)
5329
5330xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005331 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005332 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005333 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005334 "test/dwconv-microkernel-tester.h",
5335 "src/xnnpack/AlignedAllocator.h",
5336 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005337 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005338)
5339
5340xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005341 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005342 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005343 "test/f32-dwconv2d-chw.cc",
5344 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005345 "src/xnnpack/AlignedAllocator.h",
5346 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005347 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005348)
5349
5350xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005351 name = "f16_gavgpool_minmax_test",
5352 srcs = [
5353 "test/f16-gavgpool-minmax.cc",
5354 "test/gavgpool-microkernel-tester.h",
5355 "src/xnnpack/AlignedAllocator.h",
5356 ] + MICROKERNEL_TEST_HDRS,
5357 deps = MICROKERNEL_TEST_DEPS,
5358)
5359
5360xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005361 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005362 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005363 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005364 "test/gavgpool-microkernel-tester.h",
5365 "src/xnnpack/AlignedAllocator.h",
5366 ] + MICROKERNEL_TEST_HDRS,
5367 deps = MICROKERNEL_TEST_DEPS,
5368)
5369
5370xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005371 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005372 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005373 "test/f32-gavgpool-cw.cc",
5374 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005375 "src/xnnpack/AlignedAllocator.h",
5376 ] + MICROKERNEL_TEST_HDRS,
5377 deps = MICROKERNEL_TEST_DEPS,
5378)
5379
5380xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005381 name = "f32_gemm_test",
5382 srcs = [
5383 "test/f32-gemm.cc",
5384 "test/gemm-microkernel-tester.h",
5385 "src/xnnpack/AlignedAllocator.h",
5386 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005387 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005388)
5389
5390xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005391 name = "f32_gemm_relu_test",
5392 srcs = [
5393 "test/f32-gemm-relu.cc",
5394 "test/gemm-microkernel-tester.h",
5395 "src/xnnpack/AlignedAllocator.h",
5396 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005397 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005398)
5399
5400xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005401 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005402 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005403 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005404 "test/gemm-microkernel-tester.h",
5405 "src/xnnpack/AlignedAllocator.h",
5406 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005407 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005408)
5409
5410xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005411 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005412 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005413 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005414 "test/gemm-microkernel-tester.h",
5415 "src/xnnpack/AlignedAllocator.h",
5416 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005417 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005418)
5419
5420xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005421 name = "f16_hswish_test",
5422 srcs = [
5423 "test/f16-hswish.cc",
5424 "test/hswish-microkernel-tester.h",
5425 ] + MICROKERNEL_TEST_HDRS,
5426 deps = MICROKERNEL_TEST_DEPS,
5427)
5428
5429xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005430 name = "f32_hswish_test",
5431 srcs = [
5432 "test/f32-hswish.cc",
5433 "test/hswish-microkernel-tester.h",
5434 ] + MICROKERNEL_TEST_HDRS,
5435 deps = MICROKERNEL_TEST_DEPS,
5436)
5437
5438xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005439 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005440 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005441 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005442 "test/maxpool-microkernel-tester.h",
5443 ] + MICROKERNEL_TEST_HDRS,
5444 deps = MICROKERNEL_TEST_DEPS,
5445)
5446
5447xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005448 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005449 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005450 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005451 "test/avgpool-microkernel-tester.h",
5452 "src/xnnpack/AlignedAllocator.h",
5453 ] + MICROKERNEL_TEST_HDRS,
5454 deps = MICROKERNEL_TEST_DEPS,
5455)
5456
5457xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005458 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005459 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005460 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005461 "test/gemm-microkernel-tester.h",
5462 "src/xnnpack/AlignedAllocator.h",
5463 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005464 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005465)
5466
5467xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005468 name = "f16_prelu_test",
5469 srcs = [
5470 "test/f16-prelu.cc",
5471 "test/prelu-microkernel-tester.h",
5472 "src/xnnpack/AlignedAllocator.h",
5473 ] + MICROKERNEL_TEST_HDRS,
5474 deps = MICROKERNEL_TEST_DEPS,
5475)
5476
5477xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005478 name = "f32_prelu_test",
5479 srcs = [
5480 "test/f32-prelu.cc",
5481 "test/prelu-microkernel-tester.h",
5482 "src/xnnpack/AlignedAllocator.h",
5483 ] + MICROKERNEL_TEST_HDRS,
5484 deps = MICROKERNEL_TEST_DEPS,
5485)
5486
5487xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005488 name = "f32_raddexpminusmax_test",
5489 srcs = [
5490 "test/f32-raddexpminusmax.cc",
5491 "test/raddexpminusmax-microkernel-tester.h",
5492 ] + MICROKERNEL_TEST_HDRS,
5493 deps = MICROKERNEL_TEST_DEPS,
5494)
5495
5496xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005497 name = "f32_raddextexp_test",
5498 srcs = [
5499 "test/f32-raddextexp.cc",
5500 "test/raddextexp-microkernel-tester.h",
5501 ] + MICROKERNEL_TEST_HDRS,
5502 deps = MICROKERNEL_TEST_DEPS,
5503)
5504
5505xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005506 name = "f32_raddstoreexpminusmax_test",
5507 srcs = [
5508 "test/f32-raddstoreexpminusmax.cc",
5509 "test/raddstoreexpminusmax-microkernel-tester.h",
5510 ] + MICROKERNEL_TEST_HDRS,
5511 deps = MICROKERNEL_TEST_DEPS,
5512)
5513
