blob: 573f206dac4bd91442523abbf6cf426a92891925 [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 Dukhan0e521172020-11-25 13:10:04 -080063 "src/operators/depth-to-space-nhwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070064 "src/operators/fully-connected-nc.c",
65 "src/operators/global-average-pooling-ncw.c",
66 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070067 "src/operators/leaky-relu-nc.c",
68 "src/operators/max-pooling-nhwc.c",
69 "src/operators/prelu-nc.c",
Artsiom Ablavatski97918102020-10-27 15:52:59 -070070 "src/operators/resize-bilinear-nchw.c",
Marat Dukhane8265432020-04-28 18:42:59 -070071 "src/operators/resize-bilinear-nhwc.c",
72 "src/operators/sigmoid-nc.c",
73 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070074 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070075 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070076]
77
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070078SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070079 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070080 "src/subgraph/add2.c",
81 "src/subgraph/argmax-pooling-2d.c",
82 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070083 "src/subgraph/bankers-rounding.c",
84 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070085 "src/subgraph/clamp.c",
86 "src/subgraph/convolution-2d.c",
87 "src/subgraph/deconvolution-2d.c",
Artsiom Ablavatskibbe85062020-11-05 14:07:37 -080088 "src/subgraph/depth-to-space.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -080089 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070090 "src/subgraph/divide.c",
Marat Dukhana1600202020-12-01 22:17:16 -080091 "src/subgraph/elu.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070092 "src/subgraph/floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -070093 "src/subgraph/fully-connected.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070094 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070096 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070098 "src/subgraph/maximum2.c",
99 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700100 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700101 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700102 "src/subgraph/prelu.c",
103 "src/subgraph/sigmoid.c",
104 "src/subgraph/softmax.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700105 "src/subgraph/square-root.c",
106 "src/subgraph/square.c",
107 "src/subgraph/squared-difference.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700108 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700109 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700110 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700111 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700112 "src/subgraph/unpooling-2d.c",
113]
114
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800115TABLE_SRCS = [
116 "src/tables/exp2-k-over-64.c",
117 "src/tables/exp2-k-over-2048.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800118 "src/tables/exp2minus-k-over-4.c",
119 "src/tables/exp2minus-k-over-8.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800120 "src/tables/exp2minus-k-over-16.c",
Marat Dukhan1f256fc2020-09-24 21:27:14 -0700121 "src/tables/exp2minus-k-over-64.c",
122 "src/tables/exp2minus-k-over-2048.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800123]
124
Marat Dukhan08c4a432019-10-03 09:29:21 -0700125SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800126 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800127 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800128 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700129 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
130 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700131 "src/f32-clamp/gen/scalar-x1.c",
132 "src/f32-clamp/gen/scalar-x2.c",
133 "src/f32-clamp/gen/scalar-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700135 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700136 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700137 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700138 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
139 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
140 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700141 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700142 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
143 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
144 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700145 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700146 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
147 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
148 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700150 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
151 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
152 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700153 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700154 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
155 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
156 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700157 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700158 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
159 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
160 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700161 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c",
162 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c",
163 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700164 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700165 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700166 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c",
167 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c",
168 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c",
169 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c",
170 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700171 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c",
172 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c",
173 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700174 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700175 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700176 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c",
177 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c",
178 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700179 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c",
180 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c",
181 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c",
182 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700183 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700184 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c",
185 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700186 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700187 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700188 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c",
Marat Dukhan29c0c332020-10-28 22:11:00 -0700189 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c",
190 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c",
191 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c",
192 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c",
193 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c",
194 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c",
195 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c",
196 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c",
197 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c",
198 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700199 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700200 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
201 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700202 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
203 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
204 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700205 "src/f32-gemm/gen/1x4-minmax-scalar.c",
206 "src/f32-gemm/gen/1x4-relu-scalar.c",
207 "src/f32-gemm/gen/1x4-scalar.c",
208 "src/f32-gemm/gen/2x4-minmax-scalar.c",
209 "src/f32-gemm/gen/2x4-relu-scalar.c",
210 "src/f32-gemm/gen/2x4-scalar.c",
211 "src/f32-gemm/gen/4x2-minmax-scalar.c",
212 "src/f32-gemm/gen/4x2-relu-scalar.c",
213 "src/f32-gemm/gen/4x2-scalar.c",
214 "src/f32-gemm/gen/4x4-minmax-scalar.c",
215 "src/f32-gemm/gen/4x4-relu-scalar.c",
216 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700217 "src/f32-hswish/gen/hswish-scalar-x1.c",
218 "src/f32-hswish/gen/hswish-scalar-x2.c",
219 "src/f32-hswish/gen/hswish-scalar-x4.c",
XNNPACK Team21432672020-10-19 19:58:48 -0700220 "src/f32-ibilinear-chw/gen/scalar-p1.c",
221 "src/f32-ibilinear-chw/gen/scalar-p2.c",
222 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700223 "src/f32-ibilinear/gen/scalar-c1.c",
224 "src/f32-ibilinear/gen/scalar-c2.c",
225 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700226 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700227 "src/f32-igemm/gen/1x4-relu-scalar.c",
228 "src/f32-igemm/gen/1x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700229 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700230 "src/f32-igemm/gen/2x4-relu-scalar.c",
231 "src/f32-igemm/gen/2x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700232 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700233 "src/f32-igemm/gen/4x2-relu-scalar.c",
234 "src/f32-igemm/gen/4x2-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700235 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700236 "src/f32-igemm/gen/4x4-relu-scalar.c",
237 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700238 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
239 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
240 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700241 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
242 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
243 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
244 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800245 "src/f32-prelu/gen/scalar-2x1.c",
246 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800247 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800248 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700249 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800250 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
251 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700252 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800253 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800254 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700255 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800256 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
257 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700258 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700259 "src/f32-relu/gen/scalar-x1.c",
260 "src/f32-relu/gen/scalar-x2.c",
261 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700262 "src/f32-relu/gen/scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700263 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800264 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
265 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
266 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700267 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
268 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
269 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800270 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
271 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
272 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700273 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
274 "src/f32-spmm/gen/1x1-minmax-scalar.c",
275 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
276 "src/f32-spmm/gen/2x1-minmax-scalar.c",
277 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
278 "src/f32-spmm/gen/4x1-minmax-scalar.c",
279 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
280 "src/f32-spmm/gen/8x1-minmax-scalar.c",
281 "src/f32-spmm/gen/8x2-minmax-scalar.c",
282 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700283 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
284 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
285 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700286 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700287 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
288 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
289 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700290 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700291 "src/f32-vbinary/gen/vadd-scalar-x1.c",
292 "src/f32-vbinary/gen/vadd-scalar-x2.c",
293 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700294 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700295 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
296 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
297 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700298 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700299 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
300 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
301 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700302 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700303 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
304 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
305 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700306 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700307 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
308 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
309 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700310 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700311 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
312 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
313 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700314 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700315 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
316 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
317 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700318 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700319 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
320 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
321 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700322 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700323 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
324 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
325 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700326 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700327 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
328 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
329 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700330 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800331 "src/f32-vbinary/gen/vmax-scalar-x1.c",
332 "src/f32-vbinary/gen/vmax-scalar-x2.c",
333 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700334 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800335 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
336 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
337 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700338 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800339 "src/f32-vbinary/gen/vmin-scalar-x1.c",
340 "src/f32-vbinary/gen/vmin-scalar-x2.c",
341 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700342 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800343 "src/f32-vbinary/gen/vminc-scalar-x1.c",
344 "src/f32-vbinary/gen/vminc-scalar-x2.c",
345 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700346 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700347 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
348 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
349 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700350 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700351 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
352 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
353 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700354 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700355 "src/f32-vbinary/gen/vmul-scalar-x1.c",
356 "src/f32-vbinary/gen/vmul-scalar-x2.c",
357 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700358 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700359 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
360 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
361 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700362 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700363 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
364 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
365 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700366 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700367 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
368 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
369 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700370 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700371 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
372 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
373 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700374 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700375 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
376 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
377 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700378 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700379 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
380 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
381 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700382 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700383 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
384 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
385 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700386 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700387 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
388 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
389 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700390 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700391 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
392 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
393 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700394 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700395 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
396 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
397 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700398 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700399 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
400 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
401 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700402 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700403 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
404 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
405 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700406 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700407 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
408 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
409 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700410 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700411 "src/f32-vbinary/gen/vsub-scalar-x1.c",
412 "src/f32-vbinary/gen/vsub-scalar-x2.c",
413 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700414 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700415 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
416 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
417 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700418 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700419 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
420 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
421 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700422 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700423 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
424 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
425 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700426 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800427 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x1.c",
428 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x2.c",
429 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x3.c",
430 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x4.c",
431 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x5.c",
432 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x6.c",
433 "src/f32-velu/gen/velu-scalar-rr2-p6-x1.c",
434 "src/f32-velu/gen/velu-scalar-rr2-p6-x2.c",
435 "src/f32-velu/gen/velu-scalar-rr2-p6-x3.c",
436 "src/f32-velu/gen/velu-scalar-rr2-p6-x4.c",
437 "src/f32-velu/gen/velu-scalar-rr2-p6-x5.c",
438 "src/f32-velu/gen/velu-scalar-rr2-p6-x6.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700439 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
440 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
441 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700442 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
443 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
444 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700445 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
446 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
447 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700448 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
449 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
450 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
451 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
452 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
453 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
454 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
455 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
456 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700457 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
458 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
459 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700460 "src/f32-vunary/gen/vabs-scalar-x1.c",
461 "src/f32-vunary/gen/vabs-scalar-x2.c",
462 "src/f32-vunary/gen/vabs-scalar-x4.c",
463 "src/f32-vunary/gen/vneg-scalar-x1.c",
464 "src/f32-vunary/gen/vneg-scalar-x2.c",
465 "src/f32-vunary/gen/vneg-scalar-x4.c",
466 "src/f32-vunary/gen/vsqr-scalar-x1.c",
467 "src/f32-vunary/gen/vsqr-scalar-x2.c",
468 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800469 "src/math/expm1minus-scalar-rr2-lut4-p4.c",
470 "src/math/expm1minus-scalar-rr2-lut8-p3.c",
471 "src/math/expm1minus-scalar-rr2-lut8-p4.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800472 "src/math/expm1minus-scalar-rr2-lut16-p3.c",
473 "src/math/expm1minus-scalar-rr2-lut16-p4.c",
474 "src/math/expm1minus-scalar-rr2-p5.c",
475 "src/math/expm1minus-scalar-rr2-p6.c",
Frank Barchard22136062020-11-24 18:44:46 -0800476 "src/math/expminus-scalar-rr2-lut64-p2.c",
477 "src/math/expminus-scalar-rr2-lut2048-p1.c",
478 "src/math/expminus-scalar-rr2-p5.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700479 "src/math/roundd-scalar-addsub.c",
480 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700481 "src/math/roundd-scalar-floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700482 "src/math/roundne-scalar-addsub.c",
483 "src/math/roundne-scalar-nearbyint.c",
484 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700485 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700486 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700487 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700488 "src/math/roundz-scalar-addsub.c",
489 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700490 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700491 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700492 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700493 "src/math/sigmoid-scalar-rr2-p5-div.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700494 "src/qs8-requantization/fp32-scalar-lrintf.c",
495 "src/qs8-requantization/fp32-scalar-magic.c",
496 "src/qs8-requantization/precise-scalar-signed64.c",
497 "src/qs8-requantization/precise-scalar-unsigned32.c",
498 "src/qs8-requantization/precise-scalar-unsigned64.c",
499 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700500 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
501 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
502 "src/qu8-dwconv/up1x9-minmax-scalar.c",
503 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
504 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700505 "src/qu8-gemm/2x2-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700506 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700507 "src/qu8-requantization/fp32-scalar-lrintf.c",
508 "src/qu8-requantization/fp32-scalar-magic.c",
509 "src/qu8-requantization/precise-scalar-signed64.c",
510 "src/qu8-requantization/precise-scalar-unsigned32.c",
511 "src/qu8-requantization/precise-scalar-unsigned64.c",
512 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700513 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700514 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700515 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700516 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700517 "src/u8-rmax/scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700518 "src/x8-lut/scalar.c",
519 "src/x8-zip/x2-scalar.c",
520 "src/x8-zip/x3-scalar.c",
521 "src/x8-zip/x4-scalar.c",
522 "src/x8-zip/xm-scalar.c",
Marat Dukhanad71b9a2020-11-20 00:01:51 -0800523 "src/x32-depthtospace2d-chw2hwc/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700524 "src/x32-fill/scalar-float.c",
525 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700526 "src/x32-packx/x2-scalar.c",
527 "src/x32-packx/x3-scalar.c",
528 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700529 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700530 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700531 "src/x32-unpool/scalar.c",
532 "src/x32-zip/x2-scalar.c",
533 "src/x32-zip/x3-scalar.c",
534 "src/x32-zip/x4-scalar.c",
535 "src/x32-zip/xm-scalar.c",
Marat Dukhan048931b2020-11-24 20:53:54 -0800536 "src/xx-copy/memcpy.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700537]
538
Marat Dukhan436ebe62019-12-04 15:10:12 -0800539WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700540 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
541 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700542 "src/f32-clamp/gen/wasm-x1.c",
543 "src/f32-clamp/gen/wasm-x2.c",
544 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700545 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
546 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700547 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
548 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700549 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
550 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700551 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
552 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700553 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
554 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700555 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
556 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700557 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
558 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700559 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
560 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700561 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
562 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700563 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
564 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700565 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
566 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700567 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
568 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700569 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
570 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700571 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
572 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
573 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
574 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700575 "src/f32-gemm/gen/1x4-relu-wasm.c",
576 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700577 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700578 "src/f32-gemm/gen/2x4-relu-wasm.c",
579 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700580 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700581 "src/f32-gemm/gen/4x2-relu-wasm.c",
582 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700583 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700584 "src/f32-gemm/gen/4x4-relu-wasm.c",
585 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700586 "src/f32-hswish/gen/hswish-wasm-x1.c",
587 "src/f32-hswish/gen/hswish-wasm-x2.c",
588 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700589 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700590 "src/f32-igemm/gen/1x4-relu-wasm.c",
591 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700592 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700593 "src/f32-igemm/gen/2x4-relu-wasm.c",
594 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700595 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700596 "src/f32-igemm/gen/4x2-relu-wasm.c",
597 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700598 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700599 "src/f32-igemm/gen/4x4-relu-wasm.c",
600 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700601 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
602 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
603 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700604 "src/f32-prelu/gen/wasm-2x1.c",
605 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700606 "src/f32-relu/gen/wasm-x1.c",
607 "src/f32-relu/gen/wasm-x2.c",
608 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700609 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700610 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
611 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
612 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700613 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700614 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
615 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
616 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700617 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700618 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
619 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
620 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
621 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700622 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
623 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
624 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700625 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700626 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
627 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
628 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
629 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700630 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
631 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
632 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700633 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700634 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
635 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
636 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
637 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700638 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
639 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
640 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700641 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800642 "src/f32-vbinary/gen/vmax-wasm-x1.c",
643 "src/f32-vbinary/gen/vmax-wasm-x2.c",
644 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700645 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800646 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
647 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
648 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700649 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800650 "src/f32-vbinary/gen/vmin-wasm-x1.c",
651 "src/f32-vbinary/gen/vmin-wasm-x2.c",
652 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700653 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800654 "src/f32-vbinary/gen/vminc-wasm-x1.c",
655 "src/f32-vbinary/gen/vminc-wasm-x2.c",
656 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700657 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700658 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
659 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
660 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700661 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700662 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
663 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
664 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700665 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700666 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
667 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
668 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
669 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700670 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
671 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
672 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700673 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700674 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
675 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
676 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
677 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700678 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
679 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
680 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700681 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700682 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
683 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
684 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
685 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700686 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
687 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
688 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700689 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700690 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
691 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
692 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
693 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700694 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
695 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
696 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700697 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700698 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
699 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
700 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
701 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700702 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
703 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
704 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700705 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800706 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x1.c",
707 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x2.c",
708 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x3.c",
709 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x4.c",
710 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x5.c",
711 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x6.c",
712 "src/f32-velu/gen/velu-wasm-rr2-p6-x1.c",
713 "src/f32-velu/gen/velu-wasm-rr2-p6-x2.c",
714 "src/f32-velu/gen/velu-wasm-rr2-p6-x3.c",
715 "src/f32-velu/gen/velu-wasm-rr2-p6-x4.c",
716 "src/f32-velu/gen/velu-wasm-rr2-p6-x5.c",
717 "src/f32-velu/gen/velu-wasm-rr2-p6-x6.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700718 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
719 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
720 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700721 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
722 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
723 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800724]
725
Marat Dukhan290055c2020-06-09 12:24:29 -0700726WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700727 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
728 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
729 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700730 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
731 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
732 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
733 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700734 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
735 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
736 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
737 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Frank Barchard22136062020-11-24 18:44:46 -0800738 "src/f32-conv-hwc2chw/3x3s2p1c3x4-wasmsimd-2x2.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800739 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700740 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800741 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700742 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700743 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800744 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700745 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800746 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700747 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700748 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800749 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700750 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800751 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700752 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
753 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800754 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700755 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800756 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700757 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700758 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800759 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700760 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800761 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86-acc2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700762 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700763 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800764 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700765 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800766 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700767 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
768 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800769 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
770 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
771 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
772 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4.c",
773 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
774 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4.c",
775 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-3x4.c",
776 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-4x4.c",
777 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-5x4.c",
778 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-6x4.c",
Frank Barchard02bb4292020-12-15 18:25:32 -0800779 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc2.c",
780 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc3.c",
781 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc4.c",
782 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4.c",
783 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-2x4-acc2.c",
784 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-2x4.c",
785 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-3x4.c",
786 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-4x4.c",
787 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-5x4.c",
788 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-6x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800789 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
790 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
791 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
792 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4.c",
793 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
794 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4.c",
795 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-3x4.c",
796 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-4x4.c",
797 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-5x4.c",
798 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-6x4.c",
Frank Barchard02bb4292020-12-15 18:25:32 -0800799 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc2.c",
800 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc3.c",
801 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc4.c",
802 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4.c",
803 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-2x4-acc2.c",
804 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-2x4.c",
805 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-3x4.c",
806 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-4x4.c",
807 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-5x4.c",
808 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-6x4.c",
Frank Barchardc5704bf2020-12-21 23:09:00 -0800809 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
810 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
811 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
812 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4.c",
813 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
814 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4.c",
815 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-3x4.c",
816 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-4x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800817 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc2.c",
818 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc3.c",
819 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc4.c",
820 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4.c",
821 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4-acc2.c",
822 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4.c",
823 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-3x4.c",
824 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-4x4.c",
Frank Barchardcadd4222021-01-20 16:27:25 -0800825 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
826 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
827 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
828 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4.c",
829 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
830 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4.c",
831 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-3x4.c",
832 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-4x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800833 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc2.c",
834 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc3.c",
835 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc4.c",
836 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4.c",
837 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4-acc2.c",
838 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4.c",
839 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-3x4.c",
840 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-4x4.c",
Frank Barchardb20dcd62020-12-15 16:46:14 -0800841 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
842 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
843 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
844 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c",
845 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4.c",
846 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
847 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c",
848 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4.c",
849 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c",
850 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-3x4.c",
851 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-4x4-acc2.c",
852 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-4x4.c",
853 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800854 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
855 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
856 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
857 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
858 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4.c",
859 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
860 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
861 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4.c",
862 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
863 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4.c",
864 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4-acc2.c",
865 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4.c",
866 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-5x4.c",
Frank Barchardb20dcd62020-12-15 16:46:14 -0800867 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
868 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
869 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
870 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c",
871 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4.c",
872 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
873 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c",
874 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4.c",
875 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c",
876 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-3x4.c",
877 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-4x4-acc2.c",
878 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-4x4.c",
879 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800880 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
881 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
882 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
883 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
884 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4.c",
885 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
886 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
887 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4.c",
888 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
889 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4.c",
890 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4-acc2.c",
891 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4.c",
892 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-5x4.c",
Frank Barchardc6889b32020-12-21 11:27:22 -0800893 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
894 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
895 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
896 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c",
897 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c",
898 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
899 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c",
900 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4.c",
901 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c",
902 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800903 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
904 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
905 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
906 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
907 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4.c",
908 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
909 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
910 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4.c",
911 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
912 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4.c",
Frank Barchardc6889b32020-12-21 11:27:22 -0800913 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
914 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
915 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
916 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c",
917 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c",
918 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
919 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c",
920 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4.c",
921 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c",
922 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4.c",
Frank Barchard412e2f42020-12-11 11:40:50 -0800923 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
924 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
925 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
926 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
927 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4.c",
928 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
929 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
930 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4.c",
931 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
932 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700933 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
934 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700935 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
936 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
937 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
938 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800939 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-loadsplat.c",
940 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-splat.c",
941 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-loadsplat.c",
942 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700943 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
944 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800945 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-loadsplat.c",
946 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-splat.c",
947 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-loadsplat.c",
948 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700949 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
950 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800951 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-loadsplat.c",
952 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-splat.c",
953 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-loadsplat.c",
954 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700955 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
956 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800957 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-loadsplat.c",
958 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-splat.c",
959 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-loadsplat.c",
960 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700961 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
962 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800963 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-loadsplat.c",
964 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-splat.c",
965 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-loadsplat.c",
966 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700967 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
968 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800969 "src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
970 "src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
971 "src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
972 "src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700973 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
974 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
975 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
976 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800977 "src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
978 "src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
979 "src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
980 "src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700981 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
982 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
983 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
984 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
985 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
986 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800987 "src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
988 "src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
989 "src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
990 "src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700991 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
992 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
993 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
994 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -0800995 "src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
996 "src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
997 "src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
998 "src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700999 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
1000 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
1001 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
1002 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001003 "src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
1004 "src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
1005 "src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
1006 "src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001007 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
1008 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001009 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
1010 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
1011 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
XNNPACK Team965272b2020-10-23 21:10:15 -07001012 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
1013 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001014 "src/f32-ibilinear/gen/wasmsimd-c4.c",
1015 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001016 "src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
1017 "src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
1018 "src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
1019 "src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001020 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
1021 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
1022 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
1023 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001024 "src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
1025 "src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
1026 "src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
1027 "src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001028 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
1029 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
1030 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
1031 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
1032 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
1033 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001034 "src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
1035 "src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
1036 "src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
1037 "src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001038 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
1039 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
1040 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
1041 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001042 "src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
1043 "src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
1044 "src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
1045 "src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001046 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
1047 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
1048 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
1049 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001050 "src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
1051 "src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
1052 "src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
1053 "src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001054 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
1055 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -07001056 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
1057 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -07001058 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
1059 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
1060 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
1061 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Frank Barchard0725b8d2020-12-07 11:07:35 -08001062 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-arm-splat.c",
1063 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-x86-splat.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001064 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
1065 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
1066 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -07001067 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
1068 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001069 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
1070 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
1071 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
1072 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
1073 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
1074 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
1075 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -07001076 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
1077 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001078 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
1079 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
1080 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
1081 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001082 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001083 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001084 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001085 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
1086 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001087 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001088 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
1089 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001090 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001091 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
1092 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001093 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001094 "src/f32-relu/gen/wasmsimd-x4.c",
1095 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -07001096 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -07001097 "src/f32-rmax/wasmsimd-arm.c",
1098 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -07001099 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
1100 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
1101 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
1102 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
1103 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
1104 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001105 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
1106 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
1107 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
1108 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
1109 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
1110 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001111 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined-x2.c",
1112 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001113 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c",
1114 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001115 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001116 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined-x2.c",
1117 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001118 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c",
1119 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001120 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001121 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined-x2.c",
1122 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001123 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c",
1124 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001125 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001126 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined-x2.c",
1127 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001128 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c",
1129 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001130 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001131 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined-x2.c",
1132 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001133 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c",
1134 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001135 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001136 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined-x2.c",
1137 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001138 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c",
1139 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001140 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001141 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined-x2.c",
1142 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001143 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c",
1144 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001145 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001146 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined-x2.c",
1147 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001148 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c",
1149 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001150 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001151 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
1152 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001153 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001154 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
1155 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001156 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001157 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
1158 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001159 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001160 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
1161 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001162 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001163 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
1164 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001165 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001166 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
1167 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001168 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001169 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
1170 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001171 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001172 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
1173 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001174 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001175 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
1176 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001177 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001178 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
1179 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001180 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001181 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
1182 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001183 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001184 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
1185 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001186 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001187 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
1188 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001189 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001190 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
1191 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001192 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001193 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
1194 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001195 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001196 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
1197 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001198 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001199 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
1200 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001201 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001202 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
1203 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001204 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001205 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
1206 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001207 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001208 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
1209 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001210 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001211 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
1212 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001213 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001214 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
1215 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001216 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001217 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
1218 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001219 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001220 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
1221 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001222 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001223 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
1224 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001225 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001226 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
1227 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001228 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001229 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
1230 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001231 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001232 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
1233 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001234 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001235 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
1236 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001237 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001238 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
1239 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001240 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001241 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
1242 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001243 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001244 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
1245 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001246 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001247 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
1248 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001249 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001250 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
1251 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001252 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001253 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
1254 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001255 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001256 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
1257 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001258 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001259 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1260 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001261 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001262 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1263 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001264 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001265 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1266 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001267 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001268 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1269 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001270 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001271 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1272 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001273 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001274 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1275 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001276 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001277 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1278 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001279 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001280 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1281 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001282 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001283 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1284 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001285 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001286 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1287 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001288 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001289 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1290 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001291 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001292 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1293 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001294 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001295 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1296 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001297 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001298 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1299 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001300 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001301 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x4.c",
1302 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x8.c",
1303 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x12.c",
1304 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x16.c",
1305 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x20.c",
1306 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001307 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x4.c",
1308 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x8.c",
1309 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x12.c",
1310 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x16.c",
1311 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x20.c",
1312 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x24.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08001313 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x4.c",
1314 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x8.c",
1315 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x12.c",
1316 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x16.c",
1317 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x20.c",
1318 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001319 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x4.c",
1320 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x8.c",
1321 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x12.c",
1322 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x16.c",
1323 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x20.c",
1324 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001325 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1326 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1327 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1328 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001329 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001330 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001331 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001332 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001333 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1334 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1335 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1336 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001337 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1338 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1339 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1340 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1341 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1342 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1343 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1344 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1345 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1346 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001347 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1348 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001349 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1350 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1351 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1352 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1353 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1354 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001355 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-andnot.c",
1356 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-max.c",
1357 "src/math/expm1minus-wasmsimd-rr2-p6-andnot.c",
1358 "src/math/expm1minus-wasmsimd-rr2-p6-max.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001359 "src/math/roundd-wasmsimd-addsub.c",
1360 "src/math/roundd-wasmsimd-cvt.c",
1361 "src/math/roundne-wasmsimd-addsub.c",
1362 "src/math/roundu-wasmsimd-addsub.c",
1363 "src/math/roundu-wasmsimd-cvt.c",
1364 "src/math/roundz-wasmsimd-addsub.c",
1365 "src/math/roundz-wasmsimd-cvt.c",
1366 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1367 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001368 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1369 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1370 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001371 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1372 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1373 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001374 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1375 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1376 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001377 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001378 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001379 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001380 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1381 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001382 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001383 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1384 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001385 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1386 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001387 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001388 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001389 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001390 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001391 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001392 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001393 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001394 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1395 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1396 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1397 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1398 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1399 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1400 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1401 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001402 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001403 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001404 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001405 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001406 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001407 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001408 "src/x32-zip/x2-wasmsimd.c",
1409 "src/x32-zip/x3-wasmsimd.c",
1410 "src/x32-zip/x4-wasmsimd.c",
1411 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001412]
1413
Marat Dukhan08c4a432019-10-03 09:29:21 -07001414# ISA-specific micro-kernels
1415NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001416 "src/f32-argmaxpool/4x-neon-c4.c",
1417 "src/f32-argmaxpool/9p8x-neon-c4.c",
1418 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001419 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1420 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001421 "src/f32-clamp/gen/neon-x4.c",
1422 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001423 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001424 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001425 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001426 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001427 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001428 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001429 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001430 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanc7634882020-12-07 15:11:12 -08001431 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001432 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001433 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001434 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001435 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001436 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001437 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1438 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1439 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1440 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1441 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001442 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001443 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001444 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1445 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1446 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001447 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001448 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001449 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1450 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1451 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1452 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1453 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001454 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1455 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1456 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001457 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001458 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001459 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1460 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1461 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001462 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1463 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1464 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1465 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001466 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001467 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1468 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001469 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001470 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001471 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001472 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001473 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1474 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001475 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1476 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1477 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1478 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1479 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1480 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1481 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1482 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001483 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001484 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001485 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001486 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1487 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001488 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001489 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1490 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001491 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001492 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1493 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1494 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1495 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1496 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001497 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1498 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001499 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1500 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001501 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1502 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001503 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1504 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1505 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1506 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1507 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1508 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1509 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1510 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1511 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1512 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1513 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1514 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1515 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1516 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1517 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1518 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001519 "src/f32-hswish/gen/hswish-neon-x4.c",