5514xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005515 name = "f32_relu_test",
5516 srcs = [
5517 "test/f32-relu.cc",
5518 "test/vunary-microkernel-tester.h",
5519 ] + MICROKERNEL_TEST_HDRS,
5520 deps = MICROKERNEL_TEST_DEPS,
5521)
5522
5523xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005524 name = "f32_rmax_test",
5525 srcs = [
5526 "test/f32-rmax.cc",
5527 "test/rmax-microkernel-tester.h",
5528 ] + MICROKERNEL_TEST_HDRS,
5529 deps = MICROKERNEL_TEST_DEPS,
5530)
5531
5532xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005533 name = "f32_sigmoid_test",
5534 srcs = [
5535 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005536 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005537 ] + MICROKERNEL_TEST_HDRS,
5538 deps = MICROKERNEL_TEST_DEPS,
5539)
5540
5541xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005542 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005543 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005544 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005545 "test/spmm-microkernel-tester.h",
5546 "src/xnnpack/AlignedAllocator.h",
5547 ] + MICROKERNEL_TEST_HDRS,
5548 deps = MICROKERNEL_TEST_DEPS,
5549)
5550
5551xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005552 name = "f16_vadd_minmax_test",
5553 srcs = [
5554 "test/f16-vadd-minmax.cc",
5555 "test/vbinary-microkernel-tester.h",
5556 ] + MICROKERNEL_TEST_HDRS,
5557 deps = MICROKERNEL_TEST_DEPS,
5558)
5559
5560xnnpack_unit_test(
5561 name = "f16_vaddc_minmax_test",
5562 srcs = [
5563 "test/f16-vaddc-minmax.cc",
5564 "test/vbinaryc-microkernel-tester.h",
5565 ] + MICROKERNEL_TEST_HDRS,
5566 deps = MICROKERNEL_TEST_DEPS,
5567)
5568
5569xnnpack_unit_test(
5570 name = "f16_vdiv_minmax_test",
5571 srcs = [
5572 "test/f16-vdiv-minmax.cc",
5573 "test/vbinary-microkernel-tester.h",
5574 ] + MICROKERNEL_TEST_HDRS,
5575 deps = MICROKERNEL_TEST_DEPS,
5576)
5577
5578xnnpack_unit_test(
5579 name = "f16_vdivc_minmax_test",
5580 srcs = [
5581 "test/f16-vdivc-minmax.cc",
5582 "test/vbinaryc-microkernel-tester.h",
5583 ] + MICROKERNEL_TEST_HDRS,
5584 deps = MICROKERNEL_TEST_DEPS,
5585)
5586
5587xnnpack_unit_test(
5588 name = "f16_vrdivc_minmax_test",
5589 srcs = [
5590 "test/f16-vrdivc-minmax.cc",
5591 "test/vbinaryc-microkernel-tester.h",
5592 ] + MICROKERNEL_TEST_HDRS,
5593 deps = MICROKERNEL_TEST_DEPS,
5594)
5595
5596xnnpack_unit_test(
5597 name = "f16_vmax_test",
5598 srcs = [
5599 "test/f16-vmax.cc",
5600 "test/vbinary-microkernel-tester.h",
5601 ] + MICROKERNEL_TEST_HDRS,
5602 deps = MICROKERNEL_TEST_DEPS,
5603)
5604
5605xnnpack_unit_test(
5606 name = "f16_vmaxc_test",
5607 srcs = [
5608 "test/f16-vmaxc.cc",
5609 "test/vbinaryc-microkernel-tester.h",
5610 ] + MICROKERNEL_TEST_HDRS,
5611 deps = MICROKERNEL_TEST_DEPS,
5612)
5613
5614xnnpack_unit_test(
5615 name = "f16_vmin_test",
5616 srcs = [
5617 "test/f16-vmin.cc",
5618 "test/vbinary-microkernel-tester.h",
5619 ] + MICROKERNEL_TEST_HDRS,
5620 deps = MICROKERNEL_TEST_DEPS,
5621)
5622
5623xnnpack_unit_test(
5624 name = "f16_vminc_test",
5625 srcs = [
5626 "test/f16-vminc.cc",
5627 "test/vbinaryc-microkernel-tester.h",
5628 ] + MICROKERNEL_TEST_HDRS,
5629 deps = MICROKERNEL_TEST_DEPS,
5630)
5631
5632xnnpack_unit_test(
5633 name = "f16_vmul_minmax_test",
5634 srcs = [
5635 "test/f16-vmul-minmax.cc",
5636 "test/vbinary-microkernel-tester.h",
5637 ] + MICROKERNEL_TEST_HDRS,
5638 deps = MICROKERNEL_TEST_DEPS,
5639)
5640
5641xnnpack_unit_test(
5642 name = "f16_vmulc_minmax_test",
5643 srcs = [
5644 "test/f16-vmulc-minmax.cc",
5645 "test/vbinaryc-microkernel-tester.h",
5646 ] + MICROKERNEL_TEST_HDRS,
5647 deps = MICROKERNEL_TEST_DEPS,
5648)
5649
5650xnnpack_unit_test(
5651 name = "f16_vsub_minmax_test",
5652 srcs = [
5653 "test/f16-vsub-minmax.cc",
5654 "test/vbinary-microkernel-tester.h",
5655 ] + MICROKERNEL_TEST_HDRS,
5656 deps = MICROKERNEL_TEST_DEPS,
5657)
5658
5659xnnpack_unit_test(
5660 name = "f16_vsubc_minmax_test",
5661 srcs = [
5662 "test/f16-vsubc-minmax.cc",
5663 "test/vbinaryc-microkernel-tester.h",
5664 ] + MICROKERNEL_TEST_HDRS,
5665 deps = MICROKERNEL_TEST_DEPS,
5666)
5667
5668xnnpack_unit_test(
5669 name = "f16_vrsubc_minmax_test",
5670 srcs = [
5671 "test/f16-vrsubc-minmax.cc",
5672 "test/vbinaryc-microkernel-tester.h",
5673 ] + MICROKERNEL_TEST_HDRS,
5674 deps = MICROKERNEL_TEST_DEPS,
5675)
5676
5677xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005678 name = "f32_vabs_test",
5679 srcs = [
5680 "test/f32-vabs.cc",
5681 "test/vunary-microkernel-tester.h",
5682 ] + MICROKERNEL_TEST_HDRS,
5683 deps = MICROKERNEL_TEST_DEPS,
5684)
5685
5686xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005687 name = "f32_vadd_test",
5688 srcs = [
5689 "test/f32-vadd.cc",
5690 "test/vbinary-microkernel-tester.h",
5691 ] + MICROKERNEL_TEST_HDRS,
5692 deps = MICROKERNEL_TEST_DEPS,
5693)
5694
5695xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005696 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005697 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005698 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005699 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005700 ] + MICROKERNEL_TEST_HDRS,
5701 deps = MICROKERNEL_TEST_DEPS,
5702)
5703
5704xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005705 name = "f32_vadd_relu_test",
5706 srcs = [
5707 "test/f32-vadd-relu.cc",
5708 "test/vbinary-microkernel-tester.h",
5709 ] + MICROKERNEL_TEST_HDRS,
5710 deps = MICROKERNEL_TEST_DEPS,
5711)
5712
5713xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005714 name = "f32_vaddc_test",
5715 srcs = [
5716 "test/f32-vaddc.cc",
5717 "test/vbinaryc-microkernel-tester.h",
5718 ] + MICROKERNEL_TEST_HDRS,
5719 deps = MICROKERNEL_TEST_DEPS,
5720)
5721
5722xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005723 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005724 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005725 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005726 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005727 ] + MICROKERNEL_TEST_HDRS,
5728 deps = MICROKERNEL_TEST_DEPS,
5729)
5730
5731xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005732 name = "f32_vaddc_relu_test",
5733 srcs = [
5734 "test/f32-vaddc-relu.cc",
5735 "test/vbinaryc-microkernel-tester.h",
5736 ] + MICROKERNEL_TEST_HDRS,
5737 deps = MICROKERNEL_TEST_DEPS,
5738)
5739
5740xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005741 name = "f32_vdiv_test",
5742 srcs = [
5743 "test/f32-vdiv.cc",
5744 "test/vbinary-microkernel-tester.h",
5745 ] + MICROKERNEL_TEST_HDRS,
5746 deps = MICROKERNEL_TEST_DEPS,
5747)
5748