1520 "src/f32-hswish/gen/hswish-neon-x8.c",
1521 "src/f32-hswish/gen/hswish-neon-x16.c",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08001522 "src/f32-ibilinear-chw/gen/neon-p4.c",
1523 "src/f32-ibilinear-chw/gen/neon-p8.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001524 "src/f32-ibilinear/gen/neon-c4.c",
1525 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001526 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001527 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001528 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001529 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1530 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001531 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001532 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1533 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1534 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1535 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001536 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1537 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001538 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1539 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001540 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1541 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001542 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1543 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1544 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001545 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1546 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001547 "src/f32-prelu/gen/neon-1x4.c",
1548 "src/f32-prelu/gen/neon-1x8.c",
1549 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001550 "src/f32-prelu/gen/neon-2x4.c",
1551 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001552 "src/f32-prelu/gen/neon-2x16.c",
1553 "src/f32-prelu/gen/neon-4x4.c",
1554 "src/f32-prelu/gen/neon-4x8.c",
1555 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001556 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001557 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001558 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001559 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1560 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001561 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001562 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1563 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001564 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001565 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1566 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001567 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1568 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1569 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1570 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1571 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1572 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1573 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1574 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1575 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1576 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1577 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1578 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1579 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001580 "src/f32-relu/gen/neon-x4.c",
1581 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001582 "src/f32-rmax/neon.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001583 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1584 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1585 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1586 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1587 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1588 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1589 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1590 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1591 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1592 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1593 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1594 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001595 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1596 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1597 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1598 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1599 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1600 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan5b86c432020-12-06 19:15:03 -08001601 "src/f32-spmm/gen/4x1-minmax-neon-pipelined.c",
1602 "src/f32-spmm/gen/4x1-minmax-neon-x2.c",
1603 "src/f32-spmm/gen/4x1-minmax-neon.c",
1604 "src/f32-spmm/gen/8x1-minmax-neon-pipelined.c",
1605 "src/f32-spmm/gen/8x1-minmax-neon-x2.c",
1606 "src/f32-spmm/gen/8x1-minmax-neon.c",
1607 "src/f32-spmm/gen/12x1-minmax-neon.c",
1608 "src/f32-spmm/gen/16x1-minmax-neon-pipelined.c",
1609 "src/f32-spmm/gen/16x1-minmax-neon-x2.c",
1610 "src/f32-spmm/gen/16x1-minmax-neon.c",
1611 "src/f32-spmm/gen/32x1-minmax-neon-pipelined.c",
1612 "src/f32-spmm/gen/32x1-minmax-neon-x2.c",
1613 "src/f32-spmm/gen/32x1-minmax-neon.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001614 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1615 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1616 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1617 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001618 "src/f32-vbinary/gen/vmax-neon-x4.c",
1619 "src/f32-vbinary/gen/vmax-neon-x8.c",
1620 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1621 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1622 "src/f32-vbinary/gen/vmin-neon-x4.c",
1623 "src/f32-vbinary/gen/vmin-neon-x8.c",
1624 "src/f32-vbinary/gen/vminc-neon-x4.c",
1625 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001626 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1627 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1628 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1629 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1630 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1631 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001632 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1633 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1634 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1635 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001636 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1637 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1638 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1639 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001640 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x4.c",
1641 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x8.c",
1642 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x12.c",
1643 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x16.c",
1644 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x20.c",
1645 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x24.c",
1646 "src/f32-velu/gen/velu-neon-rr2-p6-x4.c",
1647 "src/f32-velu/gen/velu-neon-rr2-p6-x8.c",
1648 "src/f32-velu/gen/velu-neon-rr2-p6-x12.c",
1649 "src/f32-velu/gen/velu-neon-rr2-p6-x16.c",
1650 "src/f32-velu/gen/velu-neon-rr2-p6-x20.c",
1651 "src/f32-velu/gen/velu-neon-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001652 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1653 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001654 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1655 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001656 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1657 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001658 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1659 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1660 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1661 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1662 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1663 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001664 "src/f32-vunary/gen/vabs-neon-x4.c",
1665 "src/f32-vunary/gen/vabs-neon-x8.c",
1666 "src/f32-vunary/gen/vneg-neon-x4.c",
1667 "src/f32-vunary/gen/vneg-neon-x8.c",
1668 "src/f32-vunary/gen/vsqr-neon-x4.c",
1669 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001670 "src/math/expm1minus-neon-rr2-lut16-p3.c",
1671 "src/math/expm1minus-neon-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001672 "src/math/roundd-neon-addsub.c",
1673 "src/math/roundd-neon-cvt.c",
1674 "src/math/roundne-neon-addsub.c",
1675 "src/math/roundu-neon-addsub.c",
1676 "src/math/roundu-neon-cvt.c",
1677 "src/math/roundz-neon-addsub.c",
1678 "src/math/roundz-neon-cvt.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001679 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1680 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1681 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1682 "src/math/sqrt-neon-nr1rsqrts.c",
1683 "src/math/sqrt-neon-nr2rsqrts.c",
1684 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001685 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1686 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1687 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1688 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001689 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1690 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1691 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1692 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001693 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1694 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1695 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1696 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001697 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001698 "src/qs8-gemm/gen/1x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001699 "src/qs8-gemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001700 "src/qs8-gemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001701 "src/qs8-gemm/gen/1x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001702 "src/qs8-gemm/gen/1x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001703 "src/qs8-gemm/gen/1x8c16-minmax-neon-mlal-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001704 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1705 "src/qs8-gemm/gen/1x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001706 "src/qs8-gemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001707 "src/qs8-gemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001708 "src/qs8-gemm/gen/1x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001709 "src/qs8-gemm/gen/1x16c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001710 "src/qs8-gemm/gen/1x16c16-minmax-neon-mlal-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001711 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001712 "src/qs8-gemm/gen/2x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001713 "src/qs8-gemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001714 "src/qs8-gemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001715 "src/qs8-gemm/gen/2x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001716 "src/qs8-gemm/gen/2x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001717 "src/qs8-gemm/gen/2x8c16-minmax-neon-mlal-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001718 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1719 "src/qs8-gemm/gen/2x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001720 "src/qs8-gemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001721 "src/qs8-gemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001722 "src/qs8-gemm/gen/2x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001723 "src/qs8-gemm/gen/2x16c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001724 "src/qs8-gemm/gen/2x16c16-minmax-neon-mlal-padal.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001725 "src/qs8-gemm/gen/3x8-minmax-neon-mlal-lane.c",
1726 "src/qs8-gemm/gen/3x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001727 "src/qs8-gemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001728 "src/qs8-gemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001729 "src/qs8-gemm/gen/3x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001730 "src/qs8-gemm/gen/3x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001731 "src/qs8-gemm/gen/3x8c16-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001732 "src/qs8-gemm/gen/3x16-minmax-neon-mlal-lane.c",
1733 "src/qs8-gemm/gen/3x16-minmax-neon-mull-addw-dup.c",
1734 "src/qs8-gemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
1735 "src/qs8-gemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001736 "src/qs8-gemm/gen/3x16c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001737 "src/qs8-gemm/gen/3x16c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001738 "src/qs8-gemm/gen/3x16c16-minmax-neon-mlal-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001739 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001740 "src/qs8-gemm/gen/4x8-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001741 "src/qs8-gemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001742 "src/qs8-gemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001743 "src/qs8-gemm/gen/4x8c8-minmax-neon-mlal-padal.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001744 "src/qs8-gemm/gen/4x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001745 "src/qs8-gemm/gen/4x8c16-minmax-neon-mlal-padal.c",
Frank Barchardcadd4222021-01-20 16:27:25 -08001746 "src/qs8-gemm/gen/4x16-minmax-neon-mlal-lane.c",
1747 "src/qs8-gemm/gen/4x16-minmax-neon-mull-addw-dup.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001748 "src/qs8-gemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
Frank Barchard2302ffd2021-01-22 14:08:02 -08001749 "src/qs8-gemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001750 "src/qs8-gemm/gen/4x16c8-minmax-neon-mlal-padal.c",
Frank Barchard58990122021-01-29 16:09:56 -08001751 "src/qs8-gemm/gen/4x16c8-minmax-neon-mull-padal.c",
Frank Barchard71c4d1a2021-02-19 10:07:23 -08001752 "src/qs8-gemm/gen/4x16c16-minmax-neon-mlal-padal.c",
Frank Barchardec0bf142021-01-15 10:14:05 -08001753 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001754 "src/qs8-igemm/gen/1x8-minmax-neon-mull-addw-dup.c",
1755 "src/qs8-igemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
1756 "src/qs8-igemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001757 "src/qs8-igemm/gen/1x8c8-minmax-neon-mlal-padal.c",
1758 "src/qs8-igemm/gen/1x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001759 "src/qs8-igemm/gen/1x8c16-minmax-neon-mlal-padal.c",
1760 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1761 "src/qs8-igemm/gen/1x16-minmax-neon-mull-addw-dup.c",
1762 "src/qs8-igemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
1763 "src/qs8-igemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
1764 "src/qs8-igemm/gen/1x16c8-minmax-neon-mlal-padal.c",
1765 "src/qs8-igemm/gen/1x16c8-minmax-neon-mull-padal.c",
1766 "src/qs8-igemm/gen/1x16c16-minmax-neon-mlal-padal.c",
1767 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
1768 "src/qs8-igemm/gen/2x8-minmax-neon-mull-addw-dup.c",
1769 "src/qs8-igemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
1770 "src/qs8-igemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001771 "src/qs8-igemm/gen/2x8c8-minmax-neon-mlal-padal.c",
1772 "src/qs8-igemm/gen/2x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001773 "src/qs8-igemm/gen/2x8c16-minmax-neon-mlal-padal.c",
1774 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
1775 "src/qs8-igemm/gen/2x16-minmax-neon-mull-addw-dup.c",
1776 "src/qs8-igemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
1777 "src/qs8-igemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
1778 "src/qs8-igemm/gen/2x16c8-minmax-neon-mlal-padal.c",
1779 "src/qs8-igemm/gen/2x16c8-minmax-neon-mull-padal.c",
1780 "src/qs8-igemm/gen/2x16c16-minmax-neon-mlal-padal.c",
1781 "src/qs8-igemm/gen/3x8-minmax-neon-mlal-lane.c",
1782 "src/qs8-igemm/gen/3x8-minmax-neon-mull-addw-dup.c",
1783 "src/qs8-igemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
1784 "src/qs8-igemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001785 "src/qs8-igemm/gen/3x8c8-minmax-neon-mlal-padal.c",
1786 "src/qs8-igemm/gen/3x8c8-minmax-neon-mull-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001787 "src/qs8-igemm/gen/3x8c16-minmax-neon-mlal-padal.c",
1788 "src/qs8-igemm/gen/3x16-minmax-neon-mlal-lane.c",
1789 "src/qs8-igemm/gen/3x16-minmax-neon-mull-addw-dup.c",
1790 "src/qs8-igemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
1791 "src/qs8-igemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
1792 "src/qs8-igemm/gen/3x16c8-minmax-neon-mlal-padal.c",
1793 "src/qs8-igemm/gen/3x16c8-minmax-neon-mull-padal.c",
1794 "src/qs8-igemm/gen/3x16c16-minmax-neon-mlal-padal.c",
1795 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1796 "src/qs8-igemm/gen/4x8-minmax-neon-mull-addw-dup.c",
1797 "src/qs8-igemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
1798 "src/qs8-igemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001799 "src/qs8-igemm/gen/4x8c8-minmax-neon-mlal-padal.c",
1800 "src/qs8-igemm/gen/4x8c8-minmax-neon-mull-padal.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001801 "src/qs8-igemm/gen/4x8c16-minmax-neon-mlal-padal.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001802 "src/qs8-igemm/gen/4x16-minmax-neon-mlal-lane.c",
Frank Barchardda78da12021-03-02 14:28:00 -08001803 "src/qs8-igemm/gen/4x16-minmax-neon-mull-addw-dup.c",
Frank Barchard671d1b02021-03-10 15:42:08 -08001804 "src/qs8-igemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
1805 "src/qs8-igemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
1806 "src/qs8-igemm/gen/4x16c8-minmax-neon-mlal-padal.c",
1807 "src/qs8-igemm/gen/4x16c8-minmax-neon-mull-padal.c",
1808 "src/qs8-igemm/gen/4x16c16-minmax-neon-mlal-padal.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001809 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001810 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001811 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001812 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1813 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1814 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1815 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1816 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1817 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1818 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1819 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001820 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1821 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1822 "src/qu8-dwconv/up8x9-minmax-neon.c",
1823 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1824 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1825 "src/qu8-gemm/4x8-minmax-neon.c",
1826 "src/qu8-gemm/8x8-minmax-neon.c",
1827 "src/qu8-igemm/4x8-minmax-neon.c",
1828 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001829 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001830 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001831 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001832 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001833 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001834 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001835 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001836 "src/x8-zip/x2-neon.c",
1837 "src/x8-zip/x3-neon.c",
1838 "src/x8-zip/x4-neon.c",
1839 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001840 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001841 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001842 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001843 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001844 "src/x32-zip/x2-neon.c",
1845 "src/x32-zip/x3-neon.c",
1846 "src/x32-zip/x4-neon.c",
1847 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001848]
1849
1850NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001851 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1852 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1853 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1854 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1855 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1856 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1857 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1858 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1859 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1860 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1861 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1862 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1863 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1864 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1865 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1866 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1867 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1868 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1869 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1870 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1871 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1872 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1873 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1874 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1875 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1876 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1877 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1878 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1879 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1880 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08001881 "src/f32-ibilinear-chw/gen/neonfma-p4.c",
1882 "src/f32-ibilinear-chw/gen/neonfma-p8.c",
Frank Barchard8247e212021-02-03 18:12:33 -08001883 "src/f32-ibilinear/gen/neonfma-c4.c",
1884 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001885 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001886 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001887 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001888 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1889 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001890 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1891 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001892 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1893 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001894 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1895 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001896 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001897 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001898 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001899 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1900 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001901 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001902 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1903 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001904 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001905 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1906 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001907 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1908 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1909 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1910 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1911 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1912 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1913 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1914 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1915 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1916 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1917 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1918 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1919 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001920 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1921 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1922 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1923 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1924 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1925 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001926 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1927 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1928 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1929 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1930 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1931 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001932 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1933 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1934 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1935 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1936 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1937 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001938 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1939 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1940 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1941 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1942 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1943 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001944 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1945 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1946 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1947 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1948 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1949 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001950 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1951 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1952 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1953 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1954 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1955 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001956 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1957 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1958 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1959 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1960 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1961 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1962 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1963 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1964 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1965 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1966 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1967 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1968 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1969 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1970 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1971 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1972 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1973 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan2fa7a0c2020-12-06 19:09:02 -08001974 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1975 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
1976 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1977 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1978 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
1979 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1980 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1981 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1982 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
1983 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1984 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
1985 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
1986 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001987 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x4.c",
1988 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x8.c",
1989 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x12.c",
1990 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x16.c",
1991 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x20.c",
1992 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x24.c",
1993 "src/f32-velu/gen/velu-neonfma-rr1-p6-x4.c",
1994 "src/f32-velu/gen/velu-neonfma-rr1-p6-x8.c",
1995 "src/f32-velu/gen/velu-neonfma-rr1-p6-x12.c",
1996 "src/f32-velu/gen/velu-neonfma-rr1-p6-x16.c",
1997 "src/f32-velu/gen/velu-neonfma-rr1-p6-x20.c",
1998 "src/f32-velu/gen/velu-neonfma-rr1-p6-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001999 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
2000 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002001 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
2002 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
2003 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
2004 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
2005 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
2006 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
2007 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
2008 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
2009 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
2010 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
2011 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
2012 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
2013 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
2014 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
2015 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
2016 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
2017 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
2018 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
2019 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
2020 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002021 "src/math/exp-neonfma-rr2-lut64-p2.c",
2022 "src/math/exp-neonfma-rr2-p5.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08002023 "src/math/expm1minus-neonfma-rr1-lut16-p3.c",
2024 "src/math/expm1minus-neonfma-rr1-p6.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08002025 "src/math/expminus-neonfma-rr2-lut64-p2.c",
2026 "src/math/expminus-neonfma-rr2-lut2048-p1.c",
2027 "src/math/expminus-neonfma-rr2-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002028 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
2029 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
2030 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002031 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
2032 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
2033 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002034 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
2035 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
2036 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002037 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
2038 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
2039 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002040 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
2041 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
2042 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002043 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
2044 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
2045 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002046 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002047 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002048 "src/math/sqrt-neonfma-nr2fma.c",
2049 "src/math/sqrt-neonfma-nr2fma1adj.c",
2050 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002051]
2052
2053AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07002054 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002055 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002056 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002057 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07002058 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002059 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002060 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07002061 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002062 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002063 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
2064 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
2065 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07002066 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002067 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07002068 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
2069 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
2070 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
2071 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
2072 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07002073 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
2074 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
2075 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002076 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07002077 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002078 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
2079 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
2080 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002081 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
2082 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
2083 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
2084 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002085 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002086 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
2087 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002088 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002089 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002090 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07002091 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002092 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
2093 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07002094 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
2095 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
2096 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
2097 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
2098 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
2099 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
2100 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
2101 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07002102 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002103 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002104 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
2105 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
2106 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
2107 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
2108 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
2109 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
2110 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
2111 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
2112 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
2113 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
2114 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
2115 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
2116 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
2117 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
2118 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
2119 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
2120 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
2121 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
2122 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
2123 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08002124 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
2125 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
2126 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
2127 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
2128 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
2129 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
2130 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
2131 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
2132 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
2133 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
2134 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
2135 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002136 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
2137 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
2138 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
2139 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
2140 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
2141 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002142 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
2143 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002144 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
2145 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002146 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
2147 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002148 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
2149 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002150 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
2151 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002152 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
2153 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
2154 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
2155 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
2156 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
2157 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002158 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
2159 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002160 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002161 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002162 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002163 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002164 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08002165 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002166]
2167
Marat Dukhan8853b822020-05-07 12:19:01 -07002168NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002169 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
2170 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002171 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
2172 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
2173 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
2174 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
2175 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
2176 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002177 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002178 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002179 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002180 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002181]
2182
Marat Dukhan08c4a432019-10-03 09:29:21 -07002183AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07002184 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002185 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07002186 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
2187 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
2188 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
2189 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002190 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
2191 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
2192 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
2193 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
2194 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
2195 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
2196 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
2197 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07002198 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
2199 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002200 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
2201 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
2202 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
2203 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
2204 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
2205 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
2206 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
2207 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
2208 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2209 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2210 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2211 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2212 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2213 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2214 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2215 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002216 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002217 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
2218 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2219 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2220 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2221 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2222 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2223 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2224 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2225 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07002226 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07002227 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002228 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002229 "src/f16-relu/gen/neonfp16arith-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002230 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002231 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002232 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002233 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002234 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002235 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002236 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002237 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