5749xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005750 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005751 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005752 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005753 "test/vbinary-microkernel-tester.h",
5754 ] + MICROKERNEL_TEST_HDRS,
5755 deps = MICROKERNEL_TEST_DEPS,
5756)
5757
5758xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005759 name = "f32_vdiv_relu_test",
5760 srcs = [
5761 "test/f32-vdiv-relu.cc",
5762 "test/vbinary-microkernel-tester.h",
5763 ] + MICROKERNEL_TEST_HDRS,
5764 deps = MICROKERNEL_TEST_DEPS,
5765)
5766
5767xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005768 name = "f32_vdivc_test",
5769 srcs = [
5770 "test/f32-vdivc.cc",
5771 "test/vbinaryc-microkernel-tester.h",
5772 ] + MICROKERNEL_TEST_HDRS,
5773 deps = MICROKERNEL_TEST_DEPS,
5774)
5775
5776xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005777 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005778 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005779 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005780 "test/vbinaryc-microkernel-tester.h",
5781 ] + MICROKERNEL_TEST_HDRS,
5782 deps = MICROKERNEL_TEST_DEPS,
5783)
5784
5785xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005786 name = "f32_vdivc_relu_test",
5787 srcs = [
5788 "test/f32-vdivc-relu.cc",
5789 "test/vbinaryc-microkernel-tester.h",
5790 ] + MICROKERNEL_TEST_HDRS,
5791 deps = MICROKERNEL_TEST_DEPS,
5792)
5793
5794xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005795 name = "f32_vrdivc_test",
5796 srcs = [
5797 "test/f32-vrdivc.cc",
5798 "test/vbinaryc-microkernel-tester.h",
5799 ] + MICROKERNEL_TEST_HDRS,
5800 deps = MICROKERNEL_TEST_DEPS,
5801)
5802
5803xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005804 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005805 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005806 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005807 "test/vbinaryc-microkernel-tester.h",
5808 ] + MICROKERNEL_TEST_HDRS,
5809 deps = MICROKERNEL_TEST_DEPS,
5810)
5811
5812xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005813 name = "f32_vrdivc_relu_test",
5814 srcs = [
5815 "test/f32-vrdivc-relu.cc",
5816 "test/vbinaryc-microkernel-tester.h",
5817 ] + MICROKERNEL_TEST_HDRS,
5818 deps = MICROKERNEL_TEST_DEPS,
5819)
5820
5821xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005822 name = "f32_vmax_test",
5823 srcs = [
5824 "test/f32-vmax.cc",
5825 "test/vbinary-microkernel-tester.h",
5826 ] + MICROKERNEL_TEST_HDRS,
5827 deps = MICROKERNEL_TEST_DEPS,
5828)
5829
5830xnnpack_unit_test(
5831 name = "f32_vmaxc_test",
5832 srcs = [
5833 "test/f32-vmaxc.cc",
5834 "test/vbinaryc-microkernel-tester.h",
5835 ] + MICROKERNEL_TEST_HDRS,
5836 deps = MICROKERNEL_TEST_DEPS,
5837)
5838
5839xnnpack_unit_test(
5840 name = "f32_vmin_test",
5841 srcs = [
5842 "test/f32-vmin.cc",
5843 "test/vbinary-microkernel-tester.h",
5844 ] + MICROKERNEL_TEST_HDRS,
5845 deps = MICROKERNEL_TEST_DEPS,
5846)
5847
5848xnnpack_unit_test(
5849 name = "f32_vminc_test",
5850 srcs = [
5851 "test/f32-vminc.cc",
5852 "test/vbinaryc-microkernel-tester.h",
5853 ] + MICROKERNEL_TEST_HDRS,
5854 deps = MICROKERNEL_TEST_DEPS,
5855)
5856
5857xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005858 name = "f32_vmul_test",
5859 srcs = [
5860 "test/f32-vmul.cc",
5861 "test/vbinary-microkernel-tester.h",
5862 ] + MICROKERNEL_TEST_HDRS,
5863 deps = MICROKERNEL_TEST_DEPS,
5864)
5865
5866xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005867 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005868 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005869 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005870 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005871 ] + MICROKERNEL_TEST_HDRS,
5872 deps = MICROKERNEL_TEST_DEPS,
5873)
5874
5875xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005876 name = "f32_vmul_relu_test",
5877 srcs = [
5878 "test/f32-vmul-relu.cc",
5879 "test/vbinary-microkernel-tester.h",
5880 ] + MICROKERNEL_TEST_HDRS,
5881 deps = MICROKERNEL_TEST_DEPS,
5882)
5883
5884xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005885 name = "f32_vmulc_test",
5886 srcs = [
5887 "test/f32-vmulc.cc",
5888 "test/vbinaryc-microkernel-tester.h",
5889 ] + MICROKERNEL_TEST_HDRS,
5890 deps = MICROKERNEL_TEST_DEPS,
5891)
5892
5893xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005894 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005895 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005896 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005897 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005898 ] + MICROKERNEL_TEST_HDRS,
5899 deps = MICROKERNEL_TEST_DEPS,
5900)
5901
5902xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005903 name = "f32_vmulc_relu_test",
5904 srcs = [
5905 "test/f32-vmulc-relu.cc",
5906 "test/vbinaryc-microkernel-tester.h",
5907 ] + MICROKERNEL_TEST_HDRS,
5908 deps = MICROKERNEL_TEST_DEPS,
5909)
5910
5911xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005912 name = "f16_vmulcaddc_minmax_test",
5913 srcs = [
5914 "test/f16-vmulcaddc-minmax.cc",
5915 "test/vmulcaddc-microkernel-tester.h",
5916 "src/xnnpack/AlignedAllocator.h",
5917 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005918 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005919)
5920
5921xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005922 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005923 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005924 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005925 "test/vmulcaddc-microkernel-tester.h",
5926 "src/xnnpack/AlignedAllocator.h",
5927 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005928 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005929)
5930
5931xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005932 name = "f32_vlrelu_test",
5933 srcs = [
5934 "test/f32-vlrelu.cc",
5935 "test/vunary-microkernel-tester.h",
5936 ] + MICROKERNEL_TEST_HDRS,
5937 deps = MICROKERNEL_TEST_DEPS,
5938)
5939
5940xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005941 name = "f32_vneg_test",
5942 srcs = [
5943 "test/f32-vneg.cc",
5944 "test/vunary-microkernel-tester.h",
5945 ] + MICROKERNEL_TEST_HDRS,
5946 deps = MICROKERNEL_TEST_DEPS,
5947)
5948
5949xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005950 name = "f32_vrndne_test",
5951 srcs = [
5952 "test/f32-vrndne.cc",
5953 "test/vunary-microkernel-tester.h",
5954 ] + MICROKERNEL_TEST_HDRS,
5955 deps = MICROKERNEL_TEST_DEPS,
5956)
5957
5958xnnpack_unit_test(
5959 name = "f32_vrndz_test",
5960 srcs = [
5961 "test/f32-vrndz.cc",
5962 "test/vunary-microkernel-tester.h",
5963 ] + MICROKERNEL_TEST_HDRS,
5964 deps = MICROKERNEL_TEST_DEPS,
5965)
5966
5967xnnpack_unit_test(
5968 name = "f32_vrndu_test",
5969 srcs = [
5970 "test/f32-vrndu.cc",
5971 "test/vunary-microkernel-tester.h",
5972 ] + MICROKERNEL_TEST_HDRS,
5973 deps = MICROKERNEL_TEST_DEPS,
5974)
5975
5976xnnpack_unit_test(
5977 name = "f32_vrndd_test",
5978 srcs = [
5979 "test/f32-vrndd.cc",
5980 "test/vunary-microkernel-tester.h",
5981 ] + MICROKERNEL_TEST_HDRS,
5982 deps = MICROKERNEL_TEST_DEPS,
5983)
5984
5985xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005986 name = "f32_vscale_test",
5987 srcs = [
5988 "test/f32-vscale.cc",
5989 "test/vscale-microkernel-tester.h",
5990 ] + MICROKERNEL_TEST_HDRS,
5991 deps = MICROKERNEL_TEST_DEPS,
5992)
5993
5994xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005995 name = "f32_vscaleexpminusmax_test",
5996 srcs = [
5997 "test/f32-vscaleexpminusmax.cc",
5998 "test/vscaleexpminusmax-microkernel-tester.h",
5999 ] + MICROKERNEL_TEST_HDRS,
6000 deps = MICROKERNEL_TEST_DEPS,
6001)
6002
6003xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006004 name = "f32_vscaleextexp_test",
6005 srcs = [
6006 "test/f32-vscaleextexp.cc",
6007 "test/vscaleextexp-microkernel-tester.h",
6008 ] + MICROKERNEL_TEST_HDRS,
6009 deps = MICROKERNEL_TEST_DEPS,
6010)
6011
6012xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006013 name = "f32_vsqr_test",
6014 srcs = [
6015 "test/f32-vsqr.cc",
6016 "test/vunary-microkernel-tester.h",
6017 ] + MICROKERNEL_TEST_HDRS,
6018 deps = MICROKERNEL_TEST_DEPS,
6019)
6020
6021xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006022 name = "f32_vsqrdiff_test",
6023 srcs = [
6024 "test/f32-vsqrdiff.cc",
6025 "test/vbinary-microkernel-tester.h",
6026 ] + MICROKERNEL_TEST_HDRS,
6027 deps = MICROKERNEL_TEST_DEPS,
6028)
6029
6030xnnpack_unit_test(
6031 name = "f32_vsqrdiffc_test",
6032 srcs = [
6033 "test/f32-vsqrdiffc.cc",
6034 "test/vbinaryc-microkernel-tester.h",
6035 ] + MICROKERNEL_TEST_HDRS,
6036 deps = MICROKERNEL_TEST_DEPS,
6037)
6038
6039xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006040 name = "f32_vsqrt_test",
6041 srcs = [
6042 "test/f32-vsqrt.cc",
6043 "test/vunary-microkernel-tester.h",
6044 ] + MICROKERNEL_TEST_HDRS,
6045 deps = MICROKERNEL_TEST_DEPS,
6046)
6047
6048xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006049 name = "f32_vsub_test",
6050 srcs = [
6051 "test/f32-vsub.cc",
6052 "test/vbinary-microkernel-tester.h",
6053 ] + MICROKERNEL_TEST_HDRS,
6054 deps = MICROKERNEL_TEST_DEPS,
6055)
6056
6057xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006058 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006059 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006060 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006061 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006062 ] + MICROKERNEL_TEST_HDRS,
6063 deps = MICROKERNEL_TEST_DEPS,
6064)
6065
6066xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006067 name = "f32_vsub_relu_test",
6068 srcs = [
6069 "test/f32-vsub-relu.cc",
6070 "test/vbinary-microkernel-tester.h",
6071 ] + MICROKERNEL_TEST_HDRS,
6072 deps = MICROKERNEL_TEST_DEPS,
6073)
6074
6075xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006076 name = "f32_vsubc_test",
6077 srcs = [
6078 "test/f32-vsubc.cc",
6079 "test/vbinaryc-microkernel-tester.h",
6080 ] + MICROKERNEL_TEST_HDRS,
6081 deps = MICROKERNEL_TEST_DEPS,
6082)
6083
6084xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006085 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006086 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006087 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006088 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006089 ] + MICROKERNEL_TEST_HDRS,
6090 deps = MICROKERNEL_TEST_DEPS,
6091)
6092
6093xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006094 name = "f32_vsubc_relu_test",
6095 srcs = [
6096 "test/f32-vsubc-relu.cc",
6097 "test/vbinaryc-microkernel-tester.h",
6098 ] + MICROKERNEL_TEST_HDRS,
6099 deps = MICROKERNEL_TEST_DEPS,
6100)
6101
6102xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006103 name = "f32_vrsubc_test",
6104 srcs = [
6105 "test/f32-vrsubc.cc",
6106 "test/vbinaryc-microkernel-tester.h",
6107 ] + MICROKERNEL_TEST_HDRS,
6108 deps = MICROKERNEL_TEST_DEPS,
6109)
6110
6111xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006112 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006113 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006114 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006115 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006116 ] + MICROKERNEL_TEST_HDRS,
6117 deps = MICROKERNEL_TEST_DEPS,
6118)
6119
6120xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006121 name = "f32_vrsubc_relu_test",
6122 srcs = [
6123 "test/f32-vrsubc-relu.cc",
6124 "test/vbinaryc-microkernel-tester.h",
6125 ] + MICROKERNEL_TEST_HDRS,
6126 deps = MICROKERNEL_TEST_DEPS,
6127)
6128
6129xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006130 name = "qs8_dwconv_minmax_test",
6131 srcs = [
6132 "test/qs8-dwconv-minmax.cc",
6133 "test/dwconv-microkernel-tester.h",
6134 "src/xnnpack/AlignedAllocator.h",
6135 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6136 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6137)
6138
6139xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006140 name = "qs8_gavgpool_minmax_test",
6141 srcs = [
6142 "test/qs8-gavgpool-minmax.cc",
6143 "test/gavgpool-microkernel-tester.h",
6144 "src/xnnpack/AlignedAllocator.h",
6145 ] + MICROKERNEL_TEST_HDRS,
6146 deps = MICROKERNEL_TEST_DEPS,
6147)
6148
6149xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006150 name = "qs8_gemm_minmax_test",
6151 srcs = [
6152 "test/qs8-gemm-minmax.cc",
6153 "test/gemm-microkernel-tester.h",
6154 "src/xnnpack/AlignedAllocator.h",
6155 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6156 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6157)
6158
6159xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006160 name = "qs8_igemm_minmax_test",
6161 srcs = [
6162 "test/qs8-igemm-minmax.cc",
6163 "test/gemm-microkernel-tester.h",
6164 "src/xnnpack/AlignedAllocator.h",
6165 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6166 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6167)
6168
6169xnnpack_unit_test(
6170 name = "qs8_requantization_test",
6171 srcs = [
6172 "src/xnnpack/requantization-stubs.h",
6173 "test/qs8-requantization.cc",
6174 "test/requantization-tester.h",
6175 ] + MICROKERNEL_TEST_HDRS,
6176 deps = MICROKERNEL_TEST_DEPS,
6177)
6178
6179xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006180 name = "qs8_vadd_minmax_test",
6181 srcs = [
6182 "test/qs8-vadd-minmax.cc",
6183 "test/vadd-microkernel-tester.h",
6184 ] + MICROKERNEL_TEST_HDRS,
6185 deps = MICROKERNEL_TEST_DEPS,