2238 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
2239 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
2240 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
2241 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
2242 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
2243 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
2244 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
2245 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
2246 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
2247 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
2248 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
2249 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
2250 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
2251 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
2252 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
2253 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
2254 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
2255 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
2256 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
2257 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
2258 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
2259 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
2260 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
2261 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
2262 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
2263 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
2264 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
2265 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
2266 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
2267 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002268]
2269
Benoit Jacoba9644732020-08-13 12:48:55 -07002270NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07002271 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002272 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002273 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002274 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002275 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002276 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002277 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002278 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002279 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002280 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002281 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07002282 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002283 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002284 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002285 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002286 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002287]
2288
Marat Dukhan08c4a432019-10-03 09:29:21 -07002289SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07002290 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
2291 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002292 "src/f32-clamp/gen/sse-x4.c",
2293 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07002294 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
2295 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002296 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
2297 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
2298 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
2299 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002300 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
2301 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002302 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
2303 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
2304 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
2305 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002306 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
2307 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002308 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
2309 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
2310 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002311 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002312 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002313 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
2314 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
2315 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
2316 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
2317 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002318 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
2319 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
2320 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002321 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002322 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002323 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
2324 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
2325 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07002326 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
2327 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
2328 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
2329 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
2330 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
2331 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
2332 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
2333 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
2334 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
2335 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
2336 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
2337 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
2338 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002339 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
2340 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
2341 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
2342 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
2343 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
2344 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
2345 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
2346 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002347 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08002348 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002349 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002350 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
2351 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002352 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
2353 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
2354 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002355 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-dup.c",
2356 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-load1.c",
2357 "src/f32-gemm/gen-inc/3x8s4inc-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002358 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
2359 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
2360 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002361 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-dup.c",
2362 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-load1.c",
2363 "src/f32-gemm/gen-inc/5x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002364 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
2365 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
2366 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002367 "src/f32-gemm/gen/3x8-minmax-sse-dup.c",
2368 "src/f32-gemm/gen/3x8-minmax-sse-load1.c",
2369 "src/f32-gemm/gen/3x8s4-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002370 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
2371 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
2372 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
2373 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002374 "src/f32-gemm/gen/5x8-minmax-sse-dup.c",
2375 "src/f32-gemm/gen/5x8-minmax-sse-load1.c",
2376 "src/f32-gemm/gen/5x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002377 "src/f32-hswish/gen/hswish-sse-x4.c",
2378 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07002379 "src/f32-ibilinear/gen/sse-c4.c",
2380 "src/f32-ibilinear/gen/sse-c8.c",
Artsiom Ablavatskib3ffd582021-03-31 13:00:08 -07002381 "src/f32-ibilinear-chw/gen/sse-p4.c",
2382 "src/f32-ibilinear-chw/gen/sse-p8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002383 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
2384 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
2385 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002386 "src/f32-igemm/gen/3x8-minmax-sse-dup.c",
2387 "src/f32-igemm/gen/3x8-minmax-sse-load1.c",
2388 "src/f32-igemm/gen/3x8s4-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002389 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2390 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2391 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2392 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002393 "src/f32-igemm/gen/5x8-minmax-sse-dup.c",
2394 "src/f32-igemm/gen/5x8-minmax-sse-load1.c",
2395 "src/f32-igemm/gen/5x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002396 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2397 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2398 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002399 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002400 "src/f32-prelu/gen/sse-2x4.c",
2401 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002402 "src/f32-relu/gen/sse-x4.c",
2403 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002404 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002405 "src/f32-spmm/gen/4x1-minmax-sse.c",
2406 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002407 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002408 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002409 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2410 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2411 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2412 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2413 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2414 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2415 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2416 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002417 "src/f32-vbinary/gen/vmax-sse-x4.c",
2418 "src/f32-vbinary/gen/vmax-sse-x8.c",
2419 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2420 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2421 "src/f32-vbinary/gen/vmin-sse-x4.c",
2422 "src/f32-vbinary/gen/vmin-sse-x8.c",
2423 "src/f32-vbinary/gen/vminc-sse-x4.c",
2424 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002425 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2426 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2427 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2428 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2429 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2430 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2431 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2432 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002433 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2434 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2435 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2436 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002437 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2438 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2439 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2440 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002441 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2442 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002443 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2444 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002445 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2446 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002447 "src/f32-vunary/gen/vabs-sse-x4.c",
2448 "src/f32-vunary/gen/vabs-sse-x8.c",
2449 "src/f32-vunary/gen/vneg-sse-x4.c",
2450 "src/f32-vunary/gen/vneg-sse-x8.c",
2451 "src/f32-vunary/gen/vsqr-sse-x4.c",
2452 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002453 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002454 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002455 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002456 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002457 "src/math/sqrt-sse-hh1mac.c",
2458 "src/math/sqrt-sse-nr1mac.c",
2459 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002460 "src/x32-fill/sse.c",
2461 "src/x32-packx/x4-sse.c",
2462 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002463]
2464
2465SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002466 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002467 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002468 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan802fcae2020-12-11 14:37:25 -08002469 "src/f32-gemm/gen-inc/1x8inc-minmax-sse2-dup.c",
2470 "src/f32-gemm/gen-inc/3x8inc-minmax-sse2-dup.c",
2471 "src/f32-gemm/gen-inc/4x8inc-minmax-sse2-dup.c",
2472 "src/f32-gemm/gen-inc/5x8inc-minmax-sse2-dup.c",
2473 "src/f32-gemm/gen/1x8-minmax-sse2-dup.c",
2474 "src/f32-gemm/gen/3x8-minmax-sse2-dup.c",
2475 "src/f32-gemm/gen/4x8-minmax-sse2-dup.c",
2476 "src/f32-gemm/gen/5x8-minmax-sse2-dup.c",
2477 "src/f32-igemm/gen/1x8-minmax-sse2-dup.c",
2478 "src/f32-igemm/gen/3x8-minmax-sse2-dup.c",
2479 "src/f32-igemm/gen/4x8-minmax-sse2-dup.c",
2480 "src/f32-igemm/gen/5x8-minmax-sse2-dup.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002481 "src/f32-prelu/gen/sse2-2x4.c",
2482 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002483 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002484 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002485 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002486 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2487 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002488 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002489 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2490 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002491 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002492 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2493 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002494 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002495 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
2496 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
2497 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
2498 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
2499 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
2500 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002501 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002502 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002503 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002504 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002505 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
2506 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002507 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x4.c",
2508 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x8.c",
2509 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x12.c",
2510 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x16.c",
2511 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x20.c",
2512 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x24.c",
2513 "src/f32-velu/gen/velu-sse2-rr2-p6-x4.c",
2514 "src/f32-velu/gen/velu-sse2-rr2-p6-x8.c",
2515 "src/f32-velu/gen/velu-sse2-rr2-p6-x12.c",
2516 "src/f32-velu/gen/velu-sse2-rr2-p6-x16.c",
2517 "src/f32-velu/gen/velu-sse2-rr2-p6-x20.c",
2518 "src/f32-velu/gen/velu-sse2-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002519 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2520 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002521 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2522 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002523 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2524 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2525 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2526 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2527 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2528 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002529 "src/math/exp-sse2-rr2-lut64-p2.c",
2530 "src/math/exp-sse2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002531 "src/math/expm1minus-sse2-rr2-lut16-p3.c",
Marat Dukhana438aca2020-11-20 15:45:01 -08002532 "src/math/expm1minus-sse2-rr2-p6.c",
Frank Barchard3b800452020-11-22 12:12:35 -08002533 "src/math/expminus-sse2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002534 "src/math/roundd-sse2-cvt.c",
2535 "src/math/roundne-sse2-cvt.c",
2536 "src/math/roundu-sse2-cvt.c",
2537 "src/math/roundz-sse2-cvt.c",
2538 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2539 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2540 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2541 "src/math/sigmoid-sse2-rr2-p5-div.c",
2542 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2543 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002544 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2545 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2546 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002547 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2548 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2549 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002550 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2551 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2552 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002553 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002554 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002555 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002556 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002557 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002558 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002559 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2560 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002561 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002562 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2563 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002564 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002565 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2566 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2567 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002568 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002569 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002570 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002571 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002572 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002573 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002574 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002575 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002576 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2577 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002578 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002579 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002580 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002581 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2582 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2583 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2584 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002585 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2586 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2587 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2588 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002589 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2590 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002591 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2592 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2593 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2594 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2595 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002596 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002597 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002598 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002599 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002600 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002601 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002602 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002603 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002604 "src/x8-zip/x2-sse2.c",
2605 "src/x8-zip/x3-sse2.c",
2606 "src/x8-zip/x4-sse2.c",
2607 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002608 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002609 "src/x32-zip/x2-sse2.c",
2610 "src/x32-zip/x3-sse2.c",
2611 "src/x32-zip/x4-sse2.c",
2612 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002613]
2614
2615SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002616 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2617 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2618 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002619 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002620 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002621 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2622 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2623 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2624 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2625 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002626 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2627 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2628 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002629 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2630 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2631 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002632 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2633 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2634 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002635 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002636 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002637 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002638 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002639 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002640 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002641 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2642 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002643 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002644 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2645 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002646 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002647 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2648 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2649 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002650 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002651 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002652 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002653 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002654 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002655 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002656 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002657 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002658 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2659 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002660 "src/qs8-requantization/precise-ssse3.c",
2661 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002662 "src/qu8-requantization/precise-ssse3.c",
2663 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002664]
2665
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002666SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002667 "src/f32-prelu/gen/sse41-2x4.c",
2668 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002669 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2670 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2671 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2672 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2673 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2674 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002675 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002676 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002677 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002678 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002679 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2680 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002681 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x4.c",
2682 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x8.c",
2683 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x12.c",
2684 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x16.c",
2685 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x20.c",
2686 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x24.c",
2687 "src/f32-velu/gen/velu-sse41-rr2-p6-x4.c",
2688 "src/f32-velu/gen/velu-sse41-rr2-p6-x8.c",
2689 "src/f32-velu/gen/velu-sse41-rr2-p6-x12.c",
2690 "src/f32-velu/gen/velu-sse41-rr2-p6-x16.c",
2691 "src/f32-velu/gen/velu-sse41-rr2-p6-x20.c",
2692 "src/f32-velu/gen/velu-sse41-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002693 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2694 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002695 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2696 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002697 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2698 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2699 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2700 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2701 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2702 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2703 "src/math/roundd-sse41.c",
2704 "src/math/roundne-sse41.c",
2705 "src/math/roundu-sse41.c",
2706 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002707 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2708 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2709 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07002710 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul32.c",
2711 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul32.c",
2712 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul32.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002713 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2714 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2715 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002716 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2717 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2718 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002719 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002720 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002721 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002722 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002723 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002724 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002725 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2726 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002727 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002728 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2729 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002730 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002731 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2732 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2733 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002734 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002735 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002736 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002737 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002738 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002739 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002740 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002741 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002742 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2743 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002744 "src/qs8-requantization/fp32-sse4.c",
2745 "src/qs8-requantization/precise-sse4.c",
2746 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002747 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2748 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2749 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2750 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002751 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2752 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2753 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2754 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002755 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2756 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2757 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2758 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002759 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2760 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2761 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2762 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002763 "src/qu8-requantization/precise-sse4.c",
2764 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002765]
2766
Marat Dukhan08c4a432019-10-03 09:29:21 -07002767AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002768 "src/f32-clamp/gen/avx-x8.c",
2769 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002770 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2771 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002772 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2773 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002774 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2775 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002776 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2777 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2778 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2779 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2780 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2781 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002782 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002783 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2784 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002785 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002786 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002787 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002788 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002789 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2790 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2791 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2792 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2793 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2794 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2795 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2796 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2797 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2798 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2799 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002800 "src/f32-hswish/gen/hswish-avx-x8.c",
2801 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002802 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002803 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2804 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002805 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002806 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002807 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002808 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002809 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2810 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002811 "src/f32-prelu/gen/avx-2x8.c",
2812 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002813 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002814 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002815 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002816 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2817 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2818 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2819 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2820 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2821 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2822 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2823 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2824 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2825 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2826 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2827 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2828 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2829 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2830 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2831 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2832 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2833 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2834 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2835 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002836 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2837 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2838 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2839 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2840 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2841 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2842 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2843 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002844 "src/f32-vbinary/gen/vmax-avx-x8.c",
2845 "src/f32-vbinary/gen/vmax-avx-x16.c",
2846 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2847 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2848 "src/f32-vbinary/gen/vmin-avx-x8.c",
2849 "src/f32-vbinary/gen/vmin-avx-x16.c",
2850 "src/f32-vbinary/gen/vminc-avx-x8.c",
2851 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002852 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2853 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2854 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2855 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2856 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2857 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2858 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2859 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002860 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2861 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2862 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2863 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002864 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2865 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2866 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2867 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002868 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x8.c",
2869 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x16.c",
2870 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x24.c",
2871 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x32.c",
2872 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x40.c",
2873 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x48.c",
2874 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x8.c",
2875 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x16.c",
2876 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x24.c",
2877 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x32.c",
2878 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x40.c",
2879 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x48.c",
2880 "src/f32-velu/gen/velu-avx-rr2-p6-x8.c",
2881 "src/f32-velu/gen/velu-avx-rr2-p6-x16.c",
2882 "src/f32-velu/gen/velu-avx-rr2-p6-x24.c",
2883 "src/f32-velu/gen/velu-avx-rr2-p6-x32.c",
2884 "src/f32-velu/gen/velu-avx-rr2-p6-x40.c",
2885 "src/f32-velu/gen/velu-avx-rr2-p6-x48.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002886 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2887 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002888 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2889 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002890 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2891 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2892 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2893 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2894 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2895 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002896 "src/f32-vscale/avx-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002897 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2898 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002899 "src/f32-vunary/gen/vabs-avx-x8.c",
2900 "src/f32-vunary/gen/vabs-avx-x16.c",
2901 "src/f32-vunary/gen/vneg-avx-x8.c",
2902 "src/f32-vunary/gen/vneg-avx-x16.c",
2903 "src/f32-vunary/gen/vsqr-avx-x8.c",
2904 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhanfa0ab852021-04-02 17:30:49 -07002905 "src/qs8-dwconv/gen/up8x9-minmax-avx-mul16.c",
2906 "src/qs8-dwconv/gen/up16x9-minmax-avx-mul16.c",
2907 "src/qs8-dwconv/gen/up24x9-minmax-avx-mul16.c",
Marat Dukhan07feec82021-04-02 22:41:15 -07002908 "src/qs8-dwconv/gen/up8x9-minmax-avx-mul32.c",
2909 "src/qs8-dwconv/gen/up16x9-minmax-avx-mul32.c",
2910 "src/qs8-dwconv/gen/up24x9-minmax-avx-mul32.c",
Marat Dukhana3c16332021-04-02 15:03:27 -07002911 "src/qs8-gemm/gen/1x4c2-minmax-avx-ld64.c",
2912 "src/qs8-gemm/gen/1x4c2-minmax-avx-ld128.c",
2913 "src/qs8-gemm/gen/1x4c2-xw-minmax-avx.c",
2914 "src/qs8-gemm/gen/1x4c8-minmax-avx-ld64.c",
2915 "src/qs8-gemm/gen/1x4c8-minmax-avx-ld128.c",
2916 "src/qs8-gemm/gen/1x4c8-xw-minmax-avx.c",
2917 "src/qs8-gemm/gen/2x4c8-minmax-avx-ld64.c",
2918 "src/qs8-gemm/gen/2x4c8-minmax-avx-ld128.c",
2919 "src/qs8-gemm/gen/2x4c8-xw-minmax-avx.c",
2920 "src/qs8-gemm/gen/3x4c8-minmax-avx-ld64.c",
2921 "src/qs8-gemm/gen/3x4c8-minmax-avx-ld128.c",
2922 "src/qs8-gemm/gen/3x4c8-xw-minmax-avx.c",
2923 "src/qs8-gemm/gen/4x4c2-minmax-avx-ld64.c",
2924 "src/qs8-gemm/gen/4x4c2-minmax-avx-ld128.c",
2925 "src/qs8-gemm/gen/4x4c2-xw-minmax-avx.c",
2926 "src/qs8-igemm/gen/1x4c2-minmax-avx-ld64.c",
2927 "src/qs8-igemm/gen/1x4c2-minmax-avx-ld128.c",
2928 "src/qs8-igemm/gen/1x4c8-minmax-avx-ld64.c",
2929 "src/qs8-igemm/gen/1x4c8-minmax-avx-ld128.c",
2930 "src/qs8-igemm/gen/2x4c8-minmax-avx-ld64.c",
2931 "src/qs8-igemm/gen/2x4c8-minmax-avx-ld128.c",
2932 "src/qs8-igemm/gen/3x4c8-minmax-avx-ld64.c",
2933 "src/qs8-igemm/gen/3x4c8-minmax-avx-ld128.c",
2934 "src/qs8-igemm/gen/4x4c2-minmax-avx-ld64.c",
2935 "src/qs8-igemm/gen/4x4c2-minmax-avx-ld128.c",
Marat Dukhane9c4b962021-04-02 16:56:55 -07002936 "src/qs8-vadd/gen/minmax-avx-mul16-ld64-x8.c",
2937 "src/qs8-vadd/gen/minmax-avx-mul16-ld64-x16.c",
2938 "src/qs8-vadd/gen/minmax-avx-mul16-ld64-x24.c",
2939 "src/qs8-vadd/gen/minmax-avx-mul16-ld64-x32.c",
2940 "src/qs8-vadd/gen/minmax-avx-mul32-ld32-x8.c",
2941 "src/qs8-vadd/gen/minmax-avx-mul32-ld32-x16.c",
2942 "src/qs8-vadd/gen/minmax-avx-mul32-ld32-x24.c",
2943 "src/qs8-vadd/gen/minmax-avx-mul32-ld32-x32.c",
2944 "src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x8.c",
2945 "src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x16.c",
2946 "src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x24.c",
2947 "src/qs8-vaddc/gen/minmax-avx-mul16-ld64-x32.c",
2948 "src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x8.c",
2949 "src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x16.c",
2950 "src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x24.c",
2951 "src/qs8-vaddc/gen/minmax-avx-mul32-ld32-x32.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002952 "src/math/exp-avx-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002953 "src/math/expm1minus-avx-rr2-lut4-p4-perm.c",
2954 "src/math/expm1minus-avx-rr2-lut16-p3.c",
2955 "src/math/expm1minus-avx-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002956 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002957 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002958 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2959 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002960]
2961
Marat Dukhan1566fee2020-08-02 21:55:41 -07002962XOP_UKERNELS = [
2963 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002964 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002965 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002966 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002967 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002968 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002969 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2970 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002971 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002972 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2973 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002974 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002975 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2976 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2977 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002978 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002979 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002980 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002981 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002982 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002983 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002984 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002985 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002986 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2987 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002988 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2989 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2990 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2991 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2992 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2993 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2994 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2995 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002996]