6186)
6187
6188xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006189 name = "qs8_vaddc_minmax_test",
6190 srcs = [
6191 "test/qs8-vaddc-minmax.cc",
6192 "test/vaddc-microkernel-tester.h",
6193 ] + MICROKERNEL_TEST_HDRS,
6194 deps = MICROKERNEL_TEST_DEPS,
6195)
6196
6197xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006198 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006199 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006200 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006201 "test/avgpool-microkernel-tester.h",
6202 "src/xnnpack/AlignedAllocator.h",
6203 ] + MICROKERNEL_TEST_HDRS,
6204 deps = MICROKERNEL_TEST_DEPS,
6205)
6206
6207xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006208 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006209 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006210 "test/qu8-dwconv-minmax.cc",
6211 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006212 "src/xnnpack/AlignedAllocator.h",
6213 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006214 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006215)
6216
6217xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006218 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006219 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006220 "test/qu8-igemm-minmax.cc",
6221 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006222 "src/xnnpack/AlignedAllocator.h",
6223 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006224 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006225)
6226
6227xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006228 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006229 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006230 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006231 "test/gavgpool-microkernel-tester.h",
6232 "src/xnnpack/AlignedAllocator.h",
6233 ] + MICROKERNEL_TEST_HDRS,
6234 deps = MICROKERNEL_TEST_DEPS,
6235)
6236
6237xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006238 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006239 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006240 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006241 "test/gemm-microkernel-tester.h",
6242 "src/xnnpack/AlignedAllocator.h",
6243 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006244 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006245)
6246
6247xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006248 name = "qu8_requantization_test",
6249 srcs = [
6250 "src/xnnpack/requantization-stubs.h",
6251 "test/qu8-requantization.cc",
6252 "test/requantization-tester.h",
6253 ] + MICROKERNEL_TEST_HDRS,
6254 deps = MICROKERNEL_TEST_DEPS,
6255)
6256
6257xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006258 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006259 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006260 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006261 "test/vadd-microkernel-tester.h",
6262 ] + MICROKERNEL_TEST_HDRS,
6263 deps = MICROKERNEL_TEST_DEPS,
6264)
6265
6266xnnpack_unit_test(
6267 name = "u8_clamp_test",
6268 srcs = [
6269 "test/u8-clamp.cc",
6270 "test/clamp-microkernel-tester.h",
6271 ] + MICROKERNEL_TEST_HDRS,
6272 deps = MICROKERNEL_TEST_DEPS,
6273)
6274
6275xnnpack_unit_test(
6276 name = "u8_lut32norm_test",
6277 srcs = [
6278 "test/u8-lut32norm.cc",
6279 "test/lut-norm-microkernel-tester.h",
6280 ] + MICROKERNEL_TEST_HDRS,
6281 deps = MICROKERNEL_TEST_DEPS,
6282)
6283
6284xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006285 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006286 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006287 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006288 "test/maxpool-microkernel-tester.h",
6289 ] + MICROKERNEL_TEST_HDRS,
6290 deps = MICROKERNEL_TEST_DEPS,
6291)
6292
6293xnnpack_unit_test(
6294 name = "u8_rmax_test",
6295 srcs = [
6296 "test/u8-rmax.cc",
6297 "test/rmax-microkernel-tester.h",
6298 ] + MICROKERNEL_TEST_HDRS,
6299 deps = MICROKERNEL_TEST_DEPS,
6300)
6301
6302xnnpack_unit_test(
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006303 name = "x32_depthtospace2d_chw2hwc_test",
Yury Kartynnike7841862020-11-04 18:22:18 -08006304 srcs = [
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006305 "test/x32-depthtospace2d-chw2hwc.cc",
6306 "test/depthtospace-microkernel-tester.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08006307 ] + MICROKERNEL_TEST_HDRS,
6308 deps = MICROKERNEL_TEST_DEPS,
6309)
6310
6311xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006312 name = "x32_fill_test",
6313 srcs = [
6314 "test/x32-fill.cc",
6315 "test/fill-microkernel-tester.h",
6316 ] + MICROKERNEL_TEST_HDRS,
6317 deps = MICROKERNEL_TEST_DEPS,
6318)
6319
6320xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006321 name = "x32_packx_test",
6322 srcs = [
6323 "test/x32-packx.cc",
6324 "test/pack-microkernel-tester.h",
6325 "src/xnnpack/AlignedAllocator.h",
6326 ] + MICROKERNEL_TEST_HDRS,
6327 deps = MICROKERNEL_TEST_DEPS,
6328)
6329
6330xnnpack_unit_test(
6331 name = "x32_pad_test",
6332 srcs = [
6333 "test/x32-pad.cc",
6334 "test/pad-microkernel-tester.h",
6335 ] + MICROKERNEL_TEST_HDRS,
6336 deps = MICROKERNEL_TEST_DEPS,
6337)
6338
6339xnnpack_unit_test(
6340 name = "x32_unpool_test",
6341 srcs = [
6342 "test/x32-unpool.cc",
6343 "test/unpool-microkernel-tester.h",
6344 ] + MICROKERNEL_TEST_HDRS,
6345 deps = MICROKERNEL_TEST_DEPS,
6346)
6347
6348xnnpack_unit_test(
6349 name = "x32_zip_test",
6350 srcs = [
6351 "test/x32-zip.cc",
6352 "test/zip-microkernel-tester.h",
6353 ] + MICROKERNEL_TEST_HDRS,
6354 deps = MICROKERNEL_TEST_DEPS,
6355)
6356
6357xnnpack_unit_test(
6358 name = "x8_lut_test",
6359 srcs = [
6360 "test/x8-lut.cc",
6361 "test/lut-microkernel-tester.h",
6362 ] + MICROKERNEL_TEST_HDRS,
6363 deps = MICROKERNEL_TEST_DEPS,
6364)
6365
6366xnnpack_unit_test(
6367 name = "x8_zip_test",
6368 srcs = [
6369 "test/x8-zip.cc",
6370 "test/zip-microkernel-tester.h",
6371 ] + MICROKERNEL_TEST_HDRS,
6372 deps = MICROKERNEL_TEST_DEPS,
6373)
6374
Marat Dukhan20c3b922020-03-10 03:45:06 -07006375########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006376
6377xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006378 name = "operator_size_test",
6379 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006380 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006381)
6382
Marat Dukhan20c3b922020-03-10 03:45:06 -07006383xnnpack_binary(
6384 name = "subgraph_size_test",
6385 srcs = ["test/subgraph-size.c"],
6386 deps = [":XNNPACK"],
6387)
6388
6389########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006390
6391xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006392 name = "abs_nc_test",
6393 srcs = [
6394 "test/abs-nc.cc",
6395 "test/abs-operator-tester.h",
6396 ],
6397 deps = OPERATOR_TEST_DEPS,
6398)
6399
6400xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006401 name = "add_nd_test",
6402 srcs = [
6403 "test/add-nd.cc",
6404 "test/binary-elementwise-operator-tester.h",
6405 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006406 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006407)
6408
6409xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006410 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006411 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006412 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006413 "test/argmax-pooling-operator-tester.h",
6414 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006415 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006416)
6417
6418xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006419 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006420 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006421 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006422 "test/average-pooling-operator-tester.h",
6423 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006424 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006425)
6426
6427xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006428 name = "bankers_rounding_nc_test",
6429 srcs = [
6430 "test/bankers-rounding-nc.cc",
6431 "test/bankers-rounding-operator-tester.h",
6432 ],
6433 deps = OPERATOR_TEST_DEPS,
6434)
6435
6436xnnpack_unit_test(
6437 name = "ceiling_nc_test",
6438 srcs = [
6439 "test/ceiling-nc.cc",
6440 "test/ceiling-operator-tester.h",
6441 ],
6442 deps = OPERATOR_TEST_DEPS,
6443)
6444
6445xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006446 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006447 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006448 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006449 "test/channel-shuffle-operator-tester.h",
6450 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006451 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006452)
6453
6454xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006455 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006456 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006457 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006458 "test/clamp-operator-tester.h",
6459 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006460 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006461)
6462
6463xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006464 name = "constant_pad_nd_test",
6465 srcs = [
6466 "test/constant-pad-nd.cc",
6467 "test/constant-pad-operator-tester.h",
6468 ],
6469 deps = OPERATOR_TEST_DEPS,
6470)
6471
6472xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006473 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006474 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006475 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006476 "test/convolution-operator-tester.h",
6477 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006478 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006479)
6480
6481xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006482 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006483 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006484 "test/convolution-nchw.cc",
6485 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006486 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006487 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006488)
6489
6490xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006491 name = "copy_nc_test",
6492 srcs = [
6493 "test/copy-nc.cc",
6494 "test/copy-operator-tester.h",
6495 ],
6496 deps = OPERATOR_TEST_DEPS,
6497)
6498
6499xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006500 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006501 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006502 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006503 "test/deconvolution-operator-tester.h",
6504 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006505 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006506)
6507
6508xnnpack_unit_test(
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08006509 name = "depth_to_space_test",
6510 srcs = [
6511 "test/depth-to-space.cc",
6512 "test/depth-to-space-operator-tester.h",
6513 ] + OPERATOR_TEST_PARAMS_HDRS,
6514 deps = OPERATOR_TEST_DEPS,
6515)
6516
6517xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006518 name = "divide_nd_test",
6519 srcs = [
6520 "test/binary-elementwise-operator-tester.h",
6521 "test/divide-nd.cc",
6522 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006523 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006524)
6525
6526xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006527 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006528 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006529 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006530 "test/fully-connected-operator-tester.h",
6531 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006532 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006533)
6534
6535xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006536 name = "floor_nc_test",
6537 srcs = [
6538 "test/floor-nc.cc",
6539 "test/floor-operator-tester.h",
6540 ],
6541 deps = OPERATOR_TEST_DEPS,
6542)
6543
6544xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006545 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006546 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006547 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006548 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006549 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006550 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006551)
6552
6553xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006554 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006555 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006556 "test/global-average-pooling-ncw.cc",
6557 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006558 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006559 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006560)
6561
6562xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006563 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006564 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006565 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006566 "test/hardswish-operator-tester.h",
6567 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006568 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006569)