2997
Marat Dukhanfda12b82019-11-21 12:27:59 -08002998FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002999 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
3000 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003001 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
3002 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003003 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
3004 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003005 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
3006 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
3007 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
3008 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
3009 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
3010 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003011 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003012 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
3013 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
3014 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
3015 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003016 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003017 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
3018 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003019 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003020 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
3021 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003022 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
3023 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
3024 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003025 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
3026 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
3027 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
3028 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
3029 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
3030 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
3031 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
3032 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
3033 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
3034 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
3035 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
3036 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
3037 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
3038 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07003039 "src/f32-hswish/gen/hswish-fma3-x8.c",
3040 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003041 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003042 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
3043 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
3044 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
3045 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003046 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003047 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
3048 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003049 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003050 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
3051 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003052 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
3053 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
3054 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003055 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
3056 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
3057 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
3058 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
3059 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
3060 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
3061 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
3062 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003063 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003064 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003065 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003066]
3067
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003068AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003069 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
3070 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003071 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003072 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003073 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003074 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
3075 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003076 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003077 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
3078 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
3079 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003080 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003081 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
3082 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003083 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003084 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003085 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003086 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
3087 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003088 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003089 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
3090 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
3091 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003092 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003093 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
3094 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003095 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003096 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003097 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003098 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
3099 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003100 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003101 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
3102 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
3103 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003104 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08003105 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
3106 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
3107 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
3108 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
3109 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
3110 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
3111 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
3112 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
3113 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
3114 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
3115 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
3116 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
3117 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
3118 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
3119 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
3120 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
3121 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
3122 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
3123 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
3124 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
3125 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
3126 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
3127 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
3128 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
3129 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
3130 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
3131 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
3132 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
3133 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
3134 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003135 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x8.c",
3136 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x16.c",
3137 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x24.c",
3138 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x32.c",
3139 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x40.c",
3140 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x48.c",
3141 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x56.c",
3142 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x64.c",
3143 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x72.c",
3144 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x80.c",
3145 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x8.c",
3146 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x16.c",
3147 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x24.c",
3148 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x32.c",
3149 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x40.c",
3150 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x48.c",
3151 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x56.c",
3152 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x64.c",
3153 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x72.c",
3154 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x80.c",
3155 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x8.c",
3156 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x16.c",
3157 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x24.c",
3158 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x32.c",
3159 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x40.c",
3160 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x48.c",
3161 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x56.c",
3162 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x64.c",
3163 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x72.c",
3164 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x80.c",
3165 "src/f32-velu/gen/velu-avx2-rr1-p6-x8.c",
3166 "src/f32-velu/gen/velu-avx2-rr1-p6-x16.c",
3167 "src/f32-velu/gen/velu-avx2-rr1-p6-x24.c",
3168 "src/f32-velu/gen/velu-avx2-rr1-p6-x32.c",
3169 "src/f32-velu/gen/velu-avx2-rr1-p6-x40.c",
3170 "src/f32-velu/gen/velu-avx2-rr1-p6-x48.c",
3171 "src/f32-velu/gen/velu-avx2-rr1-p6-x56.c",
3172 "src/f32-velu/gen/velu-avx2-rr1-p6-x64.c",
3173 "src/f32-velu/gen/velu-avx2-rr1-p6-x72.c",
3174 "src/f32-velu/gen/velu-avx2-rr1-p6-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003175 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
3176 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
3177 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
3178 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
3179 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
3180 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
3181 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
3182 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
3183 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
3184 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
3185 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
3186 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
3187 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
3188 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
3189 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
3190 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
3191 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
3192 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
3193 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
3194 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
3195 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
3196 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
3197 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
3198 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003199 "src/math/exp-avx2-rr2-lut8-p3-perm.c",
3200 "src/math/exp-avx2-rr2-lut8-p4-perm.c",
3201 "src/math/exp-avx2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003202 "src/math/expm1minus-avx2-rr1-lut4-p4-perm.c",
3203 "src/math/expm1minus-avx2-rr1-lut8-p4-perm.c",
3204 "src/math/expm1minus-avx2-rr1-lut16-p3-gather.c",
3205 "src/math/expm1minus-avx2-rr1-p6.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08003206 "src/math/expminus-avx2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003207 "src/math/extexp-avx2-p5.c",
3208 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
3209 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
3210 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
3211 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
3212 "src/math/sigmoid-avx2-rr1-p5-div.c",
3213 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
3214 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
3215 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
3216 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
3217 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
3218 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
3219 "src/math/sigmoid-avx2-rr2-p5-div.c",
3220 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
3221 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003222 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003223 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003224 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
3225 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003226 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07003227 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07003228 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003229 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003230 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003231 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003232 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07003233 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07003234 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
3235 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
3236 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07003237 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
3238 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
3239 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
3240 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
3241 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
3242 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
3243 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
3244 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003245]
3246
Marat Dukhan08c4a432019-10-03 09:29:21 -07003247AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07003248 "src/f32-clamp/gen/avx512f-x16.c",
3249 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003250 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
3251 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003252 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
3253 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003254 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
3255 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003256 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
3257 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
3258 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
3259 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
3260 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
3261 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003262 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
3263 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
3264 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
3265 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
3266 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
3267 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003268 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
3269 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
3270 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
3271 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
3272 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
3273 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07003274 "src/f32-hswish/gen/hswish-avx512f-x16.c",
3275 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003276 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
3277 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
3278 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
3279 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
3280 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
3281 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07003282 "src/f32-prelu/gen/avx512f-2x16.c",
3283 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003284 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3285 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003286 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003287 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003288 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003289 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3290 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003291 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003292 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3293 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3294 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003295 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003296 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
3297 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003298 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003299 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003300 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003301 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
3302 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003303 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003304 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
3305 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
3306 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003307 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003308 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3309 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003310 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003311 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003312 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003313 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3314 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003315 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003316 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3317 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3318 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003319 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07003320 "src/f32-relu/gen/avx512f-x16.c",
3321 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003322 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003323 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
3324 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
3325 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
3326 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
3327 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
3328 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
3329 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
3330 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
3331 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
3332 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
3333 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
3334 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
3335 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
3336 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
3337 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
3338 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003339 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
3340 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
3341 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
3342 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
3343 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
3344 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
3345 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
3346 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
3347 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
3348 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
3349 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
3350 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
3351 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
3352 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
3353 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
3354 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003355 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
3356 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
3357 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
3358 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
3359 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
3360 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
3361 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
3362 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
3363 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
3364 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
3365 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
3366 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
3367 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
3368 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
3369 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
3370 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003371 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
3372 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
3373 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
3374 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
3375 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
3376 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
3377 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
3378 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08003379 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
3380 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
3381 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
3382 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
3383 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
3384 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
3385 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
3386 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003387 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
3388 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
3389 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
3390 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
3391 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
3392 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
3393 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
3394 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07003395 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
3396 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
3397 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
3398 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003399 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
3400 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
3401 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
3402 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003403 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x16.c",
3404 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x32.c",
3405 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x48.c",
3406 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x64.c",
3407 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x80.c",
3408 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x96.c",
3409 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x112.c",
3410 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x128.c",
3411 "src/f32-velu/gen/velu-avx512f-rr1-p6-x16.c",
3412 "src/f32-velu/gen/velu-avx512f-rr1-p6-x32.c",
3413 "src/f32-velu/gen/velu-avx512f-rr1-p6-x48.c",
3414 "src/f32-velu/gen/velu-avx512f-rr1-p6-x64.c",
3415 "src/f32-velu/gen/velu-avx512f-rr1-p6-x80.c",
3416 "src/f32-velu/gen/velu-avx512f-rr1-p6-x96.c",
3417 "src/f32-velu/gen/velu-avx512f-rr1-p6-x112.c",
3418 "src/f32-velu/gen/velu-avx512f-rr1-p6-x128.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07003419 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
3420 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003421 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
3422 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
3423 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
3424 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
3425 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
3426 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
3427 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
3428 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07003429 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003430 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
3431 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
3432 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
3433 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
3434 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
3435 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
3436 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
3437 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
3438 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
3439 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
3440 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
3441 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
3442 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
3443 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
3444 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
3445 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
3446 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
3447 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
3448 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
3449 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
3450 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
3451 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
3452 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
3453 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003454 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
3455 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
3456 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
3457 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
3458 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
3459 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
3460 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
3461 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07003462 "src/f32-vunary/gen/vabs-avx512f-x16.c",
3463 "src/f32-vunary/gen/vabs-avx512f-x32.c",
3464 "src/f32-vunary/gen/vneg-avx512f-x16.c",
3465 "src/f32-vunary/gen/vneg-avx512f-x32.c",
3466 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
3467 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003468 "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
3469 "src/math/exp-avx512f-rr2-lut16-p3-perm.c",
3470 "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
3471 "src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
3472 "src/math/exp-avx512f-rr2-p5-scalef.c",
3473 "src/math/exp-avx512f-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003474 "src/math/expm1minus-avx512f-rr1-lut16-p3-perm.c",
3475 "src/math/expm1minus-avx512f-rr1-p6.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07003476 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003477 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003478 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003479 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003480 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003481 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003482 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003483 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003484 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003485 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
3486 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
3487 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
3488 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
3489 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
3490 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
3491 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
3492 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
3493 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
3494 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003495 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003496 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003497 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
3498 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
3499 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
3500 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003501 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003502 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003503 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003504]
3505
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003506AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07003507 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
3508 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003509 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
3510 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
3511 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
3512 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
3513 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
3514 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
3515 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
3516 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
3517]
3518
Frank Barchardbcedc082020-08-17 18:00:51 -07003519WASM32_ASM_UKERNELS = [
3520 "src/f32-relu/wasm_shr_x1.S",
3521 "src/f32-relu/wasm_shr_x2.S",
3522 "src/f32-relu/wasm_shr_x4.S",
3523]
3524
Marat Dukhan08c4a432019-10-03 09:29:21 -07003525AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07003526 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07003527 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003528 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3529 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003530 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003531 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07003532 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003533 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003534 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3535 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003536 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
3537 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
3538 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
3539 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003540]
3541
3542AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003543 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003544 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003545 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003546 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003547 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003548 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07003549 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003550 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
3551 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
3552 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
3553 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
3554 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
3555 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
3556 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003557 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
3558 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003559 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3560 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3561 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003562 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
3563 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003564 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3565 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3566 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3567 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003568 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003569 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3570 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003571 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3572 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3573 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3574 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003575 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003576 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003577 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003578 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003579 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3580 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3581 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3582 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3583 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3584 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3585 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3586 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3587 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3588 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3589 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3590 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3591 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3592 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3593 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3594 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3595 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3596 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3597 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3598 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3599 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3600 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003601 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003602 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003603 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3604 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003605 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3606 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3607 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3608 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3609 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3610 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003611 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3612 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3613 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3614 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003615 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3616 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard4a4be4e2021-03-01 15:00:28 -08003617 "src/qs8-gemm/1x16c4-aarch64-neondot-ld32.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003618 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardb8762632021-03-25 04:00:12 -07003619 "src/qs8-gemm/1x8c8-aarch64-neon-mlal-padal.S",
Frank Barchard1dc9fef2021-03-10 23:50:19 -08003620 "src/qs8-gemm/2x8c8-aarch64-neon-mull-padal.S",
Frank Barchardcbb8e702021-03-12 01:22:55 -08003621 "src/qs8-gemm/2x8c8-aarch64-neon-mlal-padal.S",
Frank Barchard5655cb72021-03-09 20:57:14 -08003622 "src/qs8-gemm/2x8c16-aarch64-neon-mlal-padal.S",
Frank Barchard671d1b02021-03-10 15:42:08 -08003623 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barchardda78da12021-03-02 14:28:00 -08003624 "src/qs8-gemm/4x16c4-aarch64-neondot-ld32.S",
Frank Barchard671d1b02021-03-10 15:42:08 -08003625 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Frank Barchard6e35de52021-03-25 04:33:21 -07003626 "src/qs8-igemm/1x8c8-aarch64-neon-mlal-padal.S",
Frank Barchardcbb8e702021-03-12 01:22:55 -08003627 "src/qs8-igemm/2x8c8-aarch64-neon-mlal-padal.S",
Frank Barchard7ca54df2021-03-12 00:09:02 -08003628 "src/qs8-igemm/2x8c16-aarch64-neon-mlal-padal.S",
Frank Barchard89e12f82021-03-08 11:42:44 -08003629 "src/qs8-igemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barchard671d1b02021-03-10 15:42:08 -08003630 "src/qs8-igemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003631]
3632
Marat Dukhan1b354632020-03-23 12:50:22 -07003633INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003634 "src/xnnpack/argmaxpool.h",
3635 "src/xnnpack/avgpool.h",
3636 "src/xnnpack/clamp.h",
3637 "src/xnnpack/common.h",
3638 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003639 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003640 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003641 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003642 "src/xnnpack/gavgpool.h",
3643 "src/xnnpack/gemm.h",
3644 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003645 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003646 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003647 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003648 "src/xnnpack/lut.h",
3649 "src/xnnpack/math.h",
3650 "src/xnnpack/maxpool.h",
3651 "src/xnnpack/packx.h",
3652 "src/xnnpack/pad.h",
3653 "src/xnnpack/params.h",
3654 "src/xnnpack/pavgpool.h",
3655 "src/xnnpack/ppmm.h",
3656 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003657 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003658 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003659 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003660 "src/xnnpack/rmax.h",
3661 "src/xnnpack/scalar-utils.h",
3662 "src/xnnpack/spmm.h",
3663 "src/xnnpack/unpool.h",
3664 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003665 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003666 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003667 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003668 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003669 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003670 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003671 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003672]
3673
3674INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003675 "include/xnnpack.h",
3676 "src/xnnpack/allocator.h",
3677 "src/xnnpack/compute.h",
3678 "src/xnnpack/im2col.h",
3679 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003680 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003681 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003682 "src/xnnpack/operator.h",
3683 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003684 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003685 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003686 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003687 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003688]
3689
Marat Dukhan1b354632020-03-23 12:50:22 -07003690ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003691 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003692]
3693
Marat Dukhan1b354632020-03-23 12:50:22 -07003694MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003695 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003696 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003697]
3698
Marat Dukhan1b354632020-03-23 12:50:22 -07003699MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003700 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003701 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003702 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003703 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003704]
3705
3706OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003707 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003708 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003709]
3710
3711WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003712 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003713 "src/xnnpack/operator.h",
3714 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003715]
3716
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003717LOGGING_COPTS = select({
3718 # No logging in optimized mode
3719 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3720 # Full logging in debug mode
3721 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3722 # Error-only logging in default (fastbuild) mode
3723 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3724})
3725
Marat Dukhan3b59de22020-06-03 20:15:19 -07003726LOGGING_SRCS = select({
3727 # No logging in optimized mode
3728 ":optimized_build": [],
3729 "//conditions:default": [
Marat Dukhanccd3a1d2021-03-29 16:03:12 -07003730 "src/datatype-strings.c",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003731 "src/operator-strings.c",
3732 "src/subgraph-strings.c",
3733 ],
3734})
3735
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003736LOGGING_HDRS = [
3737 "src/xnnpack/log.h",
3738]
3739
Marat Dukhan08c4a432019-10-03 09:29:21 -07003740xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003741 name = "tables",
3742 srcs = TABLE_SRCS,
3743 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003744 gcc_copts = xnnpack_gcc_std_copts(),
3745 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003746)
3747
3748xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003749 name = "scalar_ukernels",
3750 srcs = SCALAR_UKERNELS,
3751 hdrs = INTERNAL_HDRS,
3752 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003753 gcc_copts = xnnpack_gcc_std_copts(),
3754 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003755 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003756 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003757 "@FP16",
3758 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003759 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003760 ],
3761)
3762
3763xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003764 name = "scalar_ukernels_test_mode",
3765 srcs = SCALAR_UKERNELS,
3766 hdrs = INTERNAL_HDRS,
3767 aarch32_copts = ["-marm"],
3768 copts = [
3769 "-UNDEBUG",
3770 "-DXNN_TEST_MODE=1",
3771 ],
3772 gcc_copts = xnnpack_gcc_std_copts(),
3773 msvc_copts = xnnpack_msvc_std_copts(),
3774 deps = [
3775 ":tables",
3776 "@FP16",
3777 "@FXdiv",
3778 "@pthreadpool",
3779 ],
3780)
3781
3782xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003783 name = "wasm_ukernels",
3784 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003785 gcc_copts = xnnpack_gcc_std_copts(),
3786 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003787 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003788 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003789 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003790 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003791 "@FP16",
3792 "@FXdiv",
3793 "@pthreadpool",
3794 ],
3795)
3796
3797xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003798 name = "wasm_ukernels_test_mode",
3799 hdrs = INTERNAL_HDRS,
3800 copts = [
3801 "-UNDEBUG",
3802 "-DXNN_TEST_MODE=1",
3803 ],
3804 gcc_copts = xnnpack_gcc_std_copts(),
3805 msvc_copts = xnnpack_msvc_std_copts(),
3806 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003807 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003808 deps = [
3809 ":tables",
3810 "@FP16",
3811 "@FXdiv",
3812 "@pthreadpool",
3813 ],
3814)
3815
3816xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003817 name = "neon_ukernels",
3818 hdrs = INTERNAL_HDRS,
3819 aarch32_copts = [
3820 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003821 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003822 "-mfpu=neon",
3823 ],
3824 aarch32_srcs = NEON_UKERNELS,
3825 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003826 gcc_copts = xnnpack_gcc_std_copts(),
3827 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003828 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003829 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003830 "@FP16",
3831 "@pthreadpool",
3832 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003833)
3834
3835xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003836 name = "neon_ukernels_test_mode",
3837 hdrs = INTERNAL_HDRS,
3838 aarch32_copts = [
3839 "-marm",
3840 "-march=armv7-a",
3841 "-mfpu=neon",
3842 ],
3843 aarch32_srcs = NEON_UKERNELS,
3844 aarch64_srcs = NEON_UKERNELS,
3845 copts = [
3846 "-UNDEBUG",
3847 "-DXNN_TEST_MODE=1",
3848 ],
3849 gcc_copts = xnnpack_gcc_std_copts(),
3850 msvc_copts = xnnpack_msvc_std_copts(),
3851 deps = [
3852 ":tables",
3853 "@FP16",
3854 "@pthreadpool",
3855 ],
3856)
3857
3858xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003859 name = "neonfma_ukernels",
3860 hdrs = INTERNAL_HDRS,
3861 aarch32_copts = [
3862 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003863 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003864 "-mfpu=neon-vfpv4",
3865 ],
3866 aarch32_srcs = NEONFMA_UKERNELS,
3867 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003868 apple_aarch32_copts = [
3869 "-mcpu=swift",
3870 "-mtune=generic",
3871 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003872 gcc_copts = xnnpack_gcc_std_copts(),
3873 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003874 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003875 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003876 "@FP16",
3877 "@pthreadpool",
3878 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003879)
3880
3881xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003882 name = "neonfma_ukernels_test_mode",
3883 hdrs = INTERNAL_HDRS,
3884 aarch32_copts = [
3885 "-marm",
3886 "-march=armv7-a",
3887 "-mfpu=neon-vfpv4",
3888 ],
3889 aarch32_srcs = NEONFMA_UKERNELS,
3890 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003891 apple_aarch32_copts = [
3892 "-mcpu=swift",
3893 "-mtune=generic",
3894 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003895 copts = [
3896 "-UNDEBUG",
3897 "-DXNN_TEST_MODE=1",
3898 ],
3899 gcc_copts = xnnpack_gcc_std_copts(),
3900 msvc_copts = xnnpack_msvc_std_copts(),
3901 deps = [
3902 ":tables",
3903 "@FP16",
3904 "@pthreadpool",
3905 ],
3906)
3907
3908xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003909 name = "neonv8_ukernels",
3910 hdrs = INTERNAL_HDRS,
3911 aarch32_copts = [
3912 "-marm",
3913 "-march=armv8-a",
3914 "-mfpu=neon-fp-armv8",
3915 ],
3916 aarch32_srcs = NEONV8_UKERNELS,
3917 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003918 apple_aarch32_copts = [
3919 "-mcpu=cyclone",
3920 "-mtune=generic",
3921 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003922 gcc_copts = xnnpack_gcc_std_copts(),
3923 msvc_copts = xnnpack_msvc_std_copts(),
3924 deps = [
3925 ":tables",
3926 "@FP16",
3927 "@pthreadpool",
3928 ],
3929)
3930
3931xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003932 name = "neonv8_ukernels_test_mode",
3933 hdrs = INTERNAL_HDRS,
3934 aarch32_copts = [
3935 "-marm",
3936 "-march=armv8-a",
3937 "-mfpu=neon-fp-armv8",
3938 ],
3939 aarch32_srcs = NEONV8_UKERNELS,
3940 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003941 apple_aarch32_copts = [
3942 "-mcpu=cyclone",
3943 "-mtune=generic",
3944 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003945 copts = [
3946 "-UNDEBUG",
3947 "-DXNN_TEST_MODE=1",
3948 ],
3949 gcc_copts = xnnpack_gcc_std_copts(),
3950 msvc_copts = xnnpack_msvc_std_copts(),
3951 deps = [
3952 ":tables",
3953 "@FP16",
3954 "@pthreadpool",
3955 ],
3956)
3957
3958xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003959 name = "neonfp16arith_ukernels",
3960 hdrs = INTERNAL_HDRS,
3961 aarch64_copts = ["-march=armv8.2-a+fp16"],
3962 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003963 gcc_copts = xnnpack_gcc_std_copts(),
3964 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003965 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003966 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003967 "@FP16",
3968 "@pthreadpool",
3969 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003970)
3971
3972xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003973 name = "neonfp16arith_ukernels_test_mode",
3974 hdrs = INTERNAL_HDRS,
3975 aarch64_copts = ["-march=armv8.2-a+fp16"],
3976 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3977 copts = [
3978 "-UNDEBUG",
3979 "-DXNN_TEST_MODE=1",
3980 ],
3981 gcc_copts = xnnpack_gcc_std_copts(),
3982 msvc_copts = xnnpack_msvc_std_copts(),
3983 deps = [
3984 ":tables",
3985 "@FP16",
3986 "@pthreadpool",
3987 ],
3988)
3989
3990xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003991 name = "neondot_ukernels",
3992 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003993 aarch32_copts = [
3994 "-marm",
3995 "-march=armv8.2-a+dotprod",
3996 "-mfpu=neon-fp-armv8",
3997 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003998 aarch32_srcs = NEONDOT_UKERNELS,
3999 aarch64_copts = ["-march=armv8.2-a+dotprod"],
4000 aarch64_srcs = NEONDOT_UKERNELS,
4001 gcc_copts = xnnpack_gcc_std_copts(),
4002 msvc_copts = xnnpack_msvc_std_copts(),
4003 deps = [
4004 ":tables",
4005 "@FP16",
4006 "@pthreadpool",
4007 ],
4008)
4009
4010xnnpack_cc_library(
4011 name = "neondot_ukernels_test_mode",
4012 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07004013 aarch32_copts = [
4014 "-marm",
4015 "-march=armv8.2-a+dotprod",
4016 "-mfpu=neon-fp-armv8",
4017 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07004018 aarch32_srcs = NEONDOT_UKERNELS,
4019 aarch64_copts = ["-march=armv8.2-a+dotprod"],
4020 aarch64_srcs = NEONDOT_UKERNELS,
4021 copts = [
4022 "-UNDEBUG",
4023 "-DXNN_TEST_MODE=1",
4024 ],
4025 gcc_copts = xnnpack_gcc_std_copts(),
4026 msvc_copts = xnnpack_msvc_std_copts(),
4027 deps = [
4028 ":tables",
4029 "@FP16",
4030 "@pthreadpool",
4031 ],
4032)
4033
4034xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004035 name = "sse2_ukernels",
4036 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004037 gcc_copts = xnnpack_gcc_std_copts(),
4038 gcc_x86_copts = ["-msse2"],
4039 msvc_copts = xnnpack_msvc_std_copts(),
4040 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004041 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004042 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004043 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004044 "@FP16",
4045 "@pthreadpool",
4046 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004047)
4048
4049xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004050 name = "sse2_ukernels_test_mode",
4051 hdrs = INTERNAL_HDRS,
4052 copts = [
4053 "-UNDEBUG",
4054 "-DXNN_TEST_MODE=1",
4055 ],
4056 gcc_copts = xnnpack_gcc_std_copts(),
4057 gcc_x86_copts = ["-msse2"],
4058 msvc_copts = xnnpack_msvc_std_copts(),
4059 msvc_x86_32_copts = ["/arch:SSE2"],
4060 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
4061 deps = [
4062 ":tables",
4063 "@FP16",
4064 "@pthreadpool",
4065 ],
4066)
4067
4068xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004069 name = "ssse3_ukernels",
4070 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004071 gcc_copts = xnnpack_gcc_std_copts(),
4072 gcc_x86_copts = ["-mssse3"],
4073 msvc_copts = xnnpack_msvc_std_copts(),
4074 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004075 x86_srcs = SSSE3_UKERNELS,
4076 deps = [
4077 ":tables",
4078 "@FP16",
4079 "@pthreadpool",
4080 ],
4081)
4082
4083xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004084 name = "ssse3_ukernels_test_mode",
4085 hdrs = INTERNAL_HDRS,
4086 copts = [
4087 "-UNDEBUG",
4088 "-DXNN_TEST_MODE=1",
4089 ],
4090 gcc_copts = xnnpack_gcc_std_copts(),
4091 gcc_x86_copts = ["-mssse3"],
4092 msvc_copts = xnnpack_msvc_std_copts(),
4093 msvc_x86_32_copts = ["/arch:SSE2"],
4094 x86_srcs = SSSE3_UKERNELS,
4095 deps = [
4096 ":tables",
4097 "@FP16",
4098 "@pthreadpool",
4099 ],
4100)
4101
4102xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004103 name = "sse41_ukernels",
4104 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004105 gcc_copts = xnnpack_gcc_std_copts(),
4106 gcc_x86_copts = ["-msse4.1"],
4107 msvc_copts = xnnpack_msvc_std_copts(),
4108 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004109 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004110 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004111 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004112 "@FP16",
4113 "@pthreadpool",
4114 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004115)
4116
4117xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004118 name = "sse41_ukernels_test_mode",
4119 hdrs = INTERNAL_HDRS,
4120 copts = [
4121 "-UNDEBUG",
4122 "-DXNN_TEST_MODE=1",
4123 ],
4124 gcc_copts = xnnpack_gcc_std_copts(),
4125 gcc_x86_copts = ["-msse4.1"],
4126 msvc_copts = xnnpack_msvc_std_copts(),
4127 msvc_x86_32_copts = ["/arch:SSE2"],
4128 x86_srcs = SSE41_UKERNELS,
4129 deps = [
4130 ":tables",
4131 "@FP16",
4132 "@pthreadpool",
4133 ],
4134)
4135
4136xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004137 name = "avx_ukernels",
4138 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004139 gcc_copts = xnnpack_gcc_std_copts(),
4140 gcc_x86_copts = ["-mavx"],
4141 msvc_copts = xnnpack_msvc_std_copts(),
4142 msvc_x86_32_copts = ["/arch:AVX"],
4143 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004144 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004145 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004146 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004147 "@FP16",
4148 "@pthreadpool",
4149 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004150)
4151
4152xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004153 name = "avx_ukernels_test_mode",
4154 hdrs = INTERNAL_HDRS,
4155 copts = [
4156 "-UNDEBUG",
4157 "-DXNN_TEST_MODE=1",
4158 ],
4159 gcc_copts = xnnpack_gcc_std_copts(),
4160 gcc_x86_copts = ["-mavx"],
4161 msvc_copts = xnnpack_msvc_std_copts(),
4162 msvc_x86_32_copts = ["/arch:AVX"],
4163 msvc_x86_64_copts = ["/arch:AVX"],
4164 x86_srcs = AVX_UKERNELS,
4165 deps = [
4166 ":tables",
4167 "@FP16",
4168 "@pthreadpool",
4169 ],
4170)
4171
4172xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07004173 name = "xop_ukernels",
4174 hdrs = INTERNAL_HDRS,
4175 gcc_copts = xnnpack_gcc_std_copts(),
4176 gcc_x86_copts = ["-mxop"],
4177 msvc_copts = xnnpack_msvc_std_copts(),
4178 msvc_x86_32_copts = ["/arch:AVX"],
4179 msvc_x86_64_copts = ["/arch:AVX"],
4180 x86_srcs = XOP_UKERNELS,
4181 deps = [
4182 ":tables",
4183 "@FP16",
4184 "@pthreadpool",
4185 ],
4186)
4187
4188xnnpack_cc_library(
4189 name = "xop_ukernels_test_mode",
4190 hdrs = INTERNAL_HDRS,
4191 copts = [
4192 "-UNDEBUG",
4193 "-DXNN_TEST_MODE=1",
4194 ],
4195 gcc_copts = xnnpack_gcc_std_copts(),
4196 gcc_x86_copts = ["-mxop"],
4197 msvc_copts = xnnpack_msvc_std_copts(),
4198 msvc_x86_32_copts = ["/arch:AVX"],
4199 msvc_x86_64_copts = ["/arch:AVX"],
4200 x86_srcs = XOP_UKERNELS,
4201 deps = [
4202 ":tables",
4203 "@FP16",
4204 "@pthreadpool",
4205 ],
4206)
4207
4208xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08004209 name = "fma3_ukernels",
4210 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004211 gcc_copts = xnnpack_gcc_std_copts(),
4212 gcc_x86_copts = ["-mfma"],
4213 msvc_copts = xnnpack_msvc_std_copts(),
4214 msvc_x86_32_copts = ["/arch:AVX"],
4215 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08004216 x86_srcs = FMA3_UKERNELS,
4217 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004218 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004219 "@FP16",
4220 "@pthreadpool",
4221 ],
4222)
4223
4224xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004225 name = "fma3_ukernels_test_mode",
4226 hdrs = INTERNAL_HDRS,
4227 copts = [
4228 "-UNDEBUG",
4229 "-DXNN_TEST_MODE=1",
4230 ],
4231 gcc_copts = xnnpack_gcc_std_copts(),
4232 gcc_x86_copts = ["-mfma"],
4233 msvc_copts = xnnpack_msvc_std_copts(),
4234 msvc_x86_32_copts = ["/arch:AVX"],
4235 msvc_x86_64_copts = ["/arch:AVX"],
4236 x86_srcs = FMA3_UKERNELS,
4237 deps = [
4238 ":tables",
4239 "@FP16",
4240 "@pthreadpool",
4241 ],
4242)
4243
4244xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004245 name = "avx2_ukernels",
4246 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004247 gcc_copts = xnnpack_gcc_std_copts(),
4248 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004249 "-mfma",
4250 "-mavx2",
4251 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004252 msvc_copts = xnnpack_msvc_std_copts(),
4253 msvc_x86_32_copts = ["/arch:AVX2"],
4254 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004255 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004256 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004257 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004258 "@FP16",
4259 "@pthreadpool",
4260 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004261)
4262
4263xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004264 name = "avx2_ukernels_test_mode",
4265 hdrs = INTERNAL_HDRS,
4266 copts = [
4267 "-UNDEBUG",
4268 "-DXNN_TEST_MODE=1",
4269 ],
4270 gcc_copts = xnnpack_gcc_std_copts(),
4271 gcc_x86_copts = [
4272 "-mfma",
4273 "-mavx2",
4274 ],
4275 msvc_copts = xnnpack_msvc_std_copts(),
4276 msvc_x86_32_copts = ["/arch:AVX2"],
4277 msvc_x86_64_copts = ["/arch:AVX2"],
4278 x86_srcs = AVX2_UKERNELS,
4279 deps = [
4280 ":tables",
4281 "@FP16",
4282 "@pthreadpool",
4283 ],
4284)
4285
4286xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004287 name = "avx512f_ukernels",
4288 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004289 gcc_copts = xnnpack_gcc_std_copts(),
4290 gcc_x86_copts = ["-mavx512f"],
4291 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4292 msvc_copts = xnnpack_msvc_std_copts(),
4293 msvc_x86_32_copts = ["/arch:AVX512"],
4294 msvc_x86_64_copts = ["/arch:AVX512"],
4295 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004296 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004297 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004298 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004299 "@FP16",
4300 "@pthreadpool",
4301 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004302)
4303
4304xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004305 name = "avx512f_ukernels_test_mode",
4306 hdrs = INTERNAL_HDRS,
4307 copts = [
4308 "-UNDEBUG",
4309 "-DXNN_TEST_MODE=1",
4310 ],
4311 gcc_copts = xnnpack_gcc_std_copts(),
4312 gcc_x86_copts = ["-mavx512f"],
4313 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4314 msvc_copts = xnnpack_msvc_std_copts(),
4315 msvc_x86_32_copts = ["/arch:AVX512"],
4316 msvc_x86_64_copts = ["/arch:AVX512"],
4317 msys_copts = ["-fno-asynchronous-unwind-tables"],
4318 x86_srcs = AVX512F_UKERNELS,
4319 deps = [
4320 ":tables",
4321 "@FP16",
4322 "@pthreadpool",
4323 ],
4324)
4325
4326xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004327 name = "avx512skx_ukernels",
4328 hdrs = INTERNAL_HDRS,
4329 gcc_copts = xnnpack_gcc_std_copts(),
4330 gcc_x86_copts = [
4331 "-mavx512f",
4332 "-mavx512cd",
4333 "-mavx512bw",
4334 "-mavx512dq",
4335 "-mavx512vl",
4336 ],
4337 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4338 msvc_copts = xnnpack_msvc_std_copts(),
4339 msvc_x86_32_copts = ["/arch:AVX512"],
4340 msvc_x86_64_copts = ["/arch:AVX512"],
4341 msys_copts = ["-fno-asynchronous-unwind-tables"],
4342 x86_srcs = AVX512SKX_UKERNELS,
4343 deps = [
4344 ":tables",
4345 "@FP16",
4346 "@pthreadpool",
4347 ],
4348)
4349
4350xnnpack_cc_library(
4351 name = "avx512skx_ukernels_test_mode",
4352 hdrs = INTERNAL_HDRS,
4353 copts = [
4354 "-UNDEBUG",
4355 "-DXNN_TEST_MODE=1",
4356 ],
4357 gcc_copts = xnnpack_gcc_std_copts(),
4358 gcc_x86_copts = [
4359 "-mavx512f",
4360 "-mavx512cd",
4361 "-mavx512bw",
4362 "-mavx512dq",
4363 "-mavx512vl",
4364 ],
4365 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4366 msvc_copts = xnnpack_msvc_std_copts(),
4367 msvc_x86_32_copts = ["/arch:AVX512"],
4368 msvc_x86_64_copts = ["/arch:AVX512"],
4369 msys_copts = ["-fno-asynchronous-unwind-tables"],
4370 x86_srcs = AVX512SKX_UKERNELS,
4371 deps = [
4372 ":tables",
4373 "@FP16",
4374 "@pthreadpool",
4375 ],
4376)
4377
4378xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004379 name = "asm_ukernels",
4380 hdrs = ["src/xnnpack/assembly.h"],
4381 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07004382 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004383 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07004384 wasm_srcs = WASM32_ASM_UKERNELS,
4385 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004386)
4387
Marat Dukhan3b59de22020-06-03 20:15:19 -07004388xnnpack_cc_library(
4389 name = "logging_utils",
4390 srcs = LOGGING_SRCS,
4391 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4392 copts = LOGGING_COPTS + [
4393 "-Isrc",
4394 "-Iinclude",
4395 ] + select({
4396 ":debug_build": [],
4397 "//conditions:default": xnnpack_min_size_copts(),
4398 }),
4399 gcc_copts = xnnpack_gcc_std_copts(),
4400 msvc_copts = xnnpack_msvc_std_copts(),
4401 visibility = xnnpack_visibility(),
4402 deps = [
4403 "@FP16",
4404 "@clog",
4405 "@pthreadpool",
4406 ],
4407)
4408
Marat Dukhan08c4a432019-10-03 09:29:21 -07004409xnnpack_aggregate_library(
4410 name = "ukernels",
4411 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004412 ":neon_ukernels",
4413 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004414 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004415 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004416 ":asm_ukernels",
4417 ],
4418 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004419 ":neon_ukernels",
4420 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004421 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004422 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004423 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004424 ":asm_ukernels",
4425 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004426 generic_deps = [
4427 ":scalar_ukernels",
4428 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004429 wasm_deps = [
4430 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004431 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004432 ],
4433 wasmsimd_deps = [
4434 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004435 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004436 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004437 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004438 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004439 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004440 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004441 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004442 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004443 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004444 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004445 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004446 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004447 ],
4448)
4449
Marat Dukhan33fcf782020-05-24 14:27:15 -07004450xnnpack_aggregate_library(
4451 name = "ukernels_test_mode",
4452 aarch32_deps = [
4453 ":neon_ukernels_test_mode",
4454 ":neonfma_ukernels_test_mode",
4455 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004456 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004457 ":asm_ukernels",
4458 ],
4459 aarch64_deps = [
4460 ":neon_ukernels_test_mode",
4461 ":neonfma_ukernels_test_mode",
4462 ":neonv8_ukernels_test_mode",
4463 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004464 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004465 ":asm_ukernels",
4466 ],
4467 generic_deps = [
4468 ":scalar_ukernels_test_mode",
4469 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004470 wasm_deps = [
4471 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004472 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004473 ],
4474 wasmsimd_deps = [
4475 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004476 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004477 ],
4478 x86_deps = [
4479 ":sse2_ukernels_test_mode",
4480 ":ssse3_ukernels_test_mode",
4481 ":sse41_ukernels_test_mode",
4482 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004483 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004484 ":fma3_ukernels_test_mode",
4485 ":avx2_ukernels_test_mode",
4486 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004487 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004488 ],
4489)
4490
Marat Dukhan08c4a432019-10-03 09:29:21 -07004491xnnpack_cc_library(
4492 name = "im2col",
4493 srcs = ["src/im2col.c"],
4494 hdrs = [
4495 "src/xnnpack/common.h",
4496 "src/xnnpack/im2col.h",
4497 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004498 gcc_copts = xnnpack_gcc_std_copts(),
4499 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004500)
4501
4502xnnpack_cc_library(
4503 name = "indirection",
4504 srcs = ["src/indirection.c"],
4505 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004506 gcc_copts = xnnpack_gcc_std_copts(),
4507 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004508 deps = [
4509 "@FP16",
4510 "@FXdiv",
4511 "@pthreadpool",
4512 ],
4513)
4514
4515xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004516 name = "indirection_test_mode",
4517 srcs = ["src/indirection.c"],
4518 hdrs = INTERNAL_HDRS,
4519 copts = [
4520 "-UNDEBUG",
4521 "-DXNN_TEST_MODE=1",
4522 ],
4523 gcc_copts = xnnpack_gcc_std_copts(),
4524 msvc_copts = xnnpack_msvc_std_copts(),
4525 deps = [
4526 "@FP16",
4527 "@FXdiv",
4528 "@pthreadpool",
4529 ],
4530)
4531
4532xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004533 name = "packing",
4534 srcs = ["src/packing.c"],
4535 hdrs = INTERNAL_HDRS,
4536 gcc_copts = xnnpack_gcc_std_copts(),
4537 msvc_copts = xnnpack_msvc_std_copts(),
4538 deps = [
4539 "@FP16",
4540 "@FXdiv",
4541 "@pthreadpool",
4542 ],
4543)
4544
4545xnnpack_cc_library(
4546 name = "packing_test_mode",
4547 srcs = ["src/packing.c"],
4548 hdrs = INTERNAL_HDRS,
4549 copts = [
4550 "-UNDEBUG",
4551 "-DXNN_TEST_MODE=1",
4552 ],
4553 gcc_copts = xnnpack_gcc_std_copts(),
4554 msvc_copts = xnnpack_msvc_std_copts(),
4555 deps = [
4556 "@FP16",
4557 "@FXdiv",
4558 "@pthreadpool",
4559 ],
4560)
4561
4562xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004563 name = "operator_run",
4564 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004565 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004566 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004567 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4568 "//conditions:default": [],
4569 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004570 gcc_copts = xnnpack_gcc_std_copts(),
4571 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004572 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004573 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004574 "@FP16",
4575 "@FXdiv",
4576 "@clog",
4577 "@pthreadpool",
4578 ],
4579)
4580
Chao Mei6ddfc602020-05-13 22:29:36 -07004581xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004582 name = "operator_run_test_mode",
4583 srcs = ["src/operator-run.c"],
4584 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4585 copts = LOGGING_COPTS + [
4586 "-UNDEBUG",
4587 "-DXNN_TEST_MODE=1",
4588 ] + select({
4589 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4590 "//conditions:default": [],
4591 }),
4592 gcc_copts = xnnpack_gcc_std_copts(),
4593 msvc_copts = xnnpack_msvc_std_copts(),
4594 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004595 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004596 "@FP16",
4597 "@FXdiv",
4598 "@clog",
4599 "@pthreadpool",
4600 ],
4601)
4602
4603xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004604 name = "memory_planner",
4605 srcs = ["src/memory-planner.c"],
4606 hdrs = INTERNAL_HDRS,
4607 defines = select({
4608 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4609 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4610 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4611 }),
4612 gcc_copts = xnnpack_gcc_std_copts(),
4613 msvc_copts = xnnpack_msvc_std_copts(),
4614 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004615 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004616 "@pthreadpool",
4617 ],
4618)
4619
Marat Dukhan33fcf782020-05-24 14:27:15 -07004620xnnpack_cc_library(
4621 name = "memory_planner_test_mode",
4622 srcs = ["src/memory-planner.c"],
4623 hdrs = INTERNAL_HDRS,
4624 copts = [
4625 "-UNDEBUG",
4626 "-DXNN_TEST_MODE=1",
4627 ],
4628 defines = select({
4629 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4630 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4631 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4632 }),
4633 gcc_copts = xnnpack_gcc_std_copts(),
4634 msvc_copts = xnnpack_msvc_std_copts(),
4635 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004636 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004637 "@pthreadpool",
4638 ],
4639)
4640
Marat Dukhan08c4a432019-10-03 09:29:21 -07004641cc_library(
4642 name = "enable_assembly",
4643 defines = select({
4644 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4645 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004646 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004647 }),
4648)
4649
Marat Dukhan9de90e02020-06-18 16:04:12 -07004650cc_library(
4651 name = "enable_sparse",
4652 defines = select({
4653 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4654 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhanb36582b2020-12-08 11:16:28 -08004655 "//conditions:default": ["XNN_ENABLE_SPARSE=1"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004656 }),
4657)
4658
Marat Dukhancf056b22019-10-07 10:26:29 -07004659xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004660 name = "operators",
4661 srcs = OPERATOR_SRCS + [
Marat Dukhan496389f2021-04-07 15:47:12 -07004662 "src/allocator.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004663 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004664 ],
4665 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004666 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004667 "-Isrc",
4668 "-Iinclude",
4669 ] + select({
4670 ":debug_build": [],
4671 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004672 }) + select({
4673 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4674 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004675 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004676 gcc_copts = xnnpack_gcc_std_copts(),
4677 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004678 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004679 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004680 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004681 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004682 "@FP16",
4683 "@FXdiv",
4684 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004685 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004686 ],
4687)
4688
Marat Dukhan10a38082020-04-17 03:58:35 -07004689xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004690 name = "operators_test_mode",
4691 srcs = OPERATOR_SRCS + [
Marat Dukhan496389f2021-04-07 15:47:12 -07004692 "src/allocator.c",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004693 "src/operator-delete.c",
4694 ],
4695 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4696 copts = LOGGING_COPTS + [
4697 "-Isrc",
4698 "-Iinclude",
4699 "-UNDEBUG",
4700 "-DXNN_TEST_MODE=1",
4701 ] + select({
4702 ":debug_build": [],
4703 "//conditions:default": xnnpack_min_size_copts(),
4704 }) + select({
4705 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4706 "//conditions:default": [],
4707 }),
4708 gcc_copts = xnnpack_gcc_std_copts(),
4709 msvc_copts = xnnpack_msvc_std_copts(),
4710 deps = [
4711 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004712 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004713 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004714 "@FP16",
4715 "@FXdiv",
4716 "@clog",
4717 "@pthreadpool",
4718 ],
4719)
4720
4721xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004722 name = "XNNPACK",
4723 srcs = [
4724 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004725 "src/runtime.c",
4726 "src/subgraph.c",
4727 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004728 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004729 hdrs = ["include/xnnpack.h"],
4730 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004731 "-Isrc",
4732 "-Iinclude",
4733 ] + select({
4734 ":debug_build": [],
4735 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004736 }) + select({
4737 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4738 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004739 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004740 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004741 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004742 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004743 visibility = xnnpack_visibility(),
4744 deps = [
4745 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004746 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004747 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004748 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004749 ":operator_run",
4750 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004751 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004752 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004753 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004754 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004755 ] + select({
4756 ":emscripten": [],
4757 "//conditions:default": ["@cpuinfo"],
4758 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004759)
4760
Marat Dukhan10a38082020-04-17 03:58:35 -07004761xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004762 name = "XNNPACK_test_mode",
4763 srcs = [
4764 "src/init.c",
4765 "src/runtime.c",
4766 "src/subgraph.c",
4767 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004768 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004769 hdrs = ["include/xnnpack.h"],
4770 copts = LOGGING_COPTS + [
4771 "-Isrc",
4772 "-Iinclude",
4773 "-UNDEBUG",
4774 "-DXNN_TEST_MODE=1",
4775 ] + select({
4776 ":debug_build": [],
4777 "//conditions:default": xnnpack_min_size_copts(),
4778 }) + select({
4779 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4780 "//conditions:default": [],
4781 }),
4782 gcc_copts = xnnpack_gcc_std_copts(),
4783 includes = ["include"],
4784 msvc_copts = xnnpack_msvc_std_copts(),
4785 visibility = xnnpack_visibility(),
4786 deps = [
4787 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004788 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004789 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004790 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004791 ":operator_run_test_mode",
4792 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004793 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004794 "@clog",
4795 "@FP16",
4796 "@pthreadpool",
4797 ] + select({
4798 ":emscripten": [],
4799 "//conditions:default": ["@cpuinfo"],
4800 }),
4801)
4802
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004803# Specialized XNNPACK version for TensorFlow Lite. Excludes operators currently
4804# not used by the TensorFlow Lite XNNPACK delegate to minimize code size.
Marat Dukhanae046f52020-06-15 13:16:14 -07004805xnnpack_cc_library(
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004806 name = "xnnpack_for_tflite",
4807 srcs = [
4808 "src/init.c",
4809 "src/runtime.c",
4810 "src/subgraph.c",
4811 "src/tensor.c",
4812 ] + SUBGRAPH_SRCS,
4813 hdrs = ["include/xnnpack.h"],
4814 copts = LOGGING_COPTS + [
4815 "-Isrc",
4816 "-Iinclude",
4817 ] + select({
4818 ":debug_build": [],
4819 "//conditions:default": xnnpack_min_size_copts(),
4820 }) + select({
4821 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4822 "//conditions:default": [],
4823 }),
4824 defines = [
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004825 "XNN_NO_QU8_OPERATORS",
4826 "XNN_NO_U8_OPERATORS",
4827 "XNN_NO_X8_OPERATORS",
4828 "XNN_NO_F16_OPERATORS",
4829 "XNN_NO_X16_OPERATORS",
Marat Dukhanb939cdb2021-03-30 18:51:51 -07004830 ] + select({
4831 ":xnn_enable_qs8_explicit_true": [],
4832 ":xnn_enable_qs8_explicit_false": ["XNN_NO_QS8_OPERATORS"],
4833 "//conditions:default": ["XNN_NO_QS8_OPERATORS"],
4834 }),
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004835 gcc_copts = xnnpack_gcc_std_copts(),
4836 includes = ["include"],
4837 msvc_copts = xnnpack_msvc_std_copts(),
4838 visibility = xnnpack_visibility(),
4839 deps = [
4840 ":enable_assembly",
4841 ":enable_sparse",
4842 ":logging_utils",
4843 ":memory_planner",
4844 ":operator_run",
4845 ":operators",
Marat Dukhand09ca262021-03-30 16:17:12 -07004846 ":ukernels",
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07004847 "@clog",
4848 "@FP16",
4849 "@pthreadpool",
4850 ] + select({
4851 ":emscripten": [],
4852 "//conditions:default": ["@cpuinfo"],
4853 }),
4854)
4855
4856# Specialized XNNPACK version for TensorFlow.js. Excludes operators currently
4857# not used by the TensorFlow.js WebAssembly backend to minimize code size.
4858xnnpack_cc_library(
4859 name = "xnnpack_for_tfjs",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004860 srcs = [
4861 "src/init.c",
4862 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004863 hdrs = ["include/xnnpack.h"],
4864 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004865 "-Isrc",
4866 "-Iinclude",
4867 ] + select({
4868 ":debug_build": [],
4869 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004870 }) + select({
4871 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4872 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004873 }),
4874 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004875 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004876 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004877 "XNN_NO_U8_OPERATORS",
4878 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004879 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004880 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004881 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004882 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004883 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004884 visibility = xnnpack_visibility(),
4885 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004886 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004887 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004888 ":operator_run",
4889 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004890 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004891 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004892 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004893 ] + select({
4894 ":emscripten": [],
4895 "//conditions:default": ["@cpuinfo"],
4896 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004897)
4898
Marat Dukhancf056b22019-10-07 10:26:29 -07004899xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004900 name = "bench_utils",
4901 srcs = ["bench/utils.cc"],
4902 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004903 deps = [
4904 "@com_google_benchmark//:benchmark",
4905 "@cpuinfo",
4906 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004907)
4908
Frank Barchard7e955972019-10-11 10:34:25 -07004909######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004910
4911xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004912 name = "qs8_gemm_bench",
4913 srcs = [
4914 "bench/gemm.h",
4915 "bench/qs8-gemm.cc",
4916 "src/xnnpack/AlignedAllocator.h",
4917 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004918 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4919 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004920)
4921
4922xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004923 name = "qs8_requantization_bench",
4924 srcs = [
4925 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004926 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004927 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004928 ] + MICROKERNEL_BENCHMARK_HDRS,
4929 deps = MICROKERNEL_BENCHMARK_DEPS,
4930)
4931
4932xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004933 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004934 srcs = [
4935 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004936 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004937 "src/xnnpack/AlignedAllocator.h",
4938 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004939 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004940 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004941)
4942
4943xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004944 name = "qu8_requantization_bench",
4945 srcs = [
4946 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004947 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004948 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004949 ] + MICROKERNEL_BENCHMARK_HDRS,
4950 deps = MICROKERNEL_BENCHMARK_DEPS,
4951)
4952
4953xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004954 name = "f16_igemm_bench",
4955 srcs = [
4956 "bench/f16-igemm.cc",
4957 "bench/conv.h",
4958 "bench/google/conv.h",
4959 "src/xnnpack/AlignedAllocator.h",
4960 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004961 deps = MICROKERNEL_BENCHMARK_DEPS + [
4962 ":indirection",
4963 ":packing",
4964 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004965)
4966
4967xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004968 name = "f16_gemm_bench",
4969 srcs = [
4970 "bench/f16-gemm.cc",
4971 "bench/gemm.h",
4972 "src/xnnpack/AlignedAllocator.h",
4973 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004974 deps = MICROKERNEL_BENCHMARK_DEPS + [
4975 ":packing",
4976 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004977)
4978
4979xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004980 name = "f16_spmm_bench",
4981 srcs = [
4982 "bench/f16-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08004983 "bench/spmm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004984 "src/xnnpack/AlignedAllocator.h",
4985 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004986 deps = MICROKERNEL_BENCHMARK_DEPS,
4987)
4988
4989xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004990 name = "f32_igemm_bench",
4991 srcs = [
4992 "bench/f32-igemm.cc",
4993 "bench/conv.h",
4994 "src/xnnpack/AlignedAllocator.h",
4995 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004996 deps = MICROKERNEL_BENCHMARK_DEPS + [
4997 ":indirection",
4998 ":packing",
4999 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005000)
5001
5002xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07005003 name = "f16_relu_bench",
5004 srcs = [
5005 "bench/f16-relu.cc",
5006 "src/xnnpack/AlignedAllocator.h",
5007 ] + MICROKERNEL_BENCHMARK_HDRS,
5008 deps = MICROKERNEL_BENCHMARK_DEPS,
5009)
5010
5011xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005012 name = "f32_conv_hwc_bench",
5013 srcs = [
5014 "bench/f32-conv-hwc.cc",
5015 "bench/dconv.h",
5016 "src/xnnpack/AlignedAllocator.h",
5017 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005018 deps = MICROKERNEL_BENCHMARK_DEPS + [
5019 ":packing",
5020 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005021)
5022
5023xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005024 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07005025 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005026 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07005027 "bench/dconv.h",
5028 "src/xnnpack/AlignedAllocator.h",
5029 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005030 deps = MICROKERNEL_BENCHMARK_DEPS + [
5031 ":packing",
5032 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07005033)
5034
5035xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07005036 name = "f16_dwconv_bench",
5037 srcs = [
5038 "bench/f16-dwconv.cc",
5039 "bench/dwconv.h",
5040 "bench/google/dwconv.h",
5041 "src/xnnpack/AlignedAllocator.h",
5042 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005043 deps = MICROKERNEL_BENCHMARK_DEPS + [
5044 ":indirection",
5045 ":packing",
5046 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07005047)