6570
6571xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006572 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006573 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006574 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006575 "test/leaky-relu-operator-tester.h",
6576 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006577 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006578)
6579
6580xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006581 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006582 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006583 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006584 "test/max-pooling-operator-tester.h",
6585 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006586 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006587)
6588
6589xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006590 name = "maximum_nd_test",
6591 srcs = [
6592 "test/binary-elementwise-operator-tester.h",
6593 "test/maximum-nd.cc",
6594 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006595 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006596)
6597
6598xnnpack_unit_test(
6599 name = "minimum_nd_test",
6600 srcs = [
6601 "test/binary-elementwise-operator-tester.h",
6602 "test/minimum-nd.cc",
6603 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006604 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006605)
6606
6607xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006608 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006609 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006610 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006611 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006612 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006613 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006614)
6615
6616xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006617 name = "negate_nc_test",
6618 srcs = [
6619 "test/negate-nc.cc",
6620 "test/negate-operator-tester.h",
6621 ],
6622 deps = OPERATOR_TEST_DEPS,
6623)
6624
6625xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006626 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006627 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006628 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006629 "test/prelu-operator-tester.h",
6630 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006631 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006632)
6633
6634xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006635 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006636 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006637 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006638 "test/resize-bilinear-operator-tester.h",
6639 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006640 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006641)
6642
6643xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07006644 name = "resize_bilinear_nchw_test",
6645 srcs = [
6646 "test/resize-bilinear-nchw.cc",
6647 "test/resize-bilinear-operator-tester.h",
6648 ] + OPERATOR_TEST_PARAMS_HDRS,
6649 deps = OPERATOR_TEST_DEPS,
6650)
6651
6652xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006653 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006654 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006655 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006656 "test/sigmoid-operator-tester.h",
6657 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006658 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006659)
6660
6661xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006662 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006663 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006664 "test/softmax-nc.cc",
6665 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006666 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006667 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006668)
6669
6670xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006671 name = "square_nc_test",
6672 srcs = [
6673 "test/square-nc.cc",
6674 "test/square-operator-tester.h",
6675 ],
6676 deps = OPERATOR_TEST_DEPS,
6677)
6678
6679xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006680 name = "square_root_nc_test",
6681 srcs = [
6682 "test/square-root-nc.cc",
6683 "test/square-root-operator-tester.h",
6684 ],
6685 deps = OPERATOR_TEST_DEPS,
6686)
6687
6688xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006689 name = "squared_difference_nd_test",
6690 srcs = [
6691 "test/binary-elementwise-operator-tester.h",
6692 "test/squared-difference-nd.cc",
6693 ],
6694 deps = OPERATOR_TEST_DEPS,
6695)
6696
6697xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006698 name = "subtract_nd_test",
6699 srcs = [
6700 "test/binary-elementwise-operator-tester.h",
6701 "test/subtract-nd.cc",
6702 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006703 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006704)
6705
6706xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006707 name = "truncation_nc_test",
6708 srcs = [
6709 "test/truncation-nc.cc",
6710 "test/truncation-operator-tester.h",
6711 ],
6712 deps = OPERATOR_TEST_DEPS,
6713)
6714
6715xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006716 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006717 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006718 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006719 "test/unpooling-operator-tester.h",
6720 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006721 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006722)
6723
Chao Mei6ddfc602020-05-13 22:29:36 -07006724############################### Misc unit tests ###############################
6725
6726xnnpack_unit_test(
6727 name = "memory_planner_test",
6728 srcs = [
6729 "test/memory-planner-test.cc",
6730 ],
6731 deps = [
6732 ":XNNPACK",
6733 ":memory_planner",
6734 ],
6735)
6736
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07006737xnnpack_unit_test(
6738 name = "subgraph_nchw_test",
6739 srcs = [
6740 "src/xnnpack/subgraph.h",
6741 "test/subgraph-nchw.cc",
6742 "test/subgraph-tester.h",
6743 ],
6744 deps = [
6745 ":XNNPACK",
6746 ],
6747)
6748
Marat Dukhan08c4a432019-10-03 09:29:21 -07006749############################# Build configurations #############################
6750
Marat Dukhanb8642352019-10-30 15:43:02 -07006751# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006752config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006753 name = "xnn_enable_assembly_explicit_true",
6754 define_values = {"xnn_enable_assembly": "true"},
6755)
6756
6757# Disables usage of assembly kernels.