5048
5049xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005050 name = "f32_dwconv_bench",
5051 srcs = [
5052 "bench/f32-dwconv.cc",
5053 "bench/dwconv.h",
5054 "src/xnnpack/AlignedAllocator.h",
5055 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005056 deps = MICROKERNEL_BENCHMARK_DEPS + [
5057 ":indirection",
5058 ":packing",
5059 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005060)
5061
5062xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005063 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005064 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005065 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005066 "bench/dwconv.h",
5067 "src/xnnpack/AlignedAllocator.h",
5068 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005069 deps = MICROKERNEL_BENCHMARK_DEPS + [
5070 ":indirection",
5071 ":packing",
5072 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005073)
5074
5075xnnpack_benchmark(
5076 name = "f32_gemm_bench",
5077 srcs = [
5078 "bench/f32-gemm.cc",
5079 "bench/gemm.h",
5080 "src/xnnpack/AlignedAllocator.h",
5081 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005082 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07005083 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005084)
5085
5086xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07005087 name = "f32_hswish_bench",
5088 srcs = [
5089 "bench/f32-hswish.cc",
5090 "src/xnnpack/AlignedAllocator.h",
5091 ] + MICROKERNEL_BENCHMARK_HDRS,
5092 deps = MICROKERNEL_BENCHMARK_DEPS,
5093)
5094
5095xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08005096 name = "f32_raddexpminusmax_bench",
5097 srcs = [
5098 "bench/f32-raddexpminusmax.cc",
5099 "src/xnnpack/AlignedAllocator.h",
5100 ] + MICROKERNEL_BENCHMARK_HDRS,
5101 deps = MICROKERNEL_BENCHMARK_DEPS,
5102)
5103
5104xnnpack_benchmark(
5105 name = "f32_raddextexp_bench",
5106 srcs = [
5107 "bench/f32-raddextexp.cc",
5108 "src/xnnpack/AlignedAllocator.h",
5109 ] + MICROKERNEL_BENCHMARK_HDRS,
5110 deps = MICROKERNEL_BENCHMARK_DEPS,
5111)
5112
5113xnnpack_benchmark(
5114 name = "f32_raddstoreexpminusmax_bench",
5115 srcs = [
5116 "bench/f32-raddstoreexpminusmax.cc",
5117 "src/xnnpack/AlignedAllocator.h",
5118 ] + MICROKERNEL_BENCHMARK_HDRS,
5119 deps = MICROKERNEL_BENCHMARK_DEPS,
5120)
5121
5122xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07005123 name = "f32_relu_bench",
5124 srcs = [
5125 "bench/f32-relu.cc",
5126 "src/xnnpack/AlignedAllocator.h",
5127 ] + MICROKERNEL_BENCHMARK_HDRS,
5128 deps = MICROKERNEL_BENCHMARK_DEPS,
5129)
5130
5131xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005132 name = "f32_rmax_bench",
5133 srcs = [
5134 "bench/f32-rmax.cc",
5135 "src/xnnpack/AlignedAllocator.h",
5136 ] + MICROKERNEL_BENCHMARK_HDRS,
5137 deps = MICROKERNEL_BENCHMARK_DEPS,
5138)
5139
5140xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08005141 name = "f32_sigmoid_bench",
5142 srcs = [
5143 "bench/f32-sigmoid.cc",
5144 "src/xnnpack/AlignedAllocator.h",
5145 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08005146 deps = MICROKERNEL_BENCHMARK_DEPS,
5147)
5148
5149xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005150 name = "f32_spmm_bench",
5151 srcs = [
5152 "bench/f32-spmm.cc",
Marat Dukhan1631e3e2020-12-06 19:29:31 -08005153 "bench/spmm.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005154 "src/xnnpack/AlignedAllocator.h",
5155 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005156 deps = MICROKERNEL_BENCHMARK_DEPS,
5157)
5158
5159xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005160 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005161 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005162 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005163 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07005164 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08005165 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07005166)
5167
5168xnnpack_benchmark(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08005169 name = "f32_velu_bench",
5170 srcs = [
5171 "bench/f32-velu.cc",
5172 "src/xnnpack/AlignedAllocator.h",
5173 ] + MICROKERNEL_BENCHMARK_HDRS,
5174 deps = MICROKERNEL_BENCHMARK_DEPS,
5175)
5176
5177xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08005178 name = "f32_vscaleexpminusmax_bench",
5179 srcs = [
5180 "bench/f32-vscaleexpminusmax.cc",
5181 "src/xnnpack/AlignedAllocator.h",
5182 ] + MICROKERNEL_BENCHMARK_HDRS,
5183 deps = MICROKERNEL_BENCHMARK_DEPS,
5184)
5185
5186xnnpack_benchmark(
5187 name = "f32_vscaleextexp_bench",
5188 srcs = [
5189 "bench/f32-vscaleextexp.cc",
5190 "src/xnnpack/AlignedAllocator.h",
5191 ] + MICROKERNEL_BENCHMARK_HDRS,
5192 deps = MICROKERNEL_BENCHMARK_DEPS,
5193)
5194
5195xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005196 name = "f32_vsqrt_bench",
5197 srcs = [
5198 "bench/f32-vsqrt.cc",
5199 "src/xnnpack/AlignedAllocator.h",
5200 ] + MICROKERNEL_BENCHMARK_HDRS,
5201 deps = MICROKERNEL_BENCHMARK_DEPS,
5202)
5203
5204xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005205 name = "f32_im2col_gemm_bench",
5206 srcs = [
5207 "bench/f32-im2col-gemm.cc",
5208 "bench/conv.h",
5209 "src/xnnpack/AlignedAllocator.h",
5210 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005211 deps = MICROKERNEL_BENCHMARK_DEPS + [
5212 ":im2col",
5213 ":packing",
5214 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005215)
5216
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005217xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005218 name = "rounding_bench",
5219 srcs = [
5220 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005221 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07005222 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005223 ] + MICROKERNEL_BENCHMARK_HDRS,
5224 deps = MICROKERNEL_BENCHMARK_DEPS,
5225)
5226
Marat Dukhan08c4a432019-10-03 09:29:21 -07005227########################### Benchmarks for operators ###########################
5228
5229xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005230 name = "average_pooling_bench",
5231 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07005232 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005233 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005234 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005235)
5236
5237xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005238 name = "bankers_rounding_bench",
5239 srcs = ["bench/bankers-rounding.cc"],
5240 copts = xnnpack_optional_tflite_copts(),
5241 tags = ["nowin32"],
5242 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5243)
5244
5245xnnpack_benchmark(
5246 name = "ceiling_bench",
5247 srcs = ["bench/ceiling.cc"],
5248 copts = xnnpack_optional_tflite_copts(),
5249 tags = ["nowin32"],
5250 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5251)
5252
5253xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005254 name = "channel_shuffle_bench",
5255 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005256 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005257)
5258
5259xnnpack_benchmark(
5260 name = "convolution_bench",
5261 srcs = ["bench/convolution.cc"],
5262 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005263 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005264 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005265)
5266
5267xnnpack_benchmark(
5268 name = "deconvolution_bench",
5269 srcs = ["bench/deconvolution.cc"],
5270 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005271 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005272 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005273)
5274
5275xnnpack_benchmark(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08005276 name = "elu_bench",
5277 srcs = ["bench/elu.cc"],
5278 copts = xnnpack_optional_tflite_copts(),
5279 tags = ["nowin32"],
5280 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5281)
5282
5283xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005284 name = "floor_bench",
5285 srcs = ["bench/floor.cc"],
5286 copts = xnnpack_optional_tflite_copts(),
5287 tags = ["nowin32"],
5288 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5289)
5290
5291xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005292 name = "global_average_pooling_bench",
5293 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005294 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005295)
5296
5297xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07005298 name = "hardswish_bench",
5299 srcs = ["bench/hardswish.cc"],
5300 copts = xnnpack_optional_tflite_copts(),
5301 tags = ["nowin32"],
5302 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5303)
5304
5305xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005306 name = "max_pooling_bench",
5307 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005308 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005309)
5310
5311xnnpack_benchmark(
5312 name = "sigmoid_bench",
5313 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08005314 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005315 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005316 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005317)
5318
5319xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07005320 name = "prelu_bench",
5321 srcs = ["bench/prelu.cc"],
5322 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005323 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005324 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07005325)
5326
5327xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005328 name = "softmax_bench",
5329 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08005330 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005331 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005332 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005333)
5334
Marat Dukhan87727142020-06-24 15:24:10 -07005335xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005336 name = "square_root_bench",
5337 srcs = ["bench/square-root.cc"],
5338 copts = xnnpack_optional_tflite_copts(),
5339 tags = ["nowin32"],
5340 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5341)
5342
5343xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005344 name = "truncation_bench",
5345 srcs = ["bench/truncation.cc"],
5346 deps = OPERATOR_BENCHMARK_DEPS,
5347)
5348
Marat Dukhanc068bb62019-10-04 13:24:39 -07005349############################# End-to-end benchmarks ############################
5350
5351cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005352 name = "fp32_mobilenet_v1",
5353 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005354 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005355 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005356 linkstatic = True,
5357 deps = [
5358 ":XNNPACK",
5359 "@pthreadpool",
5360 ],
5361)
5362
5363cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005364 name = "fp32_sparse_mobilenet_v1",
5365 srcs = ["models/fp32-sparse-mobilenet-v1.cc"],
5366 hdrs = ["models/models.h"],
5367 copts = xnnpack_std_cxxopts(),
5368 linkstatic = True,
5369 deps = [
5370 ":XNNPACK",
5371 "@pthreadpool",
5372 ],
5373)
5374
5375cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005376 name = "fp16_mobilenet_v1",
5377 srcs = ["models/fp16-mobilenet-v1.cc"],
5378 hdrs = ["models/models.h"],
5379 copts = xnnpack_std_cxxopts(),
5380 linkstatic = True,
5381 deps = [
5382 ":XNNPACK",
5383 "@FP16",
5384 "@pthreadpool",
5385 ],
5386)
5387
5388cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005389 name = "qs8_mobilenet_v1",
5390 srcs = ["models/qs8-mobilenet-v1.cc"],
5391 hdrs = ["models/models.h"],
5392 copts = xnnpack_std_cxxopts(),
5393 linkstatic = True,
5394 deps = [
5395 ":XNNPACK",
5396 "@pthreadpool",
5397 ],
5398)
5399
5400cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07005401 name = "qs8_mobilenet_v2",
5402 srcs = ["models/qs8-mobilenet-v2.cc"],
5403 hdrs = ["models/models.h"],
5404 copts = xnnpack_std_cxxopts(),
5405 linkstatic = True,
5406 deps = [
5407 ":XNNPACK",
5408 "@pthreadpool",
5409 ],
5410)
5411
5412cc_library(
Marat Dukhan12a23bb2021-03-08 08:13:21 -08005413 name = "qu8_mobilenet_v1",
5414 srcs = ["models/qu8-mobilenet-v1.cc"],
5415 hdrs = ["models/models.h"],
5416 copts = xnnpack_std_cxxopts(),
5417 linkstatic = True,
5418 deps = [
5419 ":XNNPACK",
5420 "@pthreadpool",
5421 ],
5422)
5423
5424cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005425 name = "fp32_mobilenet_v2",
5426 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005427 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005428 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005429 linkstatic = True,
5430 deps = [
5431 ":XNNPACK",
5432 "@pthreadpool",
5433 ],
5434)
5435
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005436cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005437 name = "fp32_sparse_mobilenet_v2",
5438 srcs = ["models/fp32-sparse-mobilenet-v2.cc"],
5439 hdrs = ["models/models.h"],
5440 copts = xnnpack_std_cxxopts(),
5441 linkstatic = True,
5442 deps = [
5443 ":XNNPACK",
5444 "@pthreadpool",
5445 ],
5446)
5447
5448cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005449 name = "fp16_mobilenet_v2",
5450 srcs = ["models/fp16-mobilenet-v2.cc"],
5451 hdrs = ["models/models.h"],
5452 copts = xnnpack_std_cxxopts(),
5453 linkstatic = True,
5454 deps = [
5455 ":XNNPACK",
5456 "@FP16",
5457 "@pthreadpool",
5458 ],
5459)
5460
5461cc_library(
5462 name = "fp32_mobilenet_v3_large",
5463 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005464 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005465 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005466 linkstatic = True,
5467 deps = [
5468 ":XNNPACK",
5469 "@pthreadpool",
5470 ],
5471)
5472
5473cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005474 name = "fp32_sparse_mobilenet_v3_large",
5475 srcs = ["models/fp32-sparse-mobilenet-v3-large.cc"],
5476 hdrs = ["models/models.h"],
5477 copts = xnnpack_std_cxxopts(),
5478 linkstatic = True,
5479 deps = [
5480 ":XNNPACK",
5481 "@pthreadpool",
5482 ],
5483)
5484
5485cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005486 name = "fp16_mobilenet_v3_large",
5487 srcs = ["models/fp16-mobilenet-v3-large.cc"],
5488 hdrs = ["models/models.h"],
5489 copts = xnnpack_std_cxxopts(),
5490 linkstatic = True,
5491 deps = [
5492 ":XNNPACK",
5493 "@FP16",
5494 "@pthreadpool",
5495 ],
5496)
5497
5498cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005499 name = "fp32_mobilenet_v3_small",
5500 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005501 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005502 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005503 linkstatic = True,
5504 deps = [
5505 ":XNNPACK",
5506 "@pthreadpool",
5507 ],
5508)
5509
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005510cc_library(
Marat Dukhan4cea2322021-03-09 09:35:36 -08005511 name = "fp32_sparse_mobilenet_v3_small",
5512 srcs = ["models/fp32-sparse-mobilenet-v3-small.cc"],
5513 hdrs = ["models/models.h"],
5514 copts = xnnpack_std_cxxopts(),
5515 linkstatic = True,
5516 deps = [
5517 ":XNNPACK",
5518 "@pthreadpool",
5519 ],
5520)
5521
5522cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005523 name = "fp16_mobilenet_v3_small",
5524 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5525 hdrs = ["models/models.h"],
5526 copts = xnnpack_std_cxxopts(),
5527 linkstatic = True,
5528 deps = [
5529 ":XNNPACK",
5530 "@FP16",
5531 "@pthreadpool",
5532 ],
5533)
5534
Marat Dukhanc068bb62019-10-04 13:24:39 -07005535xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005536 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005537 srcs = [
5538 "bench/f32-dwconv-e2e.cc",
5539 "bench/end2end.h",
5540 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005541 deps = MICROKERNEL_BENCHMARK_DEPS + [
5542 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005543 ":fp32_mobilenet_v1",
5544 ":fp32_mobilenet_v2",
5545 ":fp32_mobilenet_v3_large",
5546 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005547 ],
5548)
5549
5550xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005551 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005552 srcs = [
5553 "bench/f32-gemm-e2e.cc",
5554 "bench/end2end.h",
5555 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005556 deps = MICROKERNEL_BENCHMARK_DEPS + [
5557 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005558 ":fp32_mobilenet_v1",
5559 ":fp32_mobilenet_v2",
5560 ":fp32_mobilenet_v3_large",
5561 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005562 ],
5563)
5564
5565xnnpack_benchmark(
Frank Barcharddc909cb2021-02-08 13:59:31 -08005566 name = "qs8_gemm_e2e_bench",
5567 srcs = [
5568 "bench/qs8-gemm-e2e.cc",
5569 "bench/end2end.h",
5570 ] + MICROKERNEL_BENCHMARK_HDRS,
5571 deps = MICROKERNEL_BENCHMARK_DEPS + [
5572 ":XNNPACK",
5573 ":qs8_mobilenet_v1",
5574 ":qs8_mobilenet_v2",
5575 ],
5576)
5577
5578xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005579 name = "end2end_bench",
5580 srcs = ["bench/end2end.cc"],
5581 deps = [
5582 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005583 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005584 ":fp16_mobilenet_v1",
5585 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005586 ":fp16_mobilenet_v3_large",
5587 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005588 ":fp32_mobilenet_v1",
5589 ":fp32_mobilenet_v2",
5590 ":fp32_mobilenet_v3_large",
5591 ":fp32_mobilenet_v3_small",
Marat Dukhan4cea2322021-03-09 09:35:36 -08005592 ":fp32_sparse_mobilenet_v1",
5593 ":fp32_sparse_mobilenet_v2",
5594 ":fp32_sparse_mobilenet_v3_large",
5595 ":fp32_sparse_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005596 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005597 ":qs8_mobilenet_v2",
Marat Dukhan12a23bb2021-03-08 08:13:21 -08005598 ":qu8_mobilenet_v1",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005599 "@pthreadpool",
5600 ],
5601)
5602
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005603#################### Accuracy evaluation for math functions ####################
5604
5605xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005606 name = "f32_exp_ulp_eval",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005607 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005608 "eval/f32-exp-ulp.cc",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005609 "src/xnnpack/AlignedAllocator.h",
5610 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005611 deps = ACCURACY_EVAL_DEPS + [
5612 ":bench_utils",
5613 "@cpuinfo",
5614 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005615)
5616
Marat Dukhan515c9772019-10-17 18:07:57 -07005617xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005618 name = "f32_expminus_ulp_eval",
Marat Dukhan515c9772019-10-17 18:07:57 -07005619 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005620 "eval/f32-expminus-ulp.cc",
Marat Dukhan515c9772019-10-17 18:07:57 -07005621 "src/xnnpack/AlignedAllocator.h",
5622 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005623 deps = ACCURACY_EVAL_DEPS + [
5624 ":bench_utils",
5625 "@cpuinfo",
5626 ],
Marat Dukhan515c9772019-10-17 18:07:57 -07005627)
5628
Marat Dukhan98ba4412019-10-23 02:14:28 -07005629xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005630 name = "f32_expm1minus_ulp_eval",
Marat Dukhana438aca2020-11-20 15:45:01 -08005631 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005632 "eval/f32-expm1minus-ulp.cc",
Marat Dukhana438aca2020-11-20 15:45:01 -08005633 "src/xnnpack/AlignedAllocator.h",
5634 ] + ACCURACY_EVAL_HDRS,
Marat Dukhande390d42020-11-29 19:32:18 -08005635 deps = ACCURACY_EVAL_DEPS + [
5636 ":bench_utils",
Marat Dukhan3a305212020-12-06 19:24:27 -08005637 "@cpuinfo",
Marat Dukhande390d42020-11-29 19:32:18 -08005638 ],
Marat Dukhana438aca2020-11-20 15:45:01 -08005639)
5640
5641xnnpack_benchmark(
Marat Dukhancf67ec62020-12-14 09:56:37 -08005642 name = "f32_extexp_ulp_eval",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005643 srcs = [
Marat Dukhancf67ec62020-12-14 09:56:37 -08005644 "eval/f32-extexp-ulp.cc",
Marat Dukhan98ba4412019-10-23 02:14:28 -07005645 "src/xnnpack/AlignedAllocator.h",
5646 ] + ACCURACY_EVAL_HDRS,
Marat Dukhan3a305212020-12-06 19:24:27 -08005647 deps = ACCURACY_EVAL_DEPS + [
5648 ":bench_utils",
5649 "@cpuinfo",
5650 ],
Marat Dukhan98ba4412019-10-23 02:14:28 -07005651)
5652
Marat Dukhanf44f0222020-12-14 11:53:27 -08005653xnnpack_benchmark(
5654 name = "f32_sigmoid_ulp_eval",
5655 srcs = [
5656 "eval/f32-sigmoid-ulp.cc",
5657 "src/xnnpack/AlignedAllocator.h",
5658 ] + ACCURACY_EVAL_HDRS,
5659 deps = ACCURACY_EVAL_DEPS + [
5660 ":bench_utils",
5661 "@cpuinfo",
5662 ],
5663)
5664
5665xnnpack_benchmark(
5666 name = "f32_sqrt_ulp_eval",
5667 srcs = [
5668 "eval/f32-sqrt-ulp.cc",
5669 "src/xnnpack/AlignedAllocator.h",
5670 ] + ACCURACY_EVAL_HDRS,
5671 deps = ACCURACY_EVAL_DEPS + [
5672 ":bench_utils",
5673 "@cpuinfo",
5674 ],
5675)
5676
5677################### Accuracy verification for math functions ##################
5678
5679xnnpack_unit_test(
Marat Dukhanf7291fc2020-12-15 11:02:50 -08005680 name = "f32_exp_eval",
5681 srcs = [
5682 "eval/f32-exp.cc",
5683 "src/xnnpack/AlignedAllocator.h",
5684 "src/xnnpack/math-stubs.h",
5685 ] + MICROKERNEL_TEST_HDRS,
5686 automatic = False,
5687 deps = MICROKERNEL_TEST_DEPS,
5688)
5689
5690xnnpack_unit_test(
Marat Dukhanf44f0222020-12-14 11:53:27 -08005691 name = "f32_expm1minus_eval",
5692 srcs = [
5693 "eval/f32-expm1minus.cc",
5694 "src/xnnpack/AlignedAllocator.h",
5695 "src/xnnpack/math-stubs.h",
5696 ] + MICROKERNEL_TEST_HDRS,
5697 automatic = False,
5698 deps = MICROKERNEL_TEST_DEPS,
5699)
5700
Marat Dukhan8853b822020-05-07 12:19:01 -07005701xnnpack_unit_test(
Marat Dukhand28a5a22020-12-14 15:27:22 -08005702 name = "f32_expminus_eval",
5703 srcs = [
5704 "eval/f32-expminus.cc",
5705 "src/xnnpack/AlignedAllocator.h",
5706 "src/xnnpack/math-stubs.h",
5707 ] + MICROKERNEL_TEST_HDRS,
5708 automatic = False,
5709 deps = MICROKERNEL_TEST_DEPS,
5710)
5711
5712xnnpack_unit_test(
Marat Dukhan8853b822020-05-07 12:19:01 -07005713 name = "f32_roundne_eval",
5714 srcs = [
5715 "eval/f32-roundne.cc",
5716 "src/xnnpack/AlignedAllocator.h",
5717 "src/xnnpack/math-stubs.h",
5718 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005719 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005720 deps = MICROKERNEL_TEST_DEPS,
5721)
5722
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005723xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005724 name = "f32_roundd_eval",
5725 srcs = [
5726 "eval/f32-roundd.cc",
5727 "src/xnnpack/AlignedAllocator.h",
5728 "src/xnnpack/math-stubs.h",
5729 ] + MICROKERNEL_TEST_HDRS,
5730 automatic = False,
5731 deps = MICROKERNEL_TEST_DEPS,
5732)
5733
5734xnnpack_unit_test(
5735 name = "f32_roundu_eval",
5736 srcs = [
5737 "eval/f32-roundu.cc",
5738 "src/xnnpack/AlignedAllocator.h",
5739 "src/xnnpack/math-stubs.h",
5740 ] + MICROKERNEL_TEST_HDRS,
5741 automatic = False,
5742 deps = MICROKERNEL_TEST_DEPS,
5743)
5744
5745xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005746 name = "f32_roundz_eval",
5747 srcs = [
5748 "eval/f32-roundz.cc",
5749 "src/xnnpack/AlignedAllocator.h",
5750 "src/xnnpack/math-stubs.h",
5751 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005752 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005753 deps = MICROKERNEL_TEST_DEPS,
5754)
5755
Marat Dukhan08c4a432019-10-03 09:29:21 -07005756######################### Unit tests for micro-kernels #########################
5757
5758xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005759 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005760 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005761 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005762 "test/gemm-microkernel-tester.h",
5763 "src/xnnpack/AlignedAllocator.h",
5764 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005765 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005766)
5767
5768xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005769 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005770 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005771 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005772 "test/spmm-microkernel-tester.h",
5773 "src/xnnpack/AlignedAllocator.h",
5774 ] + MICROKERNEL_TEST_HDRS,
5775 deps = MICROKERNEL_TEST_DEPS,
5776)
5777
5778xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005779 name = "f32_argmaxpool_test",
5780 srcs = [
5781 "test/f32-argmaxpool.cc",
5782 "test/argmaxpool-microkernel-tester.h",
5783 "src/xnnpack/AlignedAllocator.h",
5784 ] + MICROKERNEL_TEST_HDRS,
5785 deps = MICROKERNEL_TEST_DEPS,
5786)
5787
5788xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005789 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005790 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005791 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005792 "test/avgpool-microkernel-tester.h",
5793 "src/xnnpack/AlignedAllocator.h",
5794 ] + MICROKERNEL_TEST_HDRS,
5795 deps = MICROKERNEL_TEST_DEPS,
5796)
5797
5798xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005799 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005800 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005801 "test/f32-ibilinear.cc",
5802 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005803 "src/xnnpack/AlignedAllocator.h",
5804 ] + MICROKERNEL_TEST_HDRS,
5805 deps = MICROKERNEL_TEST_DEPS,
5806)
5807
5808xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005809 name = "f32_ibilinear_chw_test",
5810 srcs = [
5811 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005812 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005813 "src/xnnpack/AlignedAllocator.h",
5814 ] + MICROKERNEL_TEST_HDRS,
5815 deps = MICROKERNEL_TEST_DEPS,
5816)
5817
5818xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005819 name = "f16_clamp_test",
5820 srcs = [
5821 "test/f16-clamp.cc",
5822 "test/clamp-microkernel-tester.h",
5823 ] + MICROKERNEL_TEST_HDRS,
5824 deps = MICROKERNEL_TEST_DEPS,
5825)
5826
5827xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005828 name = "f32_clamp_test",
5829 srcs = [
5830 "test/f32-clamp.cc",
5831 "test/clamp-microkernel-tester.h",
5832 ] + MICROKERNEL_TEST_HDRS,
5833 deps = MICROKERNEL_TEST_DEPS,
5834)
5835
5836xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005837 name = "f32_igemm_test",
5838 srcs = [
5839 "test/f32-igemm.cc",
5840 "test/gemm-microkernel-tester.h",
5841 "src/xnnpack/AlignedAllocator.h",
5842 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005843 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005844)
5845
5846xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005847 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005848 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005849 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005850 "test/gemm-microkernel-tester.h",
5851 "src/xnnpack/AlignedAllocator.h",
5852 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005853 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005854)
5855
5856xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005857 name = "f32_igemm_minmax_test",
5858 srcs = [
5859 "test/f32-igemm-minmax.cc",
5860 "test/gemm-microkernel-tester.h",
5861 "src/xnnpack/AlignedAllocator.h",
5862 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005863 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005864)
5865
5866xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005867 name = "f16_igemm_minmax_test",
5868 srcs = [
5869 "test/f16-igemm-minmax.cc",
5870 "test/gemm-microkernel-tester.h",
5871 "src/xnnpack/AlignedAllocator.h",
5872 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005873 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005874)
5875
5876xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005877 name = "f32_conv_hwc_test",
5878 srcs = [
5879 "test/f32-conv-hwc.cc",
5880 "test/conv-hwc-microkernel-tester.h",
5881 "src/xnnpack/AlignedAllocator.h",
5882 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005883 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005884)
5885
5886xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005887 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005888 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005889 "test/f32-conv-hwc2chw.cc",
5890 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005891 "src/xnnpack/AlignedAllocator.h",
5892 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005893 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005894)
5895
5896xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005897 name = "f32_dwconv_test",
5898 srcs = [
5899 "test/f32-dwconv.cc",
5900 "test/dwconv-microkernel-tester.h",
5901 "src/xnnpack/AlignedAllocator.h",
5902 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005903 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005904)
5905
5906xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005907 name = "f16_dwconv_minmax_test",
5908 srcs = [
5909 "test/f16-dwconv-minmax.cc",
5910 "test/dwconv-microkernel-tester.h",
5911 "src/xnnpack/AlignedAllocator.h",
5912 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005913 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005914)
5915
5916xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005917 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005918 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005919 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005920 "test/dwconv-microkernel-tester.h",
5921 "src/xnnpack/AlignedAllocator.h",
5922 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005923 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005924)
5925
5926xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005927 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005928 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005929 "test/f32-dwconv2d-chw.cc",
5930 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005931 "src/xnnpack/AlignedAllocator.h",
5932 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005933 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005934)
5935
5936xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005937 name = "f16_gavgpool_minmax_test",
5938 srcs = [
5939 "test/f16-gavgpool-minmax.cc",
5940 "test/gavgpool-microkernel-tester.h",
5941 "src/xnnpack/AlignedAllocator.h",
5942 ] + MICROKERNEL_TEST_HDRS,
5943 deps = MICROKERNEL_TEST_DEPS,
5944)
5945
5946xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005947 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005948 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005949 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005950 "test/gavgpool-microkernel-tester.h",
5951 "src/xnnpack/AlignedAllocator.h",
5952 ] + MICROKERNEL_TEST_HDRS,
5953 deps = MICROKERNEL_TEST_DEPS,
5954)
5955
5956xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005957 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005958 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005959 "test/f32-gavgpool-cw.cc",
5960 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005961 "src/xnnpack/AlignedAllocator.h",
5962 ] + MICROKERNEL_TEST_HDRS,
5963 deps = MICROKERNEL_TEST_DEPS,
5964)
5965
5966xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005967 name = "f32_gemm_test",
5968 srcs = [
5969 "test/f32-gemm.cc",
5970 "test/gemm-microkernel-tester.h",
5971 "src/xnnpack/AlignedAllocator.h",
5972 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005973 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005974)
5975
5976xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005977 name = "f32_gemm_relu_test",
5978 srcs = [
5979 "test/f32-gemm-relu.cc",
5980 "test/gemm-microkernel-tester.h",
5981 "src/xnnpack/AlignedAllocator.h",
5982 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005983 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005984)
5985
5986xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005987 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005988 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005989 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005990 "test/gemm-microkernel-tester.h",
5991 "src/xnnpack/AlignedAllocator.h",
5992 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005993 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005994)
5995
5996xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005997 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005998 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005999 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006000 "test/gemm-microkernel-tester.h",
6001 "src/xnnpack/AlignedAllocator.h",
6002 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006003 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006004)
6005
6006xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07006007 name = "f16_hswish_test",
6008 srcs = [
6009 "test/f16-hswish.cc",
6010 "test/hswish-microkernel-tester.h",
6011 ] + MICROKERNEL_TEST_HDRS,
6012 deps = MICROKERNEL_TEST_DEPS,
6013)
6014
6015xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006016 name = "f32_hswish_test",
6017 srcs = [
6018 "test/f32-hswish.cc",
6019 "test/hswish-microkernel-tester.h",
6020 ] + MICROKERNEL_TEST_HDRS,
6021 deps = MICROKERNEL_TEST_DEPS,
6022)
6023
6024xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006025 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006026 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006027 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006028 "test/maxpool-microkernel-tester.h",
6029 ] + MICROKERNEL_TEST_HDRS,
6030 deps = MICROKERNEL_TEST_DEPS,
6031)
6032
6033xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006034 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006035 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006036 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006037 "test/avgpool-microkernel-tester.h",
6038 "src/xnnpack/AlignedAllocator.h",
6039 ] + MICROKERNEL_TEST_HDRS,
6040 deps = MICROKERNEL_TEST_DEPS,
6041)
6042
6043xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07006044 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006045 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07006046 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006047 "test/gemm-microkernel-tester.h",
6048 "src/xnnpack/AlignedAllocator.h",
6049 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006050 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006051)
6052
6053xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07006054 name = "f16_prelu_test",
6055 srcs = [
6056 "test/f16-prelu.cc",
6057 "test/prelu-microkernel-tester.h",
6058 "src/xnnpack/AlignedAllocator.h",
6059 ] + MICROKERNEL_TEST_HDRS,
6060 deps = MICROKERNEL_TEST_DEPS,
6061)
6062
6063xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006064 name = "f32_prelu_test",
6065 srcs = [
6066 "test/f32-prelu.cc",
6067 "test/prelu-microkernel-tester.h",
6068 "src/xnnpack/AlignedAllocator.h",
6069 ] + MICROKERNEL_TEST_HDRS,
6070 deps = MICROKERNEL_TEST_DEPS,
6071)
6072
6073xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006074 name = "f32_raddexpminusmax_test",
6075 srcs = [
6076 "test/f32-raddexpminusmax.cc",
6077 "test/raddexpminusmax-microkernel-tester.h",
6078 ] + MICROKERNEL_TEST_HDRS,
6079 deps = MICROKERNEL_TEST_DEPS,
6080)
6081
6082xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006083 name = "f32_raddextexp_test",
6084 srcs = [
6085 "test/f32-raddextexp.cc",
6086 "test/raddextexp-microkernel-tester.h",
6087 ] + MICROKERNEL_TEST_HDRS,
6088 deps = MICROKERNEL_TEST_DEPS,
6089)
6090
6091xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006092 name = "f32_raddstoreexpminusmax_test",
6093 srcs = [
6094 "test/f32-raddstoreexpminusmax.cc",
6095 "test/raddstoreexpminusmax-microkernel-tester.h",
6096 ] + MICROKERNEL_TEST_HDRS,
6097 deps = MICROKERNEL_TEST_DEPS,
6098)
6099
6100xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07006101 name = "f32_relu_test",
6102 srcs = [
6103 "test/f32-relu.cc",
6104 "test/vunary-microkernel-tester.h",
6105 ] + MICROKERNEL_TEST_HDRS,
6106 deps = MICROKERNEL_TEST_DEPS,
6107)
6108
6109xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006110 name = "f32_rmax_test",
6111 srcs = [
6112 "test/f32-rmax.cc",
6113 "test/rmax-microkernel-tester.h",
6114 ] + MICROKERNEL_TEST_HDRS,
6115 deps = MICROKERNEL_TEST_DEPS,
6116)
6117
6118xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08006119 name = "f32_sigmoid_test",
6120 srcs = [
6121 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006122 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08006123 ] + MICROKERNEL_TEST_HDRS,
6124 deps = MICROKERNEL_TEST_DEPS,
6125)
6126
6127xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07006128 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006129 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07006130 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006131 "test/spmm-microkernel-tester.h",
6132 "src/xnnpack/AlignedAllocator.h",
6133 ] + MICROKERNEL_TEST_HDRS,
6134 deps = MICROKERNEL_TEST_DEPS,
6135)
6136
6137xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07006138 name = "f16_vadd_minmax_test",
6139 srcs = [
6140 "test/f16-vadd-minmax.cc",
6141 "test/vbinary-microkernel-tester.h",
6142 ] + MICROKERNEL_TEST_HDRS,
6143 deps = MICROKERNEL_TEST_DEPS,
6144)
6145
6146xnnpack_unit_test(
6147 name = "f16_vaddc_minmax_test",
6148 srcs = [
6149 "test/f16-vaddc-minmax.cc",
6150 "test/vbinaryc-microkernel-tester.h",
6151 ] + MICROKERNEL_TEST_HDRS,
6152 deps = MICROKERNEL_TEST_DEPS,
6153)
6154
6155xnnpack_unit_test(
6156 name = "f16_vdiv_minmax_test",
6157 srcs = [
6158 "test/f16-vdiv-minmax.cc",
6159 "test/vbinary-microkernel-tester.h",
6160 ] + MICROKERNEL_TEST_HDRS,
6161 deps = MICROKERNEL_TEST_DEPS,
6162)
6163
6164xnnpack_unit_test(
6165 name = "f16_vdivc_minmax_test",
6166 srcs = [
6167 "test/f16-vdivc-minmax.cc",
6168 "test/vbinaryc-microkernel-tester.h",
6169 ] + MICROKERNEL_TEST_HDRS,
6170 deps = MICROKERNEL_TEST_DEPS,
6171)
6172
6173xnnpack_unit_test(
6174 name = "f16_vrdivc_minmax_test",
6175 srcs = [
6176 "test/f16-vrdivc-minmax.cc",
6177 "test/vbinaryc-microkernel-tester.h",
6178 ] + MICROKERNEL_TEST_HDRS,
6179 deps = MICROKERNEL_TEST_DEPS,
6180)
6181
6182xnnpack_unit_test(
6183 name = "f16_vmax_test",
6184 srcs = [
6185 "test/f16-vmax.cc",
6186 "test/vbinary-microkernel-tester.h",
6187 ] + MICROKERNEL_TEST_HDRS,
6188 deps = MICROKERNEL_TEST_DEPS,
6189)
6190
6191xnnpack_unit_test(
6192 name = "f16_vmaxc_test",
6193 srcs = [
6194 "test/f16-vmaxc.cc",
6195 "test/vbinaryc-microkernel-tester.h",
6196 ] + MICROKERNEL_TEST_HDRS,
6197 deps = MICROKERNEL_TEST_DEPS,
6198)
6199
6200xnnpack_unit_test(
6201 name = "f16_vmin_test",
6202 srcs = [
6203 "test/f16-vmin.cc",
6204 "test/vbinary-microkernel-tester.h",
6205 ] + MICROKERNEL_TEST_HDRS,
6206 deps = MICROKERNEL_TEST_DEPS,
6207)
6208
6209xnnpack_unit_test(
6210 name = "f16_vminc_test",
6211 srcs = [
6212 "test/f16-vminc.cc",
6213 "test/vbinaryc-microkernel-tester.h",
6214 ] + MICROKERNEL_TEST_HDRS,
6215 deps = MICROKERNEL_TEST_DEPS,
6216)
6217
6218xnnpack_unit_test(
6219 name = "f16_vmul_minmax_test",
6220 srcs = [
6221 "test/f16-vmul-minmax.cc",
6222 "test/vbinary-microkernel-tester.h",
6223 ] + MICROKERNEL_TEST_HDRS,
6224 deps = MICROKERNEL_TEST_DEPS,
6225)
6226
6227xnnpack_unit_test(
6228 name = "f16_vmulc_minmax_test",
6229 srcs = [
6230 "test/f16-vmulc-minmax.cc",
6231 "test/vbinaryc-microkernel-tester.h",
6232 ] + MICROKERNEL_TEST_HDRS,
6233 deps = MICROKERNEL_TEST_DEPS,
6234)
6235
6236xnnpack_unit_test(
6237 name = "f16_vsub_minmax_test",
6238 srcs = [
6239 "test/f16-vsub-minmax.cc",
6240 "test/vbinary-microkernel-tester.h",
6241 ] + MICROKERNEL_TEST_HDRS,
6242 deps = MICROKERNEL_TEST_DEPS,
6243)
6244
6245xnnpack_unit_test(
6246 name = "f16_vsubc_minmax_test",
6247 srcs = [
6248 "test/f16-vsubc-minmax.cc",
6249 "test/vbinaryc-microkernel-tester.h",
6250 ] + MICROKERNEL_TEST_HDRS,
6251 deps = MICROKERNEL_TEST_DEPS,
6252)
6253
6254xnnpack_unit_test(
6255 name = "f16_vrsubc_minmax_test",
6256 srcs = [
6257 "test/f16-vrsubc-minmax.cc",
6258 "test/vbinaryc-microkernel-tester.h",
6259 ] + MICROKERNEL_TEST_HDRS,
6260 deps = MICROKERNEL_TEST_DEPS,
6261)
6262
6263xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006264 name = "f32_vabs_test",
6265 srcs = [
6266 "test/f32-vabs.cc",
6267 "test/vunary-microkernel-tester.h",
6268 ] + MICROKERNEL_TEST_HDRS,
6269 deps = MICROKERNEL_TEST_DEPS,
6270)
6271
6272xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006273 name = "f32_vadd_test",
6274 srcs = [
6275 "test/f32-vadd.cc",
6276 "test/vbinary-microkernel-tester.h",
6277 ] + MICROKERNEL_TEST_HDRS,
6278 deps = MICROKERNEL_TEST_DEPS,
6279)
6280
6281xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006282 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006283 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006284 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006285 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006286 ] + MICROKERNEL_TEST_HDRS,
6287 deps = MICROKERNEL_TEST_DEPS,
6288)
6289
6290xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006291 name = "f32_vadd_relu_test",
6292 srcs = [
6293 "test/f32-vadd-relu.cc",
6294 "test/vbinary-microkernel-tester.h",
6295 ] + MICROKERNEL_TEST_HDRS,
6296 deps = MICROKERNEL_TEST_DEPS,
6297)
6298
6299xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006300 name = "f32_vaddc_test",
6301 srcs = [
6302 "test/f32-vaddc.cc",
6303 "test/vbinaryc-microkernel-tester.h",
6304 ] + MICROKERNEL_TEST_HDRS,
6305 deps = MICROKERNEL_TEST_DEPS,
6306)
6307
6308xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006309 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006310 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006311 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006312 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006313 ] + MICROKERNEL_TEST_HDRS,
6314 deps = MICROKERNEL_TEST_DEPS,
6315)
6316
6317xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006318 name = "f32_vaddc_relu_test",
6319 srcs = [
6320 "test/f32-vaddc-relu.cc",
6321 "test/vbinaryc-microkernel-tester.h",
6322 ] + MICROKERNEL_TEST_HDRS,
6323 deps = MICROKERNEL_TEST_DEPS,
6324)
6325
6326xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006327 name = "f32_vdiv_test",
6328 srcs = [
6329 "test/f32-vdiv.cc",
6330 "test/vbinary-microkernel-tester.h",
6331 ] + MICROKERNEL_TEST_HDRS,
6332 deps = MICROKERNEL_TEST_DEPS,
6333)
6334
6335xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006336 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006337 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006338 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006339 "test/vbinary-microkernel-tester.h",
6340 ] + MICROKERNEL_TEST_HDRS,
6341 deps = MICROKERNEL_TEST_DEPS,
6342)
6343
6344xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006345 name = "f32_vdiv_relu_test",
6346 srcs = [
6347 "test/f32-vdiv-relu.cc",
6348 "test/vbinary-microkernel-tester.h",
6349 ] + MICROKERNEL_TEST_HDRS,
6350 deps = MICROKERNEL_TEST_DEPS,
6351)
6352
6353xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006354 name = "f32_vdivc_test",
6355 srcs = [
6356 "test/f32-vdivc.cc",
6357 "test/vbinaryc-microkernel-tester.h",
6358 ] + MICROKERNEL_TEST_HDRS,
6359 deps = MICROKERNEL_TEST_DEPS,
6360)
6361
6362xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006363 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006364 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006365 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006366 "test/vbinaryc-microkernel-tester.h",
6367 ] + MICROKERNEL_TEST_HDRS,
6368 deps = MICROKERNEL_TEST_DEPS,
6369)
6370
6371xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006372 name = "f32_vdivc_relu_test",
6373 srcs = [
6374 "test/f32-vdivc-relu.cc",
6375 "test/vbinaryc-microkernel-tester.h",
6376 ] + MICROKERNEL_TEST_HDRS,
6377 deps = MICROKERNEL_TEST_DEPS,
6378)
6379
6380xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006381 name = "f32_vrdivc_test",
6382 srcs = [
6383 "test/f32-vrdivc.cc",
6384 "test/vbinaryc-microkernel-tester.h",
6385 ] + MICROKERNEL_TEST_HDRS,
6386 deps = MICROKERNEL_TEST_DEPS,
6387)
6388
6389xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006390 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006391 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006392 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006393 "test/vbinaryc-microkernel-tester.h",
6394 ] + MICROKERNEL_TEST_HDRS,
6395 deps = MICROKERNEL_TEST_DEPS,
6396)
6397
6398xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006399 name = "f32_vrdivc_relu_test",
6400 srcs = [
6401 "test/f32-vrdivc-relu.cc",
6402 "test/vbinaryc-microkernel-tester.h",
6403 ] + MICROKERNEL_TEST_HDRS,
6404 deps = MICROKERNEL_TEST_DEPS,
6405)
6406
6407xnnpack_unit_test(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08006408 name = "f32_velu_test",
6409 srcs = [
6410 "test/f32-velu.cc",
6411 "test/vunary-microkernel-tester.h",
6412 ] + MICROKERNEL_TEST_HDRS,
6413 deps = MICROKERNEL_TEST_DEPS,
6414)
6415
6416xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08006417 name = "f32_vmax_test",
6418 srcs = [
6419 "test/f32-vmax.cc",
6420 "test/vbinary-microkernel-tester.h",
6421 ] + MICROKERNEL_TEST_HDRS,
6422 deps = MICROKERNEL_TEST_DEPS,
6423)
6424
6425xnnpack_unit_test(
6426 name = "f32_vmaxc_test",
6427 srcs = [
6428 "test/f32-vmaxc.cc",
6429 "test/vbinaryc-microkernel-tester.h",
6430 ] + MICROKERNEL_TEST_HDRS,
6431 deps = MICROKERNEL_TEST_DEPS,
6432)
6433
6434xnnpack_unit_test(
6435 name = "f32_vmin_test",
6436 srcs = [
6437 "test/f32-vmin.cc",
6438 "test/vbinary-microkernel-tester.h",
6439 ] + MICROKERNEL_TEST_HDRS,
6440 deps = MICROKERNEL_TEST_DEPS,
6441)
6442
6443xnnpack_unit_test(
6444 name = "f32_vminc_test",
6445 srcs = [
6446 "test/f32-vminc.cc",
6447 "test/vbinaryc-microkernel-tester.h",
6448 ] + MICROKERNEL_TEST_HDRS,
6449 deps = MICROKERNEL_TEST_DEPS,
6450)
6451
6452xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006453 name = "f32_vmul_test",
6454 srcs = [
6455 "test/f32-vmul.cc",
6456 "test/vbinary-microkernel-tester.h",
6457 ] + MICROKERNEL_TEST_HDRS,
6458 deps = MICROKERNEL_TEST_DEPS,
6459)
6460
6461xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006462 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006463 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006464 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006465 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006466 ] + MICROKERNEL_TEST_HDRS,
6467 deps = MICROKERNEL_TEST_DEPS,
6468)
6469
6470xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006471 name = "f32_vmul_relu_test",
6472 srcs = [
6473 "test/f32-vmul-relu.cc",
6474 "test/vbinary-microkernel-tester.h",
6475 ] + MICROKERNEL_TEST_HDRS,
6476 deps = MICROKERNEL_TEST_DEPS,
6477)
6478
6479xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006480 name = "f32_vmulc_test",
6481 srcs = [
6482 "test/f32-vmulc.cc",
6483 "test/vbinaryc-microkernel-tester.h",
6484 ] + MICROKERNEL_TEST_HDRS,
6485 deps = MICROKERNEL_TEST_DEPS,
6486)
6487
6488xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006489 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006490 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006491 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006492 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006493 ] + MICROKERNEL_TEST_HDRS,
6494 deps = MICROKERNEL_TEST_DEPS,
6495)
6496
6497xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006498 name = "f32_vmulc_relu_test",
6499 srcs = [
6500 "test/f32-vmulc-relu.cc",
6501 "test/vbinaryc-microkernel-tester.h",
6502 ] + MICROKERNEL_TEST_HDRS,
6503 deps = MICROKERNEL_TEST_DEPS,
6504)
6505
6506xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07006507 name = "f16_vmulcaddc_minmax_test",
6508 srcs = [
6509 "test/f16-vmulcaddc-minmax.cc",
6510 "test/vmulcaddc-microkernel-tester.h",
6511 "src/xnnpack/AlignedAllocator.h",
6512 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006513 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07006514)
6515
6516xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006517 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006518 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006519 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006520 "test/vmulcaddc-microkernel-tester.h",
6521 "src/xnnpack/AlignedAllocator.h",
6522 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006523 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006524)
6525
6526xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07006527 name = "f32_vlrelu_test",
6528 srcs = [
6529 "test/f32-vlrelu.cc",
6530 "test/vunary-microkernel-tester.h",
6531 ] + MICROKERNEL_TEST_HDRS,
6532 deps = MICROKERNEL_TEST_DEPS,
6533)
6534
6535xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006536 name = "f32_vneg_test",
6537 srcs = [
6538 "test/f32-vneg.cc",
6539 "test/vunary-microkernel-tester.h",
6540 ] + MICROKERNEL_TEST_HDRS,
6541 deps = MICROKERNEL_TEST_DEPS,
6542)
6543
6544xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07006545 name = "f32_vrndne_test",
6546 srcs = [
6547 "test/f32-vrndne.cc",
6548 "test/vunary-microkernel-tester.h",
6549 ] + MICROKERNEL_TEST_HDRS,
6550 deps = MICROKERNEL_TEST_DEPS,
6551)
6552
6553xnnpack_unit_test(
6554 name = "f32_vrndz_test",
6555 srcs = [
6556 "test/f32-vrndz.cc",
6557 "test/vunary-microkernel-tester.h",
6558 ] + MICROKERNEL_TEST_HDRS,
6559 deps = MICROKERNEL_TEST_DEPS,
6560)
6561
6562xnnpack_unit_test(
6563 name = "f32_vrndu_test",
6564 srcs = [
6565 "test/f32-vrndu.cc",
6566 "test/vunary-microkernel-tester.h",
6567 ] + MICROKERNEL_TEST_HDRS,
6568 deps = MICROKERNEL_TEST_DEPS,
6569)
6570
6571xnnpack_unit_test(
6572 name = "f32_vrndd_test",
6573 srcs = [
6574 "test/f32-vrndd.cc",
6575 "test/vunary-microkernel-tester.h",
6576 ] + MICROKERNEL_TEST_HDRS,
6577 deps = MICROKERNEL_TEST_DEPS,
6578)
6579
6580xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07006581 name = "f32_vscale_test",
6582 srcs = [
6583 "test/f32-vscale.cc",
6584 "test/vscale-microkernel-tester.h",
6585 ] + MICROKERNEL_TEST_HDRS,
6586 deps = MICROKERNEL_TEST_DEPS,
6587)
6588
6589xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006590 name = "f32_vscaleexpminusmax_test",
6591 srcs = [
6592 "test/f32-vscaleexpminusmax.cc",
6593 "test/vscaleexpminusmax-microkernel-tester.h",
6594 ] + MICROKERNEL_TEST_HDRS,
6595 deps = MICROKERNEL_TEST_DEPS,
6596)
6597
6598xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006599 name = "f32_vscaleextexp_test",
6600 srcs = [
6601 "test/f32-vscaleextexp.cc",
6602 "test/vscaleextexp-microkernel-tester.h",
6603 ] + MICROKERNEL_TEST_HDRS,
6604 deps = MICROKERNEL_TEST_DEPS,
6605)
6606
6607xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006608 name = "f32_vsqr_test",
6609 srcs = [
6610 "test/f32-vsqr.cc",
6611 "test/vunary-microkernel-tester.h",
6612 ] + MICROKERNEL_TEST_HDRS,
6613 deps = MICROKERNEL_TEST_DEPS,
6614)
6615
6616xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006617 name = "f32_vsqrdiff_test",
6618 srcs = [
6619 "test/f32-vsqrdiff.cc",
6620 "test/vbinary-microkernel-tester.h",
6621 ] + MICROKERNEL_TEST_HDRS,
6622 deps = MICROKERNEL_TEST_DEPS,
6623)
6624
6625xnnpack_unit_test(
6626 name = "f32_vsqrdiffc_test",
6627 srcs = [
6628 "test/f32-vsqrdiffc.cc",
6629 "test/vbinaryc-microkernel-tester.h",
6630 ] + MICROKERNEL_TEST_HDRS,
6631 deps = MICROKERNEL_TEST_DEPS,
6632)
6633
6634xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006635 name = "f32_vsqrt_test",
6636 srcs = [
6637 "test/f32-vsqrt.cc",
6638 "test/vunary-microkernel-tester.h",
6639 ] + MICROKERNEL_TEST_HDRS,
6640 deps = MICROKERNEL_TEST_DEPS,
6641)
6642
6643xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006644 name = "f32_vsub_test",
6645 srcs = [
6646 "test/f32-vsub.cc",
6647 "test/vbinary-microkernel-tester.h",
6648 ] + MICROKERNEL_TEST_HDRS,
6649 deps = MICROKERNEL_TEST_DEPS,
6650)
6651
6652xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006653 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006654 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006655 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006656 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006657 ] + MICROKERNEL_TEST_HDRS,
6658 deps = MICROKERNEL_TEST_DEPS,
6659)
6660
6661xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006662 name = "f32_vsub_relu_test",
6663 srcs = [
6664 "test/f32-vsub-relu.cc",
6665 "test/vbinary-microkernel-tester.h",
6666 ] + MICROKERNEL_TEST_HDRS,
6667 deps = MICROKERNEL_TEST_DEPS,
6668)
6669
6670xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006671 name = "f32_vsubc_test",
6672 srcs = [
6673 "test/f32-vsubc.cc",
6674 "test/vbinaryc-microkernel-tester.h",
6675 ] + MICROKERNEL_TEST_HDRS,
6676 deps = MICROKERNEL_TEST_DEPS,
6677)
6678
6679xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006680 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006681 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006682 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006683 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006684 ] + MICROKERNEL_TEST_HDRS,
6685 deps = MICROKERNEL_TEST_DEPS,
6686)
6687
6688xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006689 name = "f32_vsubc_relu_test",
6690 srcs = [
6691 "test/f32-vsubc-relu.cc",
6692 "test/vbinaryc-microkernel-tester.h",
6693 ] + MICROKERNEL_TEST_HDRS,
6694 deps = MICROKERNEL_TEST_DEPS,
6695)
6696
6697xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006698 name = "f32_vrsubc_test",
6699 srcs = [
6700 "test/f32-vrsubc.cc",
6701 "test/vbinaryc-microkernel-tester.h",
6702 ] + MICROKERNEL_TEST_HDRS,
6703 deps = MICROKERNEL_TEST_DEPS,
6704)
6705
6706xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006707 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006708 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006709 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006710 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006711 ] + MICROKERNEL_TEST_HDRS,
6712 deps = MICROKERNEL_TEST_DEPS,
6713)
6714
6715xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006716 name = "f32_vrsubc_relu_test",
6717 srcs = [
6718 "test/f32-vrsubc-relu.cc",
6719 "test/vbinaryc-microkernel-tester.h",
6720 ] + MICROKERNEL_TEST_HDRS,
6721 deps = MICROKERNEL_TEST_DEPS,
6722)
6723
6724xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006725 name = "qs8_dwconv_minmax_test",
6726 srcs = [
6727 "test/qs8-dwconv-minmax.cc",
6728 "test/dwconv-microkernel-tester.h",
6729 "src/xnnpack/AlignedAllocator.h",
6730 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6731 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6732)
6733
6734xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006735 name = "qs8_gavgpool_minmax_test",
6736 srcs = [
6737 "test/qs8-gavgpool-minmax.cc",
6738 "test/gavgpool-microkernel-tester.h",
6739 "src/xnnpack/AlignedAllocator.h",
6740 ] + MICROKERNEL_TEST_HDRS,
6741 deps = MICROKERNEL_TEST_DEPS,
6742)
6743
6744xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006745 name = "qs8_gemm_minmax_test",
6746 srcs = [
6747 "test/qs8-gemm-minmax.cc",
6748 "test/gemm-microkernel-tester.h",
6749 "src/xnnpack/AlignedAllocator.h",
6750 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6751 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6752)
6753
6754xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006755 name = "qs8_igemm_minmax_test",
6756 srcs = [
6757 "test/qs8-igemm-minmax.cc",
6758 "test/gemm-microkernel-tester.h",
6759 "src/xnnpack/AlignedAllocator.h",
6760 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6761 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6762)
6763
6764xnnpack_unit_test(
6765 name = "qs8_requantization_test",
6766 srcs = [
6767 "src/xnnpack/requantization-stubs.h",
6768 "test/qs8-requantization.cc",
6769 "test/requantization-tester.h",
6770 ] + MICROKERNEL_TEST_HDRS,
6771 deps = MICROKERNEL_TEST_DEPS,
6772)
6773
6774xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006775 name = "qs8_vadd_minmax_test",
6776 srcs = [
6777 "test/qs8-vadd-minmax.cc",
6778 "test/vadd-microkernel-tester.h",
6779 ] + MICROKERNEL_TEST_HDRS,
6780 deps = MICROKERNEL_TEST_DEPS,
6781)
6782
6783xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006784 name = "qs8_vaddc_minmax_test",
6785 srcs = [
6786 "test/qs8-vaddc-minmax.cc",
6787 "test/vaddc-microkernel-tester.h",
6788 ] + MICROKERNEL_TEST_HDRS,
6789 deps = MICROKERNEL_TEST_DEPS,
6790)
6791
6792xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006793 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006794 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006795 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006796 "test/avgpool-microkernel-tester.h",
6797 "src/xnnpack/AlignedAllocator.h",
6798 ] + MICROKERNEL_TEST_HDRS,
6799 deps = MICROKERNEL_TEST_DEPS,
6800)
6801
6802xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006803 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006804 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006805 "test/qu8-dwconv-minmax.cc",
6806 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006807 "src/xnnpack/AlignedAllocator.h",
6808 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006809 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006810)
6811
6812xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006813 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006814 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006815 "test/qu8-igemm-minmax.cc",
6816 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006817 "src/xnnpack/AlignedAllocator.h",
6818 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006819 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006820)
6821
6822xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006823 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006824 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006825 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006826 "test/gavgpool-microkernel-tester.h",
6827 "src/xnnpack/AlignedAllocator.h",
6828 ] + MICROKERNEL_TEST_HDRS,
6829 deps = MICROKERNEL_TEST_DEPS,
6830)
6831
6832xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006833 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006834 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006835 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006836 "test/gemm-microkernel-tester.h",
6837 "src/xnnpack/AlignedAllocator.h",
6838 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006839 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006840)
6841
6842xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006843 name = "qu8_requantization_test",
6844 srcs = [
6845 "src/xnnpack/requantization-stubs.h",
6846 "test/qu8-requantization.cc",
6847 "test/requantization-tester.h",
6848 ] + MICROKERNEL_TEST_HDRS,
6849 deps = MICROKERNEL_TEST_DEPS,
6850)
6851
6852xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006853 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006854 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006855 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006856 "test/vadd-microkernel-tester.h",
6857 ] + MICROKERNEL_TEST_HDRS,
6858 deps = MICROKERNEL_TEST_DEPS,
6859)
6860
6861xnnpack_unit_test(
6862 name = "u8_clamp_test",
6863 srcs = [
6864 "test/u8-clamp.cc",
6865 "test/clamp-microkernel-tester.h",
6866 ] + MICROKERNEL_TEST_HDRS,
6867 deps = MICROKERNEL_TEST_DEPS,
6868)
6869
6870xnnpack_unit_test(
6871 name = "u8_lut32norm_test",
6872 srcs = [
6873 "test/u8-lut32norm.cc",
6874 "test/lut-norm-microkernel-tester.h",
6875 ] + MICROKERNEL_TEST_HDRS,
6876 deps = MICROKERNEL_TEST_DEPS,
6877)
6878
6879xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006880 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006881 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006882 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006883 "test/maxpool-microkernel-tester.h",
6884 ] + MICROKERNEL_TEST_HDRS,
6885 deps = MICROKERNEL_TEST_DEPS,
6886)
6887
6888xnnpack_unit_test(
6889 name = "u8_rmax_test",
6890 srcs = [
6891 "test/u8-rmax.cc",
6892 "test/rmax-microkernel-tester.h",
6893 ] + MICROKERNEL_TEST_HDRS,
6894 deps = MICROKERNEL_TEST_DEPS,
6895)
6896
6897xnnpack_unit_test(
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006898 name = "x32_depthtospace2d_chw2hwc_test",
Yury Kartynnike7841862020-11-04 18:22:18 -08006899 srcs = [
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006900 "test/x32-depthtospace2d-chw2hwc.cc",
6901 "test/depthtospace-microkernel-tester.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08006902 ] + MICROKERNEL_TEST_HDRS,
6903 deps = MICROKERNEL_TEST_DEPS,
6904)
6905
6906xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006907 name = "x32_fill_test",
6908 srcs = [
6909 "test/x32-fill.cc",
6910 "test/fill-microkernel-tester.h",
6911 ] + MICROKERNEL_TEST_HDRS,
6912 deps = MICROKERNEL_TEST_DEPS,
6913)
6914
6915xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006916 name = "x32_packx_test",
6917 srcs = [
6918 "test/x32-packx.cc",
6919 "test/pack-microkernel-tester.h",
6920 "src/xnnpack/AlignedAllocator.h",
6921 ] + MICROKERNEL_TEST_HDRS,
6922 deps = MICROKERNEL_TEST_DEPS,
6923)
6924
6925xnnpack_unit_test(
6926 name = "x32_pad_test",
6927 srcs = [
6928 "test/x32-pad.cc",
6929 "test/pad-microkernel-tester.h",
6930 ] + MICROKERNEL_TEST_HDRS,
6931 deps = MICROKERNEL_TEST_DEPS,
6932)
6933
6934xnnpack_unit_test(
6935 name = "x32_unpool_test",
6936 srcs = [
6937 "test/x32-unpool.cc",
6938 "test/unpool-microkernel-tester.h",
6939 ] + MICROKERNEL_TEST_HDRS,
6940 deps = MICROKERNEL_TEST_DEPS,
6941)
6942
6943xnnpack_unit_test(
6944 name = "x32_zip_test",
6945 srcs = [
6946 "test/x32-zip.cc",
6947 "test/zip-microkernel-tester.h",
6948 ] + MICROKERNEL_TEST_HDRS,
6949 deps = MICROKERNEL_TEST_DEPS,
6950)
6951
6952xnnpack_unit_test(
6953 name = "x8_lut_test",
6954 srcs = [
6955 "test/x8-lut.cc",
6956 "test/lut-microkernel-tester.h",
6957 ] + MICROKERNEL_TEST_HDRS,
6958 deps = MICROKERNEL_TEST_DEPS,
6959)
6960
6961xnnpack_unit_test(
6962 name = "x8_zip_test",
6963 srcs = [
6964 "test/x8-zip.cc",
6965 "test/zip-microkernel-tester.h",
6966 ] + MICROKERNEL_TEST_HDRS,
6967 deps = MICROKERNEL_TEST_DEPS,
6968)
6969
Marat Dukhan20c3b922020-03-10 03:45:06 -07006970########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006971
6972xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006973 name = "operator_size_test",
6974 srcs = ["test/operator-size.c"],
Marat Dukhanf0cb70a2021-03-30 15:45:15 -07006975 deps = [":xnnpack_for_tfjs"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006976)
6977
Marat Dukhan20c3b922020-03-10 03:45:06 -07006978xnnpack_binary(
6979 name = "subgraph_size_test",
6980 srcs = ["test/subgraph-size.c"],
6981 deps = [":XNNPACK"],
6982)
6983
6984########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006985
6986xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006987 name = "abs_nc_test",
6988 srcs = [
6989 "test/abs-nc.cc",
6990 "test/abs-operator-tester.h",
6991 ],
6992 deps = OPERATOR_TEST_DEPS,
6993)
6994
6995xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006996 name = "add_nd_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08006997 timeout = "moderate",
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006998 srcs = [
6999 "test/add-nd.cc",
7000 "test/binary-elementwise-operator-tester.h",
7001 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007002 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007003)
7004
7005xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007006 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007007 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007008 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007009 "test/argmax-pooling-operator-tester.h",
7010 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007011 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007012)
7013
7014xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007015 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007016 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007017 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007018 "test/average-pooling-operator-tester.h",
7019 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007020 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007021)
7022
7023xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007024 name = "bankers_rounding_nc_test",
7025 srcs = [
7026 "test/bankers-rounding-nc.cc",
7027 "test/bankers-rounding-operator-tester.h",
7028 ],
7029 deps = OPERATOR_TEST_DEPS,
7030)
7031
7032xnnpack_unit_test(
7033 name = "ceiling_nc_test",
7034 srcs = [
7035 "test/ceiling-nc.cc",
7036 "test/ceiling-operator-tester.h",
7037 ],
7038 deps = OPERATOR_TEST_DEPS,
7039)
7040
7041xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007042 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007043 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007044 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007045 "test/channel-shuffle-operator-tester.h",
7046 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007047 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007048)
7049
7050xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007051 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007052 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007053 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007054 "test/clamp-operator-tester.h",
7055 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007056 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007057)
7058
7059xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07007060 name = "constant_pad_nd_test",
7061 srcs = [
7062 "test/constant-pad-nd.cc",
7063 "test/constant-pad-operator-tester.h",
7064 ],
7065 deps = OPERATOR_TEST_DEPS,
7066)
7067
7068xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007069 name = "convolution_nhwc_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007070 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007071 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007072 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007073 "test/convolution-operator-tester.h",
7074 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007075 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007076)
7077
7078xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007079 name = "convolution_nchw_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007080 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007081 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007082 "test/convolution-nchw.cc",
7083 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007084 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007085 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007086)
7087
7088xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07007089 name = "copy_nc_test",
7090 srcs = [
7091 "test/copy-nc.cc",
7092 "test/copy-operator-tester.h",
7093 ],
7094 deps = OPERATOR_TEST_DEPS,
7095)
7096
7097xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007098 name = "deconvolution_nhwc_test",
Artsiom Ablavatskic1aa2972020-12-08 11:23:34 -08007099 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007100 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007101 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007102 "test/deconvolution-operator-tester.h",
7103 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007104 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007105)
7106
7107xnnpack_unit_test(
Marat Dukhan188d1042020-11-24 23:39:40 -08007108 name = "depth_to_space_nchw2nhwc_test",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08007109 srcs = [
Marat Dukhan188d1042020-11-24 23:39:40 -08007110 "test/depth-to-space-nchw2nhwc.cc",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08007111 "test/depth-to-space-operator-tester.h",
7112 ] + OPERATOR_TEST_PARAMS_HDRS,
7113 deps = OPERATOR_TEST_DEPS,
7114)
7115
7116xnnpack_unit_test(
Marat Dukhan0e521172020-11-25 13:10:04 -08007117 name = "depth_to_space_nhwc_test",
7118 srcs = [
7119 "test/depth-to-space-nhwc.cc",
7120 "test/depth-to-space-operator-tester.h",
7121 ] + OPERATOR_TEST_PARAMS_HDRS,
7122 deps = OPERATOR_TEST_DEPS,
7123)
7124
7125xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08007126 name = "divide_nd_test",
7127 srcs = [
7128 "test/binary-elementwise-operator-tester.h",
7129 "test/divide-nd.cc",
7130 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007131 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08007132)
7133
7134xnnpack_unit_test(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08007135 name = "elu_nc_test",
7136 srcs = [
7137 "test/elu-nc.cc",
7138 "test/elu-operator-tester.h",
7139 ],
7140 deps = OPERATOR_TEST_DEPS,
7141)
7142
7143xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007144 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007145 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007146 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007147 "test/fully-connected-operator-tester.h",
7148 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007149 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007150)
7151
7152xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007153 name = "floor_nc_test",
7154 srcs = [
7155 "test/floor-nc.cc",
7156 "test/floor-operator-tester.h",
7157 ],
7158 deps = OPERATOR_TEST_DEPS,
7159)
7160
7161xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007162 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007163 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007164 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007165 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07007166 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007167 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007168)
7169
7170xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007171 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007172 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007173 "test/global-average-pooling-ncw.cc",
7174 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007175 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007176 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007177)
7178
7179xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007180 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007181 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007182 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007183 "test/hardswish-operator-tester.h",
7184 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007185 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007186)
7187
7188xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007189 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007190 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007191 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007192 "test/leaky-relu-operator-tester.h",
7193 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007194 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007195)
7196
7197xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007198 name = "max_pooling_nhwc_test",
Artsiom Ablavatski2202c812021-01-22 14:16:43 -08007199 timeout = "moderate",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007200 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007201 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007202 "test/max-pooling-operator-tester.h",
7203 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007204 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007205)
7206
7207xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08007208 name = "maximum_nd_test",
7209 srcs = [
7210 "test/binary-elementwise-operator-tester.h",
7211 "test/maximum-nd.cc",
7212 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007213 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08007214)
7215
7216xnnpack_unit_test(
7217 name = "minimum_nd_test",
7218 srcs = [
7219 "test/binary-elementwise-operator-tester.h",
7220 "test/minimum-nd.cc",
7221 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007222 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08007223)
7224
7225xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007226 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08007227 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08007228 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08007229 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08007230 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007231 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08007232)
7233
7234xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007235 name = "negate_nc_test",
7236 srcs = [
7237 "test/negate-nc.cc",
7238 "test/negate-operator-tester.h",
7239 ],
7240 deps = OPERATOR_TEST_DEPS,
7241)
7242
7243xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007244 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007245 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007246 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007247 "test/prelu-operator-tester.h",
7248 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007249 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007250)
7251
7252xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007253 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08007254 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007255 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08007256 "test/resize-bilinear-operator-tester.h",
7257 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07007258 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08007259)
7260
7261xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07007262 name = "resize_bilinear_nchw_test",
7263 srcs = [
7264 "test/resize-bilinear-nchw.cc",
7265 "test/resize-bilinear-operator-tester.h",
7266 ] + OPERATOR_TEST_PARAMS_HDRS,
7267 deps = OPERATOR_TEST_DEPS,
7268)
7269
7270xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007271 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007272 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007273 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007274 "test/sigmoid-operator-tester.h",
7275 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007276 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007277)
7278
7279xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08007280 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007281 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08007282 "test/softmax-nc.cc",
7283 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007284 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007285 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007286)
7287
7288xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07007289 name = "square_nc_test",
7290 srcs = [
7291 "test/square-nc.cc",
7292 "test/square-operator-tester.h",
7293 ],
7294 deps = OPERATOR_TEST_DEPS,
7295)
7296
7297xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07007298 name = "square_root_nc_test",
7299 srcs = [
7300 "test/square-root-nc.cc",
7301 "test/square-root-operator-tester.h",
7302 ],
7303 deps = OPERATOR_TEST_DEPS,
7304)
7305
7306xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07007307 name = "squared_difference_nd_test",
7308 srcs = [
7309 "test/binary-elementwise-operator-tester.h",
7310 "test/squared-difference-nd.cc",
7311 ],
7312 deps = OPERATOR_TEST_DEPS,
7313)
7314
7315xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007316 name = "subtract_nd_test",
7317 srcs = [
7318 "test/binary-elementwise-operator-tester.h",
7319 "test/subtract-nd.cc",
7320 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007321 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007322)
7323
7324xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007325 name = "truncation_nc_test",
7326 srcs = [
7327 "test/truncation-nc.cc",
7328 "test/truncation-operator-tester.h",
7329 ],
7330 deps = OPERATOR_TEST_DEPS,
7331)
7332
7333xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007334 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007335 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007336 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007337 "test/unpooling-operator-tester.h",
7338 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007339 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007340)
7341
Chao Mei6ddfc602020-05-13 22:29:36 -07007342############################### Misc unit tests ###############################
7343
7344xnnpack_unit_test(
7345 name = "memory_planner_test",
7346 srcs = [
7347 "test/memory-planner-test.cc",
7348 ],
7349 deps = [
7350 ":XNNPACK",
7351 ":memory_planner",
7352 ],
7353)
7354
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07007355xnnpack_unit_test(
7356 name = "subgraph_nchw_test",
7357 srcs = [
7358 "src/xnnpack/subgraph.h",
7359 "test/subgraph-nchw.cc",
7360 "test/subgraph-tester.h",
7361 ],
7362 deps = [
7363 ":XNNPACK",
7364 ],
7365)
7366
Marat Dukhan08c4a432019-10-03 09:29:21 -07007367############################# Build configurations #############################
7368
Marat Dukhanb8642352019-10-30 15:43:02 -07007369# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07007370config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007371 name = "xnn_enable_assembly_explicit_true",
7372 define_values = {"xnn_enable_assembly": "true"},
7373)
7374
7375# Disables usage of assembly kernels.
7376config_setting(
7377 name = "xnn_enable_assembly_explicit_false",
7378 define_values = {"xnn_enable_assembly": "false"},
7379)
7380
Marat Dukhan9de90e02020-06-18 16:04:12 -07007381# Enables usage of sparse inference.
7382config_setting(
7383 name = "xnn_enable_sparse_explicit_true",
7384 define_values = {"xnn_enable_sparse": "true"},
7385)
7386
7387# Disables usage of sparse inference.
7388config_setting(
7389 name = "xnn_enable_sparse_explicit_false",
7390 define_values = {"xnn_enable_sparse": "false"},
7391)
7392
Marat Dukhan05702cf2020-03-26 15:41:33 -07007393# Disables usage of HMP-aware optimizations.
7394config_setting(
7395 name = "xnn_enable_hmp_explicit_false",
7396 define_values = {"xnn_enable_hmp": "false"},
7397)
7398
Chao Mei6ddfc602020-05-13 22:29:36 -07007399# Enable usage of optimized memory allocation
7400config_setting(
7401 name = "xnn_enable_memopt_explicit_true",
Marat Dukhan03f46212021-03-30 21:29:49 -07007402 define_values = {"xnn_enable_memopt": "true"},
Chao Mei6ddfc602020-05-13 22:29:36 -07007403)
7404
7405# Disable usage of optimized memory allocation
7406config_setting(
7407 name = "xnn_enable_memopt_explicit_false",
Marat Dukhan03f46212021-03-30 21:29:49 -07007408 define_values = {"xnn_enable_memopt": "false"},
Chao Mei6ddfc602020-05-13 22:29:36 -07007409)
7410
Marat Dukhanb939cdb2021-03-30 18:51:51 -07007411# Enable QS8 inference in TFLite-specific version
7412config_setting(
7413 name = "xnn_enable_qs8_explicit_true",
7414 define_values = {"xnn_enable_qs8": "true"},
7415)
7416
7417# Disable QS8 inference in TFLite-specific version
7418config_setting(
7419 name = "xnn_enable_qs8_explicit_false",
7420 define_values = {"xnn_enable_qs8": "false"},
7421)
7422
Marat Dukhanb8642352019-10-30 15:43:02 -07007423# Builds with -c dbg
7424config_setting(
7425 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007426 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07007427 "compilation_mode": "dbg",
7428 },
7429)
7430
7431# Builds with -c opt
7432config_setting(
7433 name = "optimized_build",
7434 values = {
7435 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007436 },
7437)
7438
7439config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007440 name = "linux_k8",
7441 values = {"cpu": "k8"},
7442)
7443
7444config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007445 name = "linux_arm",
7446 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07007447)
7448
7449config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07007450 name = "linux_armeabi",
7451 values = {"cpu": "armeabi"},
7452)
7453
7454config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07007455 name = "linux_armhf",
7456 values = {"cpu": "armhf"},
7457)
7458
7459config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07007460 name = "linux_armv7a",
7461 values = {"cpu": "armv7a"},
7462)
7463
7464config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007465 name = "linux_aarch64",
7466 values = {"cpu": "aarch64"},
7467)
7468
7469config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007470 name = "android",
7471 values = {"crosstool_top": "//external:android/crosstool"},
7472)
7473
7474config_setting(
7475 name = "android_armv7",
7476 values = {
7477 "crosstool_top": "//external:android/crosstool",
7478 "cpu": "armeabi-v7a",
7479 },
7480)
7481
7482config_setting(
7483 name = "android_arm64",
7484 values = {
7485 "crosstool_top": "//external:android/crosstool",
7486 "cpu": "arm64-v8a",
7487 },
7488)
7489
7490config_setting(
7491 name = "android_x86",
7492 values = {
7493 "crosstool_top": "//external:android/crosstool",
7494 "cpu": "x86",
7495 },
7496)
7497
7498config_setting(
7499 name = "android_x86_64",
7500 values = {
7501 "crosstool_top": "//external:android/crosstool",
7502 "cpu": "x86_64",
7503 },
7504)
7505
7506config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007507 name = "windows_x86_64",
7508 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007509)
7510
7511config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007512 name = "windows_x86_64_clang",
7513 values = {
7514 "compiler": "clang-cl",
7515 "cpu": "x64_windows",
7516 },
7517)
7518
7519config_setting(
7520 name = "windows_x86_64_mingw",
7521 values = {
7522 "compiler": "mingw-gcc",
7523 "cpu": "x64_windows",
7524 },
7525)
7526
7527config_setting(
7528 name = "windows_x86_64_msys",
7529 values = {
7530 "compiler": "msys-gcc",
7531 "cpu": "x64_windows",
7532 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007533)
7534
7535config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07007536 name = "macos_x86_64",
7537 values = {
7538 "apple_platform_type": "macos",
7539 "cpu": "darwin",
7540 },
7541)
7542
7543config_setting(
Simon Maurerae33ab82021-03-03 23:38:22 +01007544 name = "macos_arm64",
7545 values = {
7546 "apple_platform_type": "macos",
7547 "cpu": "darwin_arm64",
7548 },
7549)
7550
7551config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007552 name = "emscripten",
XNNPACK Team3bfbdaf2021-03-29 15:26:23 -07007553 values = {"crosstool_top": "@emsdk//emscripten_toolchain:everything"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07007554)
7555
7556config_setting(
7557 name = "emscripten_wasm",
7558 values = {
XNNPACK Team3bfbdaf2021-03-29 15:26:23 -07007559 "crosstool_top": "@emsdk//emscripten_toolchain:everything",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007560 "cpu": "wasm",
7561 },
7562)
7563
7564config_setting(
7565 name = "emscripten_wasmsimd",
7566 values = {
XNNPACK Team3bfbdaf2021-03-29 15:26:23 -07007567 "crosstool_top": "@emsdk//emscripten_toolchain:everything",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007568 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07007569 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007570 },
7571)
7572
7573config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007574 name = "ios_armv7",
7575 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007576 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007577 "cpu": "ios_armv7",
7578 },
7579)
7580
7581config_setting(
7582 name = "ios_arm64",
7583 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007584 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007585 "cpu": "ios_arm64",
7586 },
7587)
7588
7589config_setting(
7590 name = "ios_arm64e",
7591 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007592 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007593 "cpu": "ios_arm64e",
7594 },
7595)
7596
7597config_setting(
7598 name = "ios_x86",
7599 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007600 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007601 "cpu": "ios_i386",
7602 },
7603)
7604
7605config_setting(
7606 name = "ios_x86_64",
7607 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007608 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007609 "cpu": "ios_x86_64",
7610 },
7611)
7612
7613config_setting(
7614 name = "watchos_armv7k",
7615 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007616 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007617 "cpu": "watchos_armv7k",
7618 },
7619)
7620
7621config_setting(
7622 name = "watchos_arm64_32",
7623 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007624 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007625 "cpu": "watchos_arm64_32",
7626 },
7627)
7628
7629config_setting(
7630 name = "watchos_x86",
7631 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007632 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007633 "cpu": "watchos_i386",
7634 },
7635)
7636
7637config_setting(
7638 name = "watchos_x86_64",
7639 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007640 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007641 "cpu": "watchos_x86_64",
7642 },
7643)
7644
7645config_setting(
7646 name = "tvos_arm64",
7647 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007648 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007649 "cpu": "tvos_arm64",
7650 },
7651)
7652
7653config_setting(
7654 name = "tvos_x86_64",
7655 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007656 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007657 "cpu": "tvos_x86_64",
7658 },
7659)