6758config_setting(
6759 name = "xnn_enable_assembly_explicit_false",
6760 define_values = {"xnn_enable_assembly": "false"},
6761)
6762
Marat Dukhan9de90e02020-06-18 16:04:12 -07006763# Enables usage of sparse inference.
6764config_setting(
6765 name = "xnn_enable_sparse_explicit_true",
6766 define_values = {"xnn_enable_sparse": "true"},
6767)
6768
6769# Disables usage of sparse inference.
6770config_setting(
6771 name = "xnn_enable_sparse_explicit_false",
6772 define_values = {"xnn_enable_sparse": "false"},
6773)
6774
Marat Dukhan05702cf2020-03-26 15:41:33 -07006775# Disables usage of HMP-aware optimizations.
6776config_setting(
6777 name = "xnn_enable_hmp_explicit_false",
6778 define_values = {"xnn_enable_hmp": "false"},
6779)
6780
Chao Mei6ddfc602020-05-13 22:29:36 -07006781# Enable usage of optimized memory allocation
6782config_setting(
6783 name = "xnn_enable_memopt_explicit_true",
6784 define_values = {"xnnpack_enable_memopt": "true"},
6785)
6786
6787# Disable usage of optimized memory allocation
6788config_setting(
6789 name = "xnn_enable_memopt_explicit_false",
6790 define_values = {"xnnpack_enable_memopt": "false"},
6791)
6792
Marat Dukhanb8642352019-10-30 15:43:02 -07006793# Builds with -c dbg
6794config_setting(
6795 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006796 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006797 "compilation_mode": "dbg",
6798 },
6799)
6800
6801# Builds with -c opt
6802config_setting(
6803 name = "optimized_build",
6804 values = {
6805 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006806 },
6807)
6808
6809config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006810 name = "linux_k8",
6811 values = {"cpu": "k8"},
6812)
6813
6814config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006815 name = "linux_arm",
6816 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006817)
6818
6819config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006820 name = "linux_armeabi",
6821 values = {"cpu": "armeabi"},
6822)
6823
6824config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006825 name = "linux_armhf",
6826 values = {"cpu": "armhf"},
6827)
6828
6829config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006830 name = "linux_armv7a",
6831 values = {"cpu": "armv7a"},
6832)
6833
6834config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006835 name = "linux_aarch64",
6836 values = {"cpu": "aarch64"},
6837)
6838
6839config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006840 name = "android",
6841 values = {"crosstool_top": "//external:android/crosstool"},
6842)
6843
6844config_setting(
6845 name = "android_armv7",
6846 values = {
6847 "crosstool_top": "//external:android/crosstool",
6848 "cpu": "armeabi-v7a",
6849 },
6850)
6851
6852config_setting(
6853 name = "android_arm64",
6854 values = {
6855 "crosstool_top": "//external:android/crosstool",
6856 "cpu": "arm64-v8a",
6857 },
6858)
6859
6860config_setting(
6861 name = "android_x86",
6862 values = {
6863 "crosstool_top": "//external:android/crosstool",
6864 "cpu": "x86",
6865 },
6866)
6867
6868config_setting(
6869 name = "android_x86_64",
6870 values = {
6871 "crosstool_top": "//external:android/crosstool",
6872 "cpu": "x86_64",
6873 },
6874)
6875
6876config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006877 name = "windows_x86_64",
6878 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006879)
6880
6881config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006882 name = "windows_x86_64_clang",
6883 values = {
6884 "compiler": "clang-cl",
6885 "cpu": "x64_windows",
6886 },
6887)
6888
6889config_setting(
6890 name = "windows_x86_64_mingw",
6891 values = {
6892 "compiler": "mingw-gcc",
6893 "cpu": "x64_windows",
6894 },
6895)
6896
6897config_setting(
6898 name = "windows_x86_64_msys",
6899 values = {
6900 "compiler": "msys-gcc",
6901 "cpu": "x64_windows",
6902 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006903)
6904
6905config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006906 name = "macos_x86_64",
6907 values = {
6908 "apple_platform_type": "macos",
6909 "cpu": "darwin",
6910 },
6911)
6912
6913config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006914 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006915 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006916)
6917
6918config_setting(
6919 name = "emscripten_wasm",
6920 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006921 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006922 "cpu": "wasm",
6923 },
6924)
6925
6926config_setting(
6927 name = "emscripten_wasmsimd",
6928 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006929 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006930 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006931 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006932 },
6933)
6934
6935config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006936 name = "ios_armv7",
6937 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006938 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006939 "cpu": "ios_armv7",
6940 },
6941)
6942
6943config_setting(
6944 name = "ios_arm64",
6945 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006946 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006947 "cpu": "ios_arm64",
6948 },
6949)
6950
6951config_setting(
6952 name = "ios_arm64e",
6953 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006954 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006955 "cpu": "ios_arm64e",
6956 },
6957)
6958
6959config_setting(
6960 name = "ios_x86",
6961 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006962 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006963 "cpu": "ios_i386",
6964 },
6965)
6966
6967config_setting(
6968 name = "ios_x86_64",
6969 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006970 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006971 "cpu": "ios_x86_64",
6972 },
6973)
6974
6975config_setting(
6976 name = "watchos_armv7k",
6977 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006978 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006979 "cpu": "watchos_armv7k",
6980 },
6981)
6982
6983config_setting(
6984 name = "watchos_arm64_32",
6985 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006986 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006987 "cpu": "watchos_arm64_32",
6988 },
6989)
6990
6991config_setting(
6992 name = "watchos_x86",
6993 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006994 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006995 "cpu": "watchos_i386",
6996 },
6997)
6998
6999config_setting(
7000 name = "watchos_x86_64",
7001 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007002 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007003 "cpu": "watchos_x86_64",
7004 },
7005)
7006
7007config_setting(
7008 name = "tvos_arm64",
7009 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007010 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007011 "cpu": "tvos_arm64",
7012 },
7013)
7014
7015config_setting(
7016 name = "tvos_x86_64",
7017 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007018 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007019 "cpu": "tvos_x86_64",
7020 },
7021